Scalable nonlinear Cox modeling via random Fourier features with analytic uncertainty Authors: Fahrettin Kaya Published: 2026 ABSTRACT BackgroundThe Cox proportional hazards model often fails to capture complex biomedical risk structures, such as U-shaped biomarker associations, due to its assumption of linearity between the log-hazard and covariates. While existing kernel-based generalizations offer the necessary flexibility, their \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{3}\right)$$\end{document} computational complexity limits applicability in large-scale cohort studies. Furthermore, most non-linear machine learning methods lack closed-form analytical measures of uncertainty for individual predictions..MethodsA novel Random Fourier Features-based Cox regression approach (RFF-Cox) is presented to model non-linear risk relationships within a scalable framework. By mapping stationary kernels into a finite-dimensional explicit feature space, the method reduces computational complexity to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n{m}^{2}\right)$$\end{document}. Model parameters are estimated via the Newton–Raphson algorithm on a ridge-regularized partial likelihood, while the bandwidth parameter (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}) is automatically optimized using a marginal likelihood criterion based on the Laplace approximation. Uncertainty quantification is performed via the Fisher information matrix, with a multivariate Delta method propagating variance from both parameter estimation and baseline hazard estimation. A Taylor-expansion-based inference framework enables covariate-level hazard ratio estimation, Wald-type significance testing, and formal interaction detection. Performance was evaluated using controlled simulations and four real-world datasets with sample sizes ranging from 500 to 9,105.ResultsIn simulation scenarios, RFF-Cox degenerated to classical Cox estimates under linearity (RMSE: 0.111 vs. 0.110) while demonstrating a marked accuracy advantage under non-linearity (RMSE: 0.137 vs. 0.314), including the recovery of U-shaped risk functions. The true SBP × Smoking interaction was detected (estimate −1.093, p < 0.001). Analytical 95% CI coverage reached 95.9% in the linear scenario and 81.9% in the non-linear scenario. In real-world applications, the model exhibited discriminatory power competitive with Random Survival Forests and XGBoost with favorable computational efficiency. IPCW-weighted calibration analyses on the METABRIC dataset yielded low Integrated Calibration Error (ICI < 0.05) across most time horizons, though calibration slope deviations were noted at longer follow-up periods. Moreover, uncertainty in individual predictions, quantified via analytical confidence intervals, varied meaningfully across risk groups.ConclusionsRFF-Cox provides a practical survival analysis framework that degenerates to the classical Cox model under linearity while offering non-linear modelling capabilities, computational efficiency, and formal inference tools — including covariate-level interpretation and interaction detection — that are not readily available in tree-based or deep learning survival methods.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12874-026-02925-x. Introduction The Cox proportional hazards model is the most widely used semi-parametric framework in survival analysis [1]. By estimating covariate effects via partial likelihood without specifying the baseline hazard function parametrically, it has become the standard approach in clinical research [2]. However, the model assumes a linear relationship between the log-hazard and covariates. This assumption often proves inadequate for capturing non-linear biomedical relationships, such as genetic interactions, dose-dependent saturation effects, or the U-shaped risk associations frequently observed in biomarkers. To relax the linearity assumption, kernel-based Cox regression was proposed [3]. In this approach, the log-hazard function is defined within a Reproducing Kernel Hilbert Space (RKHS), and by the Representer Theorem, the solution is expressed as a finite combination of kernel functions [4]. Nevertheless, this method requires a dense kernel matrix, and its computational cost scales cubically with the number of observations. This limitation hinders its applicability to large cohort datasets [5, 6]. Random Fourier Features (RFF) mitigate the computational burden of kernel methods by approximating stationary kernels in a finite-dimensional feature space, an approach grounded in sampling from the spectral density of the kernel function [7, 8]. Despite this potential, the integration of RFF-based approaches with censored survival data remains limited in the literature. Two primary methodological barriers currently impede applications in this domain. First, the performance of the Radial Basis Function (RBF) kernel is highly sensitive to the bandwidth parameter (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}). Yet, determining an optimal value for this parameter in the context of censored data structures typically necessitates cross-validation, a process that incurs substantial computational costs. Second, and more broadly, many non-linear survival models fail to provide closed-form analytical measures of uncertainty for individual predictions. Consequently, assessing the uncertainty of model predictions often relies on computationally expensive resampling techniques, such as the bootstrap. Several approaches have addressed non-linearity in the Cox model. Kernel Cox regression [3] provides theoretical flexibility but incurs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{3}\right)$$\end{document} computational cost for large cohorts. DeepSurv [9] offers universal approximation but lacks closed-form uncertainty measures, requiring resampling for confidence intervals. Tree-based methods (RSF, XGBoost) capture interactions implicitly [10] but do not provide formal hypothesis testing. The RFF-Cox framework addresses these gaps by combining \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n{m}^{2}\right)$$\end{document} scalability, analytical uncertainty quantification through the Fisher information matrix, and a Taylor-expansion-based inference framework enabling covariate-level Wald tests and interaction detection. This study proposes the RFF-Cox framework, designed to derive individual risk estimates from censored survival data in a manner that is more flexible, efficient, and interpretable. By leveraging random Fourier features, the method effectively models complex, non-linear risk relationships, thereby overcoming the limitations of classical linear approaches while offering a scalable and rapid prediction process suitable for large-scale datasets. Furthermore, the framework introduces a Taylor-expansion-based inference procedure that bridges the classical linear Cox model and the non-linear RFF-Cox model, enabling covariate-level hazard ratio estimation, Wald-type significance testing, and formal interaction detection [11, 12]. The framework also enables the quantitative assessment of uncertainty in individual survival estimates through a multivariate Delta method that propagates variance from both parameter estimation and baseline hazard estimation [13, 14], allowing for the derivation of analytical confidence intervals (CIs) for each prediction. In these respects, RFF-Cox offers a practical survival modeling framework capable of capturing non-linear risk structures, operating with high computational efficiency, and explicitly quantifying individual prediction uncertainty. Theoretical background This section systematically summarizes the theoretical foundations underpinning the proposed method: the classical Cox model, kernel-based generalizations, and the Random Fourier Features (RFF) approach. Cox proportional hazards model The reference model in survival analysis is the Proportional Hazards (PH) model, introduced by Cox [1]. Throughout this work, d is assumed to be fixed. For a covariate vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\in\:{R}^{d}$$\end{document}, the model defines the conditional hazard function as: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h(t \mid x) = h_0(t) \exp(x^\top \beta_{\textit{cox}})$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{0}\left(t\right)$$\end{document} denotes the non-parametric baseline hazard function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta_{cox}\in{R}^{d}$$\end{document} represents the coefficient vector capturing the constant effects of covariates on the log-hazard. The most distinct feature of this model is that the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta_{cox}$$\end{document} can be estimated via the partial likelihood method without specifying the form of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{0}\left(t\right)$$\end{document}. This framework was later extended by Norman Breslow to estimate the baseline hazard function [15]. However, the assumption of a linear relationship (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}^{\top}\beta_{cox}$$\end{document}) between the log-hazard and covariates can be a limiting factor when modeling complex non-linear phenomena, such as genetic epistasis, saturation points in dose-response curves, or U-shaped risk associations observed in biomarkers. While Therneau and Grambsch have detailed the standard applications [2], the linearity constraint remains a known limitation. Functional generalization and Kernel Approach When the linear predictor in the Cox model is generalized to a nonparametric function, the log-hazard function is expressed as: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{log}h\left(t\mid{x}\right)=\mathrm{log}{h}_{0}\left(t\right)+\eta\:\left({x}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left(\cdot\:\right)$$\end{document} is an unknown function defining the non-linear relationship between covariates and risk. This problem can be formulated as a regularized optimization problem within a Reproducing Kernel Hilbert Space (RKHS). The Representer Theorem, established by Kimeldorf and Wahba [4], demonstrates that for such problems, the optimal solution can be expressed as a linear combination of a finite number of kernel functions [4, 5]. Lin et al. (2011) adapted this approach to the Cox model, proposing kernel-based Cox regression and showing that the non-linear risk function could be estimated within the RKHS framework via partial likelihood maximization [3]. Random fourier feature approximation of shift-invariant Kernels While kernel methods offer a powerful framework for modeling non-linear relationships, their classical implementations require computing and storing a dense \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\:\times\:n$$\end{document} Gram matrix for n observations. This results in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{2}\right)$$\end{document} memory and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{3}\right)$$\end{document} computational complexity [6, 7]. Due to this computational burden, approximating kernel functions via finite-dimensional feature maps has gained prominence. For shift-invariant kernels, the theoretical basis of this approach relies on Bochner’s Theorem [16]. According to the theorem, any continuous, positive-definite kernel \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\left({x}-{{x}}^{{{\prime\:}}}\right)$$\end{document} can be expressed as the Fourier transform of a probability measure [6, 16]: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\left({x},{{x}}^{{{\prime\:}}}\right)=k\left({x}-{{x}}^{{{\prime\:}}}\right)={\int}_{{\mathrm{R}}^{d}}p\left({\omega\:}\right){e}^{i{{\omega\:}}^{{\top\:}}\left({x}-{{x}}^{{{\prime\:}}}\right)}d{\omega\:}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({\omega\:}\right)$$\end{document} is a probability distribution representing the spectral density of the kernel function. By taking the real part of the complex exponential and applying Euler’s formula and trigonometric identities, this expression can be transformed into the following expected value form involving a random phase variable \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:b\sim\:\mathrm{Uniform}\left(\mathrm{0,2}{\uppi\:}\right)$$\end{document}: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\left({x},{{x}}^{{{\prime\:}}}\right)={E}_{{\omega\:},b}\left[\mathrm{cos}\left({{\omega\:}}^{{\top\:}}{x}+b\right)\mathrm{cos}\left({\{\omega\:}^{{\top}}{{x}}^{{{\prime\:}}}+b\right)\right]$$\end{document} Rahimi and Recht proposed reducing this infinite-dimensional integral representation to a finite-dimensional feature map via Monte Carlo sampling [7]. Specifically, for the Radial Basis Function (RBF) Gaussian kernel: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\left({x},{{x}}^{{{\prime\:}}}\right)=\mathrm{exp}\left(-\frac{\parallel{x}-{{x}}^{{{\prime\:}}}\parallel^{2}}{2{{\upsigma\:}}^{2}}\right)$$\end{document} the spectral density distribution is Gaussian: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left({\omega\:}\right)=\mathcal{N}\left(0,{{\upsigma\:}}^{-2}{{I}}_{d}\right)$$\end{document}. Due to the inverse variance transformation in the Fourier pair, sampling the weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}$$\end{document} satisfies the condition \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Var}\left({w}\right)={{\upsigma\:}}^{-2}$$\end{document}, which corresponds directly to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{w}}_{j}\sim\:\mathcal{N}\left(0,{{\upsigma\:}}^{-2}{{I}}_{d}\right)$$\end{document}, for j = 1,…,m, in practice. By drawing m independent frequency vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\omega\:}}_{j}$$\end{document} and phases \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{j}$$\end{document} from this spectral distribution, the following finite-dimensional feature map is constructed: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi\:\left({x}\right)=\sqrt{\frac{2}{m}}\:\left[\begin{array}{c}{cos}\left({{\omega\:}}_{1}^{{\top\:}}{x}+{b}_{1}\right)\\\:.\\\:.\\\:.\\\:{cos}\left({{\omega\:}}_{m}^{{\top\:}}{x}+{b}_{m}\right)\end{array}\right]\:\in\:{R}^{m},$$\end{document} By the Law of Large Numbers, this transformation ensures that the inner product in the feature space converges to the original kernel value. This approach is based on the principle of approximating infinite-dimensional kernel maps with finite-dimensional stochastic projections [11]: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}&{\mathrm{E}}_{{\omega\:},b}\left[\phi\:{\left({x}\right)}^{{\top\:}}\phi\:\left({{x}}^{{\prime\:}}\right)\right]=\mathrm{k}\left({x},{{x}}^{{{\prime\:}}}\right)\\&\Rightarrow \:k\left({x},{{x}}^{{{\prime\:}}}\right)\approx\:\:\phi\:{\left({x}\right)}^{{\top\:}}\phi\:\left({{x}}^{{\prime\:}}\right)\end{aligned}$$\end{document} For a data matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}\in\:{R}^{n\times\:d}$$\end{document}, computing the explicit feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Phi\:}\in\:{R}^{n\times\:m}$$\end{document} has a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{O}\left(ndm\right)$$\end{document}. This approach eliminates the necessity of constructing an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\:\times\:n$$\end{document} dense kernel matrix, reducing computational complexity from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{O}\left({n}^{3}\right)$$\end{document} in classical methods to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{O}\left(n{m}^{2}+{m}^{3}\right)$$\end{document}. When \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\ll\:n$$\end{document}, this structure provides near-linear scalability for large datasets [7, 17]. Furthermore, since the kernel method is thereby mapped to an explicit linear feature space, it allows for the direct application of standard linear optimization algorithms. These advantages constitute the core computational efficiency and practical applicability of the proposed RFF-Cox method. Proposed method: RFF-cox framework This section details the RFF-Cox method, developed by integrating the theoretical components presented in previous sections. The proposed framework introduces three key innovations: scalable optimization, data-driven automatic bandwidth selection, and asymptotic uncertainty quantification. This approach reduces the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{3}\right)$$\end{document} computational cost inherent in classical kernel methods, eigen-decomposition and matrix symmetrization, and ensures statistical reliability through full covariance matrix derivation. RFF-based cox model formulation The linear component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{x}}^{\top\:}{{\beta\:}}_{{c}{o}{x}}$$\end{document} in the classical Cox model is redefined in the explicit feature space obtained via RFF as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi\:{\left({x}\right)}^{{\top\:}}{\beta\:}$$\end{document}: 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h\left(t\mid{x}\right)={h}_{0}\left(t\right)\mathrm{exp}\left(\phi\:{\left({x}\right)}^{{\top\:}}{\beta\:}\right)$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upphi\:}\left({x}\right)\in\:{R}^{m}$$\end{document} denotes the Random Fourier Features transformation defined in Sect. Random Fourier Feature Approximation of Shift-Invariant Kernels, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}\in\:{R}^{m}$$\end{document} is the weight vector to be estimated. This formulation preserves the proportional hazards (PH) assumption, as the hazard ratio between any two individuals remains independent of time. Thus, the non-linear risk structure is reduced to a linear parameter estimation problem in an m-dimensional feature space (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\ll\:n$$\end{document}). The partial likelihood function is expressed as: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l\left({\beta\:}\right)={\sum\:}_{i:{E}_{i}=1}\left[\phi\:{\left({{x}}_{{i}}\right)}^{{\top\:}}{\beta\:}-\mathrm{log}\left({\sum\:}_{j\in\:R\left({t}_{i}\right)}\mathrm{exp}\left(\phi\:{\left({{x}}_{j}\right)}^{{\top\:}}{\beta\:}\right)\right)\right]$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E_{i}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} denotes the event indicator for observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i,t_{i}$$\end{document} denotes the observed time, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R\left(t_{i}\right)=j:t_{j}\ge\:t_{i}$$\end{document} denotes the risk set at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t_{i}$$\end{document} Defining the model dimension by the number of features m rather than the sample size n eliminates the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{2}\right)$$\end{document} memory requirement associated with dense kernel matrices, thereby enhancing the scalability of the method. Newton–Raphson optimization for penalized partial likelihood To prevent overfitting and improve estimation stability, a Ridge (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{2}$$\end{document}) penalty is added to the objective function [2, 18, 19]. The penalized log-likelihood function to be optimized is defined as: 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{\mathrm{p}}\left({\beta\:}\right)=l\left({\beta\:}\right)-\frac{{\uplambda\:}}{2}\left|\right|{\beta\:}{\left|\right|}_{2}^{2},\:$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:>\:0$$\end{document} is the parameter controlling the regularization strength. The optimization is performed using the Newton–Raphson algorithm, which utilizes second-order derivative information. The iterative update, using the gradient vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}\left({\beta\:}\right)$$\end{document} and the Hessian matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}\left({\beta\:}\right)$$\end{document}, is given by: 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\beta\:}}^{\left(k+1\right)}={{\beta\:}}^{\left(k\right)}-{H}{\left({{\beta\:}}^{\left(k\right)}\right)}^{-1}{U}\left({{\beta\:}}^{\left(k\right)}\right)$$\end{document} Since the dimension of the model parameters is limited to m following the Random Fourier Features transformation, the cost of computing the gradient and Hessian at each iteration is approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n{m}^{2}\right)$$\end{document}, while inverting the Hessian matrix requires \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({m}^{3}\right)$$\end{document} complexity [7, 17, 19]. In cases where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\ll\:\:n$$\end{document}, this approach offers a scalable solution for large datasets compared to classical kernel methods. To reduce computational costs, risk set sums are calculated in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n\right)$$\end{document} complexity using the reverse cumulative sum method [2]. To prevent floating-point overflow in exponential computations, values are clipped to the range [− 700,700] as a standard IEEE 754 safeguard. Within this framework, the variance weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\mu\:}}_{{i}}\:$$\end{document} forming the Fisher information matrix for the i-th observation is derived according to the Breslow method: 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\upmu\:}}_{i}={r}_{i}{\sum\:}_{j:{t}_{j}\le\:{t}_{i},{E}_{j}=1}\frac{1}{R\left({t}_{j}\right)}={r}_{i}{H}_{i},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{i}=\mathrm{exp}\left(\phi\:{\left({{x}}_{{i}}\right)}^{{\top\:}}{\beta\:}\right)$$\end{document} represents the individual risk score, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R\left(t_{j}\right)={\varSigma\:}_{l\in\:R\left(t_{j}\right)}r_{\mathrm{l}}$$\end{document} is the total risk score at event time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t_{j}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H_{i}={\varSigma\:}_{j:t_{j}\le\:t_{i},E_{j}=1}1/R\left(t_{j}\right)$$\end{document} is the cumulative Breslow weight that accumulates the inverse risk-set contributions from all events at or before \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t_{i}$$\end{document}. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{i}$$\end{document} represents the observation-specific variance contribution in the Hessian matrix of the partial likelihood. This value is non-zero not only for events (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}=1$$\end{document}) but also for censored observations (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}=0$$\end{document}) that contribute while present in the risk set; this is critical for accurately reflecting the covariance structure of the Fisher information matrix. This sum is computed simultaneously for the entire sample using vectorized cumulative sum operators rather than loop-based procedures. Using the obtained vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}$$\end{document}, and following the Hessian decomposition of the Cox partial likelihood [2, 14], the observed Fisher information matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}\left({\beta\:}\right)=-{H}\left({\beta\:}\right)$$\end{document} can be approximately expressed in the form of a variance-weighted design matrix, thanks to the linear model structure after RFF transformation (see Appendix A for the full derivation): 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}\left({\beta\:}\right)={{\Phi\:}}^{{\top\:}}\mathrm{diag}\left({\mu\:}\right){\Phi\:}+\lambda\:{{I}}_{m}$$\end{document} This constitutes a Breslow-type approximation obtained by omitting the positive semi-definite rank-1 correction terms from the exact Hessian (Appendix A, Eq. A5–A6). Since the omitted terms are positive semi-definite, this approximation overestimates the Fisher information matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{I}}_{approx}\geq{{I}}_{exact}$$\end{document}) which may lead to marginally anti-conservative confidence intervals. To facilitate efficient inversion and uncertainty quantification, the eigendecomposition of the regularized information matrix is computed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}\left({\beta\:}\right)={V}{\Lambda\:}{{V}}^{{\top\:}}$$\end{document} where V is the orthogonal matrix of eigenvectors and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Lambda\:}=\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left({\lambda\:}_{1},\dots\:,{\lambda\:}_{m}\right)$$\end{document} contains the corresponding positive eigenvalues. In numerical implementation, to avoid constructing an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\times\:n$$\end{document} diagonal matrix, the expression \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}^{\top\:}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left({\mu\:}\right)$$\end{document} is evaluated by first constructing the row-weighted matrix whose i-th row is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{i}\cdot{\varphi\:\left({x}_{i}\right)}^{{\top\:}}$$\end{document} and then computing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}^{\top\:}$$\end{document} times this weighted matrix, thereby reducing memory usage. Here, Φ represents the design matrix. Since the symmetry of the matrix may be compromised due to numerical rounding errors, the matrix to be used in the Newton step is symmetrized as: 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{I}}_{\mathrm{sym}}=\frac{1}{2}\left({I}+{{I}}^{{\top\:}}\right)$$\end{document} Using this matrix, the Newton–Raphson update is performed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\beta\:}}^{\left(k+1\right)}={{\beta\:}}^{\left(k\right)}+{{I}}_{\mathrm{sym}}^{-1}{U}\left({{\beta\:}}^{\left(k\right)}\right)$$\end{document} This stabilization aids the reliable convergence of optimization in the high-dimensional RFF space. According to maximum likelihood theory, the inverse of the Fisher information matrix at the point of convergence provides the approximate covariance matrix of the parameter estimates. Thus, 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Var}\left({\beta\:}\right)={{I}}_{\mathrm{sym}}^{-1}$$\end{document} is obtained. Using this covariance matrix, both standard errors of parameter estimates and confidence intervals for linear predictors can be calculated. Consequently, the model not only produces point estimates but also enables the quantitative assessment of statistical uncertainty for individual risk predictions. Automatic bandwidth selection via laplace approximation The performance of the Radial Basis Function (RBF) kernel depends significantly on the bandwidth parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}. In censored survival data, traditional cross-validation methods incur high computational costs. In this study, a Laplace approximation based on penalized partial likelihood is adopted for selecting the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upsigma\:}$$\end{document} parameter. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{D}=\{\left({x}_{i},\hspace{0.17em}{t}_{i},\hspace{0.17em}{E}_{i}\right){\}}_{i=1}^{n}$$\end{document} denote the observed survival dataset, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{x}}_{{i}}\in\:{R}^{d}$$\end{document} is the covariate vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{t}_{i}>0$$\end{document} the observed time, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} the event indicator for subject i. Under a Gaussian prior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}\sim\:\mathcal{N}\left(0,\hspace{0.17em}{{\uplambda\:}}^{-1}{{I}}_{m}\right)$$\end{document} the penalized log-likelihood\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{p}\left({\beta\:}\right)=l\left({\beta\:}\right)-\frac{{\uplambda\:}}{2}\left|\right|{\beta\:}{\left|\right|}_{2}^{2}$$\end{document} (Eq. 10) corresponds to the log-posterior density up to an additive constant. Since the marginal likelihood of the model, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\left(D\mid{\upsigma\:}\right)=\int\:\mathrm{exp}\left({l}_{p}\left({\beta\:}\right)\right)d{\beta\:}$$\end{document}, cannot be solved in closed form, it is approximated using a second-order Taylor expansion around the MAP estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}$$\end{document} [20]: 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{log}p\left(D\mid{\upsigma\:}\right)\approx\:{l}_{\mathrm{p}}\left(\widehat{{{\beta\:}}_{{\upsigma\:}}}\right)-\frac{1}{2}\mathrm{log}\mathrm{det}{I}\left(\widehat{{{\beta\:}}_{{\upsigma\:}}}\right)+C$$\end{document} This expression can be interpreted as an evidence maximization criterion, where the bandwidth parameter is selected to maximize the approximate marginal likelihood. The subscript \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\beta\:}_{{\upsigma\:}}}$$\end{document} indicates that the MAP estimate depends on the bandwidth parameter through the RFF feature map. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{l}_{\mathrm{p}}$$\end{document} denotes the penalized log-likelihood, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I\left(\cdot\:\right)$$\end{document} is the Fisher information matrix, and C is a constant term. The determinant term (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{log}\mathrm{det}{I}$$\end{document}) acts as a regularization component penalizing model complexity; thus, the resulting score serves as a model selection criterion balancing data fit and model simplicity. For numerical stability, the determinant is calculated via eigen-decomposition [21] (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}={V}{\Lambda\:}{{V}}^{{\top\:}}$$\end{document}) where V is the matrix of eigenvectors and {\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\uplambda\:}}_{j}$$\end{document}} are the positive eigenvalues of I, as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:\mathrm{log}{{\uplambda\:}}_{j}$$\end{document} rather than by direct computation. The optimal bandwidth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\upsigma\:}}^{*}$$\end{document} is determined via numerical optimization over a bounded range in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:log-{\upsigma\:}$$\end{document} space to maximize this score, ensuring positive-valued and scale-invariant exploration of the bandwidth parameter. To efficiently scan the search space, the median heuristic is employed, consistent with the suggestion by Schölkopf and Smola [6]. By calculating the median of pairwise Euclidean distances between data points (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\upsigma\:}}_{\mathrm{med}}$$\end{document}), the search range is restricted to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[0.1{{\upsigma\:}}_{\mathrm{med}},10{{\upsigma\:}}_{\mathrm{med}}\right]$$\end{document}. This constraint prevents the evaluation of meaningless extreme values and accelerates convergence. To further reduce computational load, random subsampling (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}^{{\prime\:}}=min\left(n,300\right)$$\end{document}) is applied for large-scale datasets (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\:>\:300$$\end{document}), balancing computational efficiency with adequate representation of the censoring distribution. During the optimization process, the Newton–Raphson iteration for each candidate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} value is limited to only 5 steps; the goal is not full convergence but rapid comparison of relative scores among σ candidates, an early stopping strategy justified because the purpose is relative ranking rather than full convergence at each candidate. Empirical evaluations confirming the robustness of both approximations are presented in Appendix B (Table S1). Baseline hazard and survival function estimation Once the model coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\beta\:}}$$\end{document} are obtained, the non-parametric baseline cumulative hazard function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{H}_{0}}\left(t\right)$$\end{document} is calculated using the Breslow estimator [15]. In the presence of tied events, the Breslow approximation is applied: 17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{H}_{0}}\left(t\right)={\sum\:}_{i:{T}_{i}\le\:t,{E}_{i}=1}\frac{{d}_{i}}{{\sum\:}_{j\in\:R\left({T}_{i}\right)}\mathrm{exp}\left(\phi\:{\left({{x}}_{{j}}\right)}^{{\top\:}}\widehat{{\beta\:}}\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}$$\end{document} denotes the number of events occurring at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R\left({T}_{i}\right)$$\end{document} represents the risk set at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i}$$\end{document}. The individual survival function for a new covariate vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}$$\end{document} is then derived as: 18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{S}\left(t\mid{x}\right)=\mathrm{exp}\left(-\widehat{{H}_{0}}\left(t\right)\mathrm{exp}\left(\phi\:{\left({x}\right)}^{{\top\:}}\widehat{{\beta\:}}\right)\right)$$\end{document} To facilitate downstream uncertainty quantification, the variance of the baseline cumulative hazard estimator is additionally required. To account for tied events and ensure numerical stability in finite samples, a Greenwood-type variance estimator is employed for the baseline cumulative hazard. The variance of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{H}_{0}}\left(t\right)$$\end{document} is computed as: 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Var}}\left(\widehat{{H}_{0}}\left(t\right)\right)=\sum\:_{i:{T}_{i}\le\:t,{E}_{i}=1}\frac{{d}_{i}}{R\left({T}_{i}\right)\left(R\right({T}_{i})-{d}_{i})}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}$$\end{document} is the number of events at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R\left({T}_{i}\right)$$\end{document} is the size of the risk set. This formulation reduces to the standard Nelson-Aalen [22, 23] variance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({d}_{i}/R{\left({T}_{i}\right)}^{2}\right)$$\end{document} when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}\ll\:R\left({T}_{i}\right)$$\end{document} but provides better coverage properties when ties are present or risk sets are small. This estimator is consistent with the large-sample counting-process theory established by Andersen and Gill (1982) [13] and is standard in the survival-analysis literature [14]. In the present implementation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Var}}\left(\widehat{{H}_{0}}\left(t\right)\right)$$\end{document} is computed at each unique event time and linearly interpolated for arbitrary prediction horizons. To maintain interpretability in point estimates and ensure compatibility with standard practices in the literature, the linear predictor is used directly in the form \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{exp}\left(\eta\:\right)$$\end{document} without applying any moment corrections. Uncertainty quantification and confidence intervals Reliable deployment of the proposed RFF-Cox framework in clinical decision support systems necessitates the systematic quantification of uncertainty for individual risk predictions. Observed information matrix and stabilized covariance estimation Maximum likelihood estimators obtained under the Cox partial likelihood function are asymptotically normally distributed in large samples [1, 2]. Consequently, the covariance matrix of the parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}$$\end{document} is approximated by the inverse of the observed information matrix, derived from the second derivatives of the Cox partial log-likelihood function [14]: 20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Cov}}\left(\widehat{{\beta\:}}\right)={I}{\left(\widehat{{\beta\:}}\right)}^{-1}$$\end{document} When Random Fourier Features (RFF) are employed, the model reduces to a linear parameter structure, and the Fisher information matrix takes the variance-weighted form given in Eq. 13. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}^{\top\:}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left({\mu\:}\right){\Phi\:}$$\end{document} is positive semi-definite and the ridge penalty \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:{{I}}_{m}$$\end{document} contributes strictly positive eigenvalues, positive definiteness is guaranteed for any \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:>0$$\end{document}. The covariance matrix is therefore obtained directly as: 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\Sigma\:}}={{V}{\Lambda\:}}^{-1}{{V}}^{{\top\:}}$$\end{document} Unlike diagonal approximations, this study utilizes the full covariance structure, which preserves correlations between parameters, providing a more robust foundation for calculating Wald statistics. Finally, the variance of the linear predictor for a new observation is calculated using: 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Var}\left(\eta\:\left({x}\right)\right)={\upphi\:}{\left({x}\right)}^{{\top\:}}\widehat{{\Sigma\:}}{\upphi\:}\left({x}\right)$$\end{document} Since this expression accounts for the entire covariance structure among parameters, it directly enables the calculation of confidence intervals for individual risk predictions. Variance propagation via the multivariate Delta method The survival function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{S}\left(t\mid{x}\right)$$\end{document} depends on two quantities that are estimated with sampling uncertainty: the linear predictor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left({x}\right)=\phi\:{\left({x}\right)}^{{\top\:}}\widehat{{\beta\:}}$$\end{document} and the baseline cumulative hazard \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{H}_{0}}\left(t\right)$$\end{document}. To obtain a comprehensive variance estimate, the multivariate Delta method ([12, 24]) is employed, propagating variability from both sources simultaneously: 23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Var}}\left(\widehat{S}\right)\approx\:{\left(\frac{\partial\:S}{\partial\:\eta\:}\right)}^{2}\cdot\:\hspace{0.17em}\widehat{\mathrm{Var}}\left(\eta\:\right)+{\left(\frac{\partial\:S}{\partial\:{H}_{0}}\right)}^{2}\cdot\:\hspace{0.17em}\widehat{\mathrm{Var}}\left(\widehat{{H}_{0}}\right)$$\end{document} The cross-term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Cov(\eta\:,\:H_0)$$\end{document} is omitted under the standard working assumption that parameter estimation uncertainty and baseline hazard variability are approximately independent for moderate sample sizes [14]. While both quantities are estimated from the same data, this independence approximation is commonly adopted in the survival-analysis literature and becomes increasingly accurate as the sample size grows. Where the partial derivatives, evaluated at the estimated quantities, are: 24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{\partial\:S}{\partial\:\eta\:}=-\widehat{{H}_{0}}\left(t\right)\hspace{0.17em}\mathrm{exp}\left(\eta\:\right)\hspace{0.17em}\widehat{S},\hspace{1em}\hspace{1em}\frac{\partial\:S}{\partial\:{H}_{0}}=-\mathrm{exp}\left(\eta\:\right)\hspace{0.17em}\widehat{S}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Var}}\left(\eta\:\right)=\phi\:{\left({x}\right)}^{{\top\:}}\widehat{{\Sigma\:}}\hspace{0.17em}\phi\:\left({x}\right)$$\end{document} is the variance of the linear predictor derived in Sect. Observed information matrix and stabilized covariance estimation, (Eq. 22), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Var}}\left(\widehat{{H}_{0}}\right)$$\end{document} is the Greenwood-type variance estimator defined in Eq. 19. The first term captures parameter estimation uncertainty propagated through the log-hazard structure, whereas the second term reflects the inherent variability of the non-parametric baseline hazard estimate. Inclusion of the baseline hazard variance yields more conservative—and, in small risk sets, more realistic—confidence intervals than those obtained from the single-source approximation that accounts only for parameter uncertainty. The survival function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S\left(t\mid{x}\right)$$\end{document} is naturally defined within the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\mathrm{0,1}\right)$$\end{document} range and exhibits a distinctly asymmetric distribution. Therefore, a confidence interval based on the log-log transformation is used, which inherently preserves the asymmetric structure and prevents estimates from exceeding probability bounds [25]. The transformation is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:u=\mathrm{ln}\left(-\mathrm{ln}\left(\widehat{S}\right)\right)$$\end{document}. Using the Delta method, the standard error \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\updelta\:}$$\end{document} of this transformation is derived as follows. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:u=\mathrm{ln}\left(-\mathrm{ln}\left(S\right)\right)$$\end{document}, the derivative with respect to S is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{du}{dS}=\frac{1}{S\cdot\:\mathrm{ln}\left(S\right)}$$\end{document} Applying the Delta method, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Var}\left(u\right)\approx\:{\left(\frac{du}{dS}\right)}^{2}\cdot\:\mathrm{Var}\left(S\right)$$\end{document}, yielding: 25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:=\frac{\sqrt{\widehat{\mathrm{Var}}\left(\widehat{S}\right)}}{\widehat{S}\left|\mathrm{ln}\left(\widehat{S}\right)\right|+\epsilon}$$\end{document} The absolute value ensures a positive standard error, as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:du/d\widehat{S}\:<\:0$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S\in\:\left(\mathrm{0,1}\right)$$\end{document}. The additive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document} prevents division by zero as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{S}\:\to\:\:1$$\end{document}. Consequently, the 95% confidence intervals are constructed as: 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[{S}_{lo},\hspace{0.17em}{S}_{up}\right]=\left[\mathrm{e}\mathrm{x}\mathrm{p}\left(-\mathrm{exp}\left(u+1.96\hspace{0.17em}\delta\:\right)\right),\mathrm{e}\mathrm{x}\mathrm{p}\left(-\mathrm{exp}\left(u-1.96\hspace{0.17em}\delta\:\right)\right)\right]$$\end{document} For the Hazard Ratio (HR), the standard Wald approach (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{exp}\left(\eta\:\pm\:1.96\cdot\:\mathrm{SE}\right)$$\end{document}) is adopted. Clinical interpretation: equivalent coefficients and interaction inference Although the RFF-Cox model operates in an m-dimensional random feature space, clinical interpretation frequently requires covariate-level effect estimates expressed in the original d-dimensional input space. This section presents two complementary procedures that bridge this representational gap: equivalent linear coefficients for main effect inference (Sect. Equivalent linear coefficients and bias correction) and second-order partial derivatives for interaction detection (Sect. Inference on second-order effects). Equivalent linear coefficients and bias correction To enable covariate-level interpretation, a first-order Taylor expansion of the feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi\:$$\end{document}(x) is employed around the training sample mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{{x}}$$\end{document}[11]: 27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left({x}\right)=\phi\:{\left({x}\right)}^{{\top\:}}\widehat{{\beta\:}}\approx\:\phi\:{\left(\stackrel{-}{{x}}\right)}^{{\top\:}}\widehat{{\beta\:}}+{\left({x}-\stackrel{-}{{x}}\right)}^{{\top\:}}{{J}}^{{\top\:}}\widehat{{\beta\:}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}\in\:{R}^{m\times\:d}$$\end{document} denotes the Jacobian matrix of the feature map evaluated at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{{x}}$$\end{document}. For the RBF-based Random Fourier Feature mapping defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\phi\:}_{j}\left({x}\right)=\sqrt{2/m}\mathrm{cos}\left({{\upomega\:}}_{j}^{{\top\:}}{x}+{b}_{j}\right)$$\end{document}, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(j,k\right)$$\end{document}-th element of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}$$\end{document} is given by: 28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{J}}_{jk}={\left.\frac{\partial\:{{\upphi\:}}_{j}}{\partial\:{x}_{k}}\right|}_{{x}=\stackrel{-}{{x}}}=-\sqrt{\frac{2}{m}}\mathrm{sin}\left({{\upomega\:}}_{j}^{{\top\:}}\stackrel{-}{{x}}+{b}_{j}\right)\hspace{0.17em}{{\upomega\:}}_{jk}$$\end{document} By defining the equivalent linear coefficient vector as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\beta\:}}_{\mathrm{equiv}}\:={{J}}^{{\top\:}}\widehat{{\beta\:}}\in\:{R}^{d}$$\end{document}, the linear predictor can be re-expressed as: 29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left({x}\right)\approx\:{{x}}^{{\top\:}}{\beta\:}_{\mathrm{equiv}}\:+{b}_{\mathrm{bias}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{\mathrm{bias}}=\phi\:{\left(\stackrel{-}{{x}}\right)}^{{\top\:}}\widehat{{\beta\:}}-{\stackrel{-}{{x}}}^{{\top\:}}{{\beta\:}}_{\mathrm{equiv}}\:$$\end{document} represents the residual intercept arising from the first-order approximation. This decomposition preserves the total linear predictor while enabling the interpretation of each covariate’s marginal effect through the corresponding element of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\beta\:}}_{\mathrm{equiv}}$$\end{document}, analogous to the coefficient vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}$$\end{document} in the classical Cox model [1, 2]. Consequently, the bias-corrected hazard ratio for the j-th covariate is computed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{exp}\left({\beta\:}_{\mathrm{equiv},j}\right)$$\end{document}. For uncertainty quantification of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\beta\:}}_{\mathrm{equiv}}\:$$\end{document}, the covariance matrix is derived via the standard Jacobian variance-propagation formula [12]: 30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\mathrm{Cov}}{({\beta\:}}_{\mathrm{equiv}})={{J}}^{{\top\:}}\widehat{{\Sigma\:}}{J}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\Sigma\:}}\in\:{R}^{m\times\:m}$$\end{document} is the stabilized covariance matrix obtained in Eq. 21. Standard errors, Wald-type Z-statistics, and two-sided p-values for each equivalent coefficient are subsequently computed from the diagonal elements of this \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d\:\times\:d$$\end{document} matrix. This procedure facilitates formal hypothesis testing of individual covariate effects while fully accounting for the correlation structure among the parameters in the RFF domain. Inference on second-order effects To test for covariate interactions, the cross-partial derivative \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\partial\:}^{2}{\upeta\:}/\partial\:{x}_{j}\partial\:{x}_{k}$$\end{document} is estimated using a centered finite-difference stencil applied to the model’s linear predictor: 31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{\partial\:}^{2}\widehat{\eta\:}}{\partial\:{x}_{j}\partial\:{x}_{k}}=\frac{\widehat{\eta\:}\left({x}+{{\upepsilon\:}}_{j}+{{\upepsilon\:}}_{k}\right)-\widehat{\eta\:}\left({x}+{{\upepsilon\:}}_{j}-{{\upepsilon\:}}_{k}\right)-\widehat{\eta\:}\left({x}-{{\upepsilon\:}}_{j}+{{\upepsilon\:}}_{k}\right)+\widehat{\eta\:}\left({x}-{{\upepsilon\:}}_{j}-{{\upepsilon\:}}_{k}\right)}{4{{\upepsilon\:}}^{2}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\upepsilon\:}}_{j}\in\:{R}^{d}$$\end{document} denotes a perturbation vector with value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document} in the j-th position and zeros elsewhere. All interaction derivatives are evaluated at the covariate mean unless otherwise stated. These derivatives quantify local second-order interaction behavior around the evaluation point. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{\eta\:}\left({x}\right)=\phi\:{\left({x}\right)}^{{\top\:}}\widehat{{\beta\:}}$$\end{document} is linear in the model parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\beta\:}}$$\end{document}, the delta method (Sect. Variance propagation via the multivariate Delta method) applies directly. The Jacobian of the cross-partial derivative with respect to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\beta\:}}$$\end{document} is given by: 32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{J}}_{jk}=\frac{{\upphi\:}\left({x}+{{\upepsilon\:}}_{j}+{{\upepsilon\:}}_{k}\right)-{\upphi\:}\left({x}+{{\upepsilon\:}}_{j}-{{\upepsilon\:}}_{k}\right)-{\upphi\:}\left({x}-{{\upepsilon\:}}_{j}+{{\upepsilon\:}}_{k}\right)+{\upphi\:}\left({x}-{{\upepsilon\:}}_{j}-{{\upepsilon\:}}_{k}\right)}{4{{\upepsilon\:}}^{2}}$$\end{document} The variance of the interaction estimate is then computed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Var}\left({\partial\:}^{2}\widehat{{\upeta\:}}/\partial\:{x}_{j}\partial\:{x}_{k}\right)={{J}}_{jk}^{{\top\:}}\widehat{{\Sigma\:}}{{J}}_{jk},$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{{\Sigma\:}}$$\end{document} is the stabilized covariance matrix from Eq. 21. A Wald statistic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Z=\left({\partial\:}^{2}\widehat{{\upeta\:}}/\partial\:{x}_{j}\partial\:{x}_{k}\right)/\mathrm{SE}$$\end{document} is subsequently calculated to formally test the significance of the interaction effect. When evaluating all pairwise interactions among d covariates, Bonferroni correction is applied to control the family-wise error rate at the nominal level. Algorithmic summary The integrated operational workflow of the proposed RFF-Cox framework, synthesizing the theoretical components detailed above—from RFF-based feature mapping and penalized Newton–Raphson optimization to Laplace-driven hyperparameter tuning and stabilized uncertainty quantification—is systematically outlined in Algorithm 1. This procedural summary consolidates the scalable estimation steps required for robust survival analysis in large-scale cohorts. Algorithm 1 is provided in Appendix D. Empirical evaluation and performance validation In this section, the performance of the proposed RFF-Cox model is evaluated across six datasets (two simulated and four real-world). The analyses focus on four key criteria: (i) the ability to capture non-linear risk relationships and covariate interactions, (ii) empirical coverage of analytical confidence intervals, (iii) computational cost, and (iv) calibration of prediction uncertainty. Experimental setup and data characteristics To comprehensively assess the model’s performance across diverse data structures, we utilized a total of six datasets: two simulated datasets designed to evaluate specific modeling capabilities, and four real-world datasets representing fields such as hematology, intensive care, cardiovascular diseases, and oncology. These include the FLCHAIN hematology cohort provided by Therneau [26] (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n\:=\mathrm{7,874}$$\end{document}), the SUPPORT2 intensive care data developed by Knaus et al. [27] (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n=\mathrm{9,105}$$\end{document}), the WHAS cardiovascular data reported by Hosmer and Lemeshow [28] ( n=500), and the METABRIC breast cancer genomic cohort published by Curtis et al. [29] (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n=\mathrm{1,904}$$\end{document}). Across these six datasets, sample sizes range from 500 to 9,105, the number of covariates varies between 6 and 52, and event rates span from 27.5% to 77.6%. This diversity allows for rigorous testing of the model’s performance under varying data scales and censoring levels. The first simulated dataset (non-linear; n = 2,000, d = 6) was generated to evaluate the model’s capacity to capture non-linear risk structures in a controlled environment. Covariates were generated from specific distributions and standardized: Age (Uniform), BMI and Blood Pressure (Normal), Smoking Status (Gamma), LDL Cholesterol (Normal), and CRP (Log-Normal). The log-normal distribution was chosen for CRP as right-skewed biomarker distributions are frequently modeled using log-normal frameworks in the literature [30]. The true log-hazard function was defined as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}\:\eta\:\left(x\right)&=1.3\mathrm{tanh}\left(\mathrm{Age}\right)+0.9{\mathrm{BMI}}^{2}-1.1\left(\mathrm{SBP}\times\:\mathrm{Smoking}\right)\\&+0.5\mathrm{LDL}+0.7\mathrm{log}\left(1+{\mathrm{CRP}}^{2}\right)\end{aligned}$$\end{document} This structure incorporates three types of non-linear effects: (i) saturation (tanh), (ii) U-shaped relationships (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}^{2}$$\end{document}), and (iii) covariate interactions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{1\:}{x}_{2}$$\end{document}). Event times were generated from a Weibull-based process (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\uplambda}}_{0}=0.008$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:=\:1.6$$\end{document}). Censoring times were drawn from an exponential distribution independent of the event process, yielding an approximate censoring rate of 77.6%. The second simulated dataset (linear; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n=\mathrm{2,000},\:d=6$$\end{document}) was constructed to serve as a controlled baseline in which the proportional hazards assumption holds and the true log-hazard function takes a strictly linear form: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left(x\right)=0.5{x}_{1}-0.8{x}_{2}+0.6{x}_{3}+0.4{x}_{4}-0.5{x}_{5}+1.2{x}_{6}$$\end{document} Event times were generated from a Weibull-based process (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda_{0}=0.01,\:\rho\:=1.5$$\end{document}), with an approximate event rate of 45.0%. This dataset evaluates whether the RFF approximation introduces performance penalty when the true relationship is linear. Five methods were evaluated under identical training–test splits and 5-fold cross-validation: RFF-Cox, CoxPH, Random Survival Forest (RSF), XGBoost, and DeepSurv. DeepSurv is a deep learning-based survival model that extends the Cox framework with deep neural networks [9]. XGBoost is a scalable gradient boosting framework widely adopted for survival analysis [31]. For the RFF-Cox model, the number of random features was set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m=\lceil\:\sqrt{n}\rceil\:$$\end{document} and the regularization parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} was selected from the grid {\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1{0}^{-3},1{0}^{-2},1{0}^{-1},1{0}^{0}$$\end{document}}. For competing methods, hyperparameters were optimized via grid search: RSF (trees: {50, 100}, max depth: {5, 10, None}, min samples per leaf: {5, 10}); XGBoost (trees: {100, 200}, learning rate: {0.01, 0.05, 0.1}, max depth: {3, 5}); DeepSurv (hidden layers: {[64, 32], [64, 32, 16], [128, 64]}, learning rate: {\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1{0}^{-4},1{0}^{-3}$$\end{document}}, dropout: {0.1, 0.2}). Performance metrics included Harrell’s C-index and the integrated time-dependent AUC (iAUC). Table 1 summarizes the dataset characteristics and the automatically selected RFF-Cox hyperparameters for each dataset. Table 1Dataset characteristics and RFF-Cox hyperparameter optimization resultsDataset n dEvent (%)mλ (L2)Log-Risk SE (Mean ± SD)RFF Bandwidth σ (Mean ± SD)Simulated2,000677.64010⁻³0.16 ± 0.017.62 ± 2.15Linear sim2,000645.04010⁻³0.13 ± 0.0030.81 ± 0.54SUPPORT29,1055268.18510⁻³0.12 ± 0.0056.16 ± 3.88FLCHAIN7,874827.57910⁰0.08 ± 0.006.91 ± 0.12WHAS500500643.02010⁻³0.22 ± 0.0228.57 ± 0.18METABRIC1,904957.93910⁻³0.13 ± 0.0029.39 ± 2.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:n$$\end{document} denotes the sample size; d, the number of covariates; m, the dimension of Random Fourier Features; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}, the ridge regularization coefficient. Log-Risk SE represents the mean standard error of the log-risk (linear predictor) estimates calculated across the test samples. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} indicates the RFF kernel bandwidth automatically selected via the Laplace approximation Table 1Dataset characteristics and RFF-Cox hyperparameter optimization resultsDataset n dEvent (%)mλ (L2)Log-Risk SE (Mean ± SD)RFF Bandwidth σ (Mean ± SD)Simulated2,000677.64010⁻³0.16 ± 0.017.62 ± 2.15Linear sim2,000645.04010⁻³0.13 ± 0.0030.81 ± 0.54SUPPORT29,1055268.18510⁻³0.12 ± 0.0056.16 ± 3.88FLCHAIN7,874827.57910⁰0.08 ± 0.006.91 ± 0.12WHAS500500643.02010⁻³0.22 ± 0.0228.57 ± 0.18METABRIC1,904957.93910⁻³0.13 ± 0.0029.39 ± 2.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:n$$\end{document} denotes the sample size; d, the number of covariates; m, the dimension of Random Fourier Features; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}, the ridge regularization coefficient. Log-Risk SE represents the mean standard error of the log-risk (linear predictor) estimates calculated across the test samples. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} indicates the RFF kernel bandwidth automatically selected via the Laplace approximation Dataset characteristics and RFF-Cox hyperparameter optimization results \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:n$$\end{document} denotes the sample size; d, the number of covariates; m, the dimension of Random Fourier Features; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}, the ridge regularization coefficient. Log-Risk SE represents the mean standard error of the log-risk (linear predictor) estimates calculated across the test samples. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} indicates the RFF kernel bandwidth automatically selected via the Laplace approximation Model discrimination and computational efficiency The discriminatory performance of the models was evaluated using 5-fold cross-validation with Harrell’s C-index and time-dependent integrated AUC (iAUC) metrics [32, 33]. All C-index values are reported as mean ± SD across folds. To mitigate the potential bias of high censoring rates on the C-statistic, variants based on Inverse Probability of Censoring Weighting (IPCW) have been proposed in the literature [34]; adopting this approach ensures methodological consistency with the calibration analyses performed in this study. Table 2 presents the performance values of the models across datasets along with their mean training times. In the synthetic dataset containing non-linear components, the RFF-Cox model achieved the highest iAUC (0.92) and a C-index of 0.83, jointly with DeepSurv (iAUC = 0.92). In contrast, the CoxPH model, which relies on the linearity assumption, yielded lower performance on the same data (C-index = 0.69). These results confirm that the RFF-Cox model offers a distinct advantage in capturing non-linear risk structures. Table 2Discriminatory performance and mean training times (seconds) of the modelsDatasetModelC-indexiAUCFit Time (s)SimulatedRFF-Cox0.83 ± 0.010.92 ± 0.010.12 ± 0.02CoxPH0.69 ± 0.020.77 ± 0.020.17 ± 0.02RSF0.79 ± 0.020.88 ± 0.0214.25 ± 0.38XGBoost0.81 ± 0.020.90 ± 0.020.34 ± 0.01DeepSurv0.83 ± 0.010.92 ± 0.012.09 ± 0.05Linear SimRFF-Cox0.82 ± 0.010.89 ± 0.010.16 ± 0.01CoxPH0.82 ± 0.010.89 ± 0.020.14 ± 0.01RSF0.81 ± 0.020.87 ± 0.0211.78 ± 0.10XGBoost0.81 ± 0.010.88 ± 0.020.22 ± 0.01DeepSurv0.82 ± 0.010.88 ± 0.021.67 ± 0.02SUPPORT2RFF-Cox0.81 ± 0.010.86 ± 0.011.05 ± 0.13CoxPH0.82 ± 0.010.87 ± 0.010.79 ± 0.02RSF0.80 ± 0.010.85 ± 0.0123.19 ± 0.97XGBoost0.80 ± 0.090.86 ± 0.080.28 ± 0.02DeepSurv0.85 ± 0.010.89 ± 0.0111.88 ± 0.32FLCHAINRFF-Cox0.80 ± 0.010.82 ± 0.030.74 ± 0.11CoxPH0.80 ± 0.010.82 ± 0.030.76 ± 0.10RSF0.79 ± 0.010.82 ± 0.0322.68 ± 0.19XGBoost0.80 ± 0.010.82 ± 0.030.14 ± 0.01DeepSurv0.80 ± 0.010.82 ± 0.038.21 ± 0.19WHAS500RFF-Cox0.75 ± 0.020.77 ± 0.030.05 ± 0.00CoxPH0.75 ± 0.020.77 ± 0.030.03 ± 0.00RSF0.75 ± 0.020.78 ± 0.030.54 ± 0.00XGBoost0.75 ± 0.010.78 ± 0.040.05 ± 0.00DeepSurv0.75 ± 0.020.78 ± 0.010.61 ± 0.01METABRICRFF-Cox0.64 ± 0.010.69 ± 0.010.15 ± 0.03CoxPH0.63 ± 0.010.68 ± 0.020.13 ± 0.00RSF0.65 ± 0.020.70 ± 0.015.14 ± 0.11XGBoost0.65 ± 0.010.69 ± 0.010.23 ± 0.01DeepSurv0.64 ± 0.010.68 ± 0.011.69 ± 0.07± values represent standard deviation across 5-fold cross-validation runs Table 2Discriminatory performance and mean training times (seconds) of the modelsDatasetModelC-indexiAUCFit Time (s)SimulatedRFF-Cox0.83 ± 0.010.92 ± 0.010.12 ± 0.02CoxPH0.69 ± 0.020.77 ± 0.020.17 ± 0.02RSF0.79 ± 0.020.88 ± 0.0214.25 ± 0.38XGBoost0.81 ± 0.020.90 ± 0.020.34 ± 0.01DeepSurv0.83 ± 0.010.92 ± 0.012.09 ± 0.05Linear SimRFF-Cox0.82 ± 0.010.89 ± 0.010.16 ± 0.01CoxPH0.82 ± 0.010.89 ± 0.020.14 ± 0.01RSF0.81 ± 0.020.87 ± 0.0211.78 ± 0.10XGBoost0.81 ± 0.010.88 ± 0.020.22 ± 0.01DeepSurv0.82 ± 0.010.88 ± 0.021.67 ± 0.02SUPPORT2RFF-Cox0.81 ± 0.010.86 ± 0.011.05 ± 0.13CoxPH0.82 ± 0.010.87 ± 0.010.79 ± 0.02RSF0.80 ± 0.010.85 ± 0.0123.19 ± 0.97XGBoost0.80 ± 0.090.86 ± 0.080.28 ± 0.02DeepSurv0.85 ± 0.010.89 ± 0.0111.88 ± 0.32FLCHAINRFF-Cox0.80 ± 0.010.82 ± 0.030.74 ± 0.11CoxPH0.80 ± 0.010.82 ± 0.030.76 ± 0.10RSF0.79 ± 0.010.82 ± 0.0322.68 ± 0.19XGBoost0.80 ± 0.010.82 ± 0.030.14 ± 0.01DeepSurv0.80 ± 0.010.82 ± 0.038.21 ± 0.19WHAS500RFF-Cox0.75 ± 0.020.77 ± 0.030.05 ± 0.00CoxPH0.75 ± 0.020.77 ± 0.030.03 ± 0.00RSF0.75 ± 0.020.78 ± 0.030.54 ± 0.00XGBoost0.75 ± 0.010.78 ± 0.040.05 ± 0.00DeepSurv0.75 ± 0.020.78 ± 0.010.61 ± 0.01METABRICRFF-Cox0.64 ± 0.010.69 ± 0.010.15 ± 0.03CoxPH0.63 ± 0.010.68 ± 0.020.13 ± 0.00RSF0.65 ± 0.020.70 ± 0.015.14 ± 0.11XGBoost0.65 ± 0.010.69 ± 0.010.23 ± 0.01DeepSurv0.64 ± 0.010.68 ± 0.011.69 ± 0.07± values represent standard deviation across 5-fold cross-validation runs Discriminatory performance and mean training times (seconds) of the models ± values represent standard deviation across 5-fold cross-validation runs An examination of Table 2 reveals that while most methods exhibit comparable discriminatory performance (C-index and iAUC) across the real-world datasets, there are substantial differences in training times. RSF and DeepSurv require significantly longer training durations, particularly for large datasets, whereas the RFF-Cox model delivers similar discriminatory power at a fraction of the computational cost. For instance, although DeepSurv achieved the highest C-index on the SUPPORT2 dataset (0.85), RFF-Cox attained a competitive C-index of 0.81 while completing training more than 11 times faster (1.05 s vs. 11.88 s); relative to RSF (23.19 s), RFF-Cox was more than 22 times faster. In the small-sample WHAS dataset, performance metrics for all methods were closely aligned, with comparably fast training times across all approaches. Non-linear relationship detection: main effects, interactions Partial dependence analysis To evaluate the method’s capacity to capture non-linear risk functions, a partial dependence analysis was conducted on the simulation dataset. The simulation design explicitly incorporated a U-shaped risk function for the BMI variable, where risk increases at both low and high BMI values while remaining lower at intermediate levels. As shown in Fig. 1, the risk curve estimated by the RFF-Cox model closely tracks the true simulation function. The predicted curve exhibits a smooth structure across the entire BMI range, displaying only limited deviations in the tail regions where data density is sparse. Fig. 1Partial dependence of the normalized estimated log-risk on standardized BMI in the simulated test dataset Partial dependence of the normalized estimated log-risk on standardized BMI in the simulated test dataset In contrast, the CoxPH model, constrained by its linearity assumption, failed to capture this U-shaped relationship, yielding an approximately linear trend. While tree-based methods (GB and RSF) captured the general trend, their prediction functions displayed a more step-like structure, characteristic of ensemble tree algorithms. Wald test for main effects Covariate-level inference is performed using equivalent linear coefficients, defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{\mathrm{equiv},j}={\left.\partial\:\widehat{\eta\:}/\partial\:{x}_{j}\right|}_{\stackrel{-}{{x}}}.\:$$\end{document}Standard errors are derived via the RFF Jacobian and the stabilized covariance matrix (Sect. Observed information matrix and stabilized covariance estimation). Table 3 presents the results for both simulation scenarios. Table 3Wald test results: (a) non-linear and (b) linear simulated datasetsCoxPHRFF-Cox(a) Non-linear scenario:Covariate β true β SEp β SE p Age1.3000.5780.031< 0.0011.0010.047< 0.001 BMI0.000−0.0040.0400.9260.0230.0480.629 SBP0.0000.0400.0300.174−0.0330.0510.516 Smoking0.000−0.0240.0310.4450.0450.0760.549 LDL0.5000.3230.029< 0.0010.5990.053< 0.001 CRP0.0000.1920.031< 0.0010.0980.0570.082(b) Linear scenario: X10.5000.5130.040< 0.0010.5150.041< 0.001 X2−0.800−0.8140.044< 0.001−0.8160.042< 0.001 X30.6000.3900.039< 0.0010.3910.039< 0.001 X40.4000.2510.038< 0.0010.2470.039< 0.001 X5−0.500−0.4980.039< 0.001−0.4900.039< 0.001 X61.2001.2750.050< 0.0011.2760.044< 0.001 Table 3Wald test results: (a) non-linear and (b) linear simulated datasetsCoxPHRFF-Cox(a) Non-linear scenario:Covariate β true β SEp β SE p Age1.3000.5780.031< 0.0011.0010.047< 0.001 BMI0.000−0.0040.0400.9260.0230.0480.629 SBP0.0000.0400.0300.174−0.0330.0510.516 Smoking0.000−0.0240.0310.4450.0450.0760.549 LDL0.5000.3230.029< 0.0010.5990.053< 0.001 CRP0.0000.1920.031< 0.0010.0980.0570.082(b) Linear scenario: X10.5000.5130.040< 0.0010.5150.041< 0.001 X2−0.800−0.8140.044< 0.001−0.8160.042< 0.001 X30.6000.3900.039< 0.0010.3910.039< 0.001 X40.4000.2510.038< 0.0010.2470.039< 0.001 X5−0.500−0.4980.039< 0.001−0.4900.039< 0.001 X61.2001.2750.050< 0.0011.2760.044< 0.001 Wald test results: (a) non-linear and (b) linear simulated datasets In the linear scenario, both methods yield virtually identical estimates (RMSE: 0.111 vs. 0.110), confirming that the first-order Taylor expansion recovers classical Cox coefficients when the true hazard is linear. In the non-linear scenario, RFF-Cox outperforms CoxPH (RMSE: 0.137 vs. 0.314). CoxPH underestimates the Age effect by 56% due to forcing a saturated non-linear trend into a linear form and produces a spurious CRP association (p < 0.001) by misinterpreting the true log(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1+CR{P}^{2}$$\end{document}) relationship. RFF-Cox corrects both errors, recovering the Age effect more accurately and identifying CRP as non-significant. Interaction detection via second-order partial derivatives Cross-partial derivatives \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\partial\:}^{2}\widehat{\eta\:}/\partial\:x_j\partial\:x_k$$\end{document} are estimated at reference points using centered finite differences, with significance assessed via the Delta method (Sect. Inference on second-order effects). Applied to the non-linear dataset (true interaction: SBP × Smoking, (coefficient − 1.1), this procedure identified only the SBP × Smoking pair as significant after Bonferroni correction across 15 pairwise tests (Table 4). The estimated coefficient (− 1.093) closely matches the true value. Robustness across the covariate space is confirmed in Table 5, where the interaction remains significant (|Z| > 9) at all reference points. Table 4Pairwise interaction tests with p < 0.10 (non-linear dataset; 15 tests total)Pair∂²η̂SEZ p SBP × Smoking−1.0930.036−30.28< 0.001BMI × LDL−0.0610.033−1.830.067Smoking × LDL0.0600.0341.760.078 Table 4Pairwise interaction tests with p < 0.10 (non-linear dataset; 15 tests total)Pair∂²η̂SEZ p SBP × Smoking−1.0930.036−30.28< 0.001BMI × LDL−0.0610.033−1.830.067Smoking × LDL0.0600.0341.760.078 Pairwise interaction tests with p < 0.10 (non-linear dataset; 15 tests total) Table 5SBP × Smoking interaction at different reference points (scaled space)ReferenceSBPSmoking∂²η̂SEZ p Low(− 1σ)−1.0−1.0−0.8990.096−9.39< 0.001Mean(x̄)0.00.0−1.0930.036−30.28< 0.001High(+ 1σ)+ 1.0+ 1.0−1.1490.065−17.81< 0.001 Table 5SBP × Smoking interaction at different reference points (scaled space)ReferenceSBPSmoking∂²η̂SEZ p Low(− 1σ)−1.0−1.0−0.8990.096−9.39< 0.001Mean(x̄)0.00.0−1.0930.036−30.28< 0.001High(+ 1σ)+ 1.0+ 1.0−1.1490.065−17.81< 0.001 SBP × Smoking interaction at different reference points (scaled space) Hyperparameter sensitivity and coverage analysis Simulation-based coverage and hyperparameter grid search To evaluate the joint impact of the number of random features ( m) and the ridge regularization coefficient (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}) on discrimination and uncertainty quantification, a grid search was conducted over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\in\:\left\{\mathrm{20,40,80,120,160}\right\}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:\in\:{\{10}^{-3},{10}^{-2},{10}^{-1},{10}^{0}\}$$\end{document} on both the non-linear and linear simulated datasets ( n = 2,000 each, standardized covariates). Kernel bandwidth (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}) was selected via 5-fold stratified cross-validation within each configuration. Discrimination was assessed via cross-validated Harrell’s C-index. The reliability of analytical confidence intervals was evaluated through empirical coverage against the known true survival function, averaged across 15 time points spanning the 5th–95th percentile of event times. Bootstrap-based coverage (B = 50 resamples) served as a non-parametric reference. In each bootstrap iteration, training observations were re-sampled; additionally, the random Fourier feature weights (ω and phases b) were re-drawn to capture the Monte Carlo variance inherent in the RFF approximation [7]. Table 6 presents selected configurations, with C-index and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} values reported as mean ± SD across 5-fold cross-validation. Figure 2 displays the complete grid search outcomes. Table 6Grid search results across\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\times\:\:\lambda\:$$\end{document}configurationsNon-LinearLinearmλσC-indexAnCovBtCovσC-indexAnCov200.0016.07 ± 2.790.779 ± 0.02318.6%73.8%32.77 ± 1.150.822 ± 0.01090.6%90.7%201.02.60 ± 0.250.756 ± 0.02216.2%52.7%4.00 ± 0.240.820 ± 0.00949.2%96.9%400.0017.37 ± 2.260.827 ± 0.01459.2%92.9%33.33 ± 0.400.822 ± 0.01093.1%87.5%400.014.59 ± 0.660.822 ± 0.01551.8%93.2%17.96 ± 2.510.822 ± 0.01095.5%92.3%800.0018.45 ± 0.680.838 ± 0.01480.4%80.2%18.40 ± 12.580.820 ± 0.01395.3%88.6%800.13.10 ± 0.270.835 ± 0.01475.9%90.2%8.08 ± 0.480.822 ± 0.01098.9%94.1%1200.015.40 ± 0.200.837 ± 0.01581.0%82.6%19.72 ± 1.000.822 ± 0.01095.6%89.9%1600.015.41 ± 0.590.838 ± 0.01486.8%77.3%19.79 ± 0.560.822 ± 0.01096.9%89.8%CoxPH——0.677———0.813——m, RFF dimension; λ, ridge regularization coefficient; σ, kernel bandwidth (geometric mean ± SD across 5 folds); C, cross-validated C-index (mean ± SD); AnCov, empirical coverage of analytical 95% CIs; BtCov, bootstrap-based coverage Table 6Grid search results across\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\times\:\:\lambda\:$$\end{document}configurationsNon-LinearLinearmλσC-indexAnCovBtCovσC-indexAnCov200.0016.07 ± 2.790.779 ± 0.02318.6%73.8%32.77 ± 1.150.822 ± 0.01090.6%90.7%201.02.60 ± 0.250.756 ± 0.02216.2%52.7%4.00 ± 0.240.820 ± 0.00949.2%96.9%400.0017.37 ± 2.260.827 ± 0.01459.2%92.9%33.33 ± 0.400.822 ± 0.01093.1%87.5%400.014.59 ± 0.660.822 ± 0.01551.8%93.2%17.96 ± 2.510.822 ± 0.01095.5%92.3%800.0018.45 ± 0.680.838 ± 0.01480.4%80.2%18.40 ± 12.580.820 ± 0.01395.3%88.6%800.13.10 ± 0.270.835 ± 0.01475.9%90.2%8.08 ± 0.480.822 ± 0.01098.9%94.1%1200.015.40 ± 0.200.837 ± 0.01581.0%82.6%19.72 ± 1.000.822 ± 0.01095.6%89.9%1600.015.41 ± 0.590.838 ± 0.01486.8%77.3%19.79 ± 0.560.822 ± 0.01096.9%89.8%CoxPH——0.677———0.813——m, RFF dimension; λ, ridge regularization coefficient; σ, kernel bandwidth (geometric mean ± SD across 5 folds); C, cross-validated C-index (mean ± SD); AnCov, empirical coverage of analytical 95% CIs; BtCov, bootstrap-based coverage Grid search results across\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\times\:\:\lambda\:$$\end{document}configurations m, RFF dimension; λ, ridge regularization coefficient; σ, kernel bandwidth (geometric mean ± SD across 5 folds); C, cross-validated C-index (mean ± SD); AnCov, empirical coverage of analytical 95% CIs; BtCov, bootstrap-based coverage Fig. 2Grid search results. Upper: non-linear; lower: linear. a, c Coverage of analytical (solid) and bootstrap (dashed) 95% CIs vs. m. b, d C-index heatmaps Grid search results. Upper: non-linear; lower: linear. a, c Coverage of analytical (solid) and bootstrap (dashed) 95% CIs vs. m. b, d C-index heatmaps Discrimination. In the non-linear scenario (Fig. 2b), m exerts a dominant effect on discriminatory performance. At m = 20, C-index remains below 0.78 regardless of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}; performance improves markedly at m ≥ 80 (C-index ≈ 0.83–0.84), with diminishing returns beyond this threshold. The regularization parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} has a comparatively modest influence, shifting C-index by at most 0.02 within each m stratum. In the linear scenario (Fig. 2d), C-index is invariant across all configurations (0.82 ± 0.01). Coverage at Optimal Configuration. The optimal configuration identified by the grid search ( m = 80, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.001 for both scenarios) was evaluated on an independent test set (n_test = 400) across 15 time points (Table 7). In the non-linear scenario, analytical coverage (81.9%) closely matches bootstrap coverage (82.8%), confirming that the RFF approximation has stabilized and the Delta method provides a reliable variance estimate (Fig. 3). The mean prediction bias is negligible (|bias| = 0.004). However, analytical intervals are on average 2.6 times wider than bootstrap intervals (AnW = 0.251 vs. BtW = 0.097), reflecting the chain-rule amplification of linear predictor variance through the survival function. In the linear scenario, analytical coverage (95.9%) approaches the nominal 95% level and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} coverage (97.8%) is near-nominal. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document}-level coverage was assessed after centering both the true and predicted values by their respective means, as the linear predictor in PH models is identifiable only up to a location shift. Table 7Coverage and CI width at optimal configurationsMetricNon-linearLinearS(t) Coverage AnCov (mean)81.9%95.9% BtCov (mean)82.8%91.1% AnW (mean)0.2510.133 BtW (mean)0.0970.074 |S(t) bias| (mean)0.0040.012η Coverage (centered) Coverage76.0%97.8% Correlation0.9880.999 RMSE0.4000.290 Mean SE0.1850.135Averaged across 15 time points. AnCov, BtCov: analytical and bootstrap coverage; AnW, BtW: corresponding mean interval widths Table 7Coverage and CI width at optimal configurationsMetricNon-linearLinearS(t) Coverage AnCov (mean)81.9%95.9% BtCov (mean)82.8%91.1% AnW (mean)0.2510.133 BtW (mean)0.0970.074 |S(t) bias| (mean)0.0040.012η Coverage (centered) Coverage76.0%97.8% Correlation0.9880.999 RMSE0.4000.290 Mean SE0.1850.135Averaged across 15 time points. AnCov, BtCov: analytical and bootstrap coverage; AnW, BtW: corresponding mean interval widths Coverage and CI width at optimal configurations Averaged across 15 time points. AnCov, BtCov: analytical and bootstrap coverage; AnW, BtW: corresponding mean interval widths Fig. 3Coverage analysis at optimal configurations Coverage analysis at optimal configurations Coverage Sensitivity to m, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}$$\end{document}, and Approximation Method: Coverage is sensitive to both m and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} (Fig. 2a and c). In the non-linear setting, coverage at m = 20 is uniformly poor (16–19%), reflecting insufficient feature space dimensionality. Strong regularization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 1.0) adversely affects coverage: at m = 80, increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} from 0.001 to 1.0 reduces coverage from 80% to 53%, as ridge shrinkage compresses the eigenvalues of the Fisher information matrix and underestimates prediction variance [35]. In the linear setting, at m ≥ 40 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} ≤ 0.1, analytical coverage consistently exceeds 95%. At low m (e.g., m = 20, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.001), bootstrap coverage substantially exceeds analytical coverage in the non-linear scenario (73.8% vs. 18.6%), because the bootstrap re-samples ω and b, incorporating the Monte Carlo variance of the RFF approximation, which scales as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({m}^{-1}{/}^{2}\right)$$\end{document} [7]. As m increases, the two approaches converge; at m = 80, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.001, they are nearly identical (80.2% vs. 80.4%). At m = 160 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.01, analytical coverage (86.8%) exceeds bootstrap coverage (77.3%), as re-sampling ω and b occasionally yields feature configurations poorly aligned with the test covariates. Hyperparameter Sensitivity Analysis (METABRIC) To investigate the impact of model hyperparameters on performance, a grid search was conducted over the RFF dimension (m) and the ridge regularization coefficient (λ) using the METABRIC breast cancer survival dataset. Model performance was evaluated using 5-fold stratified cross-validation. The left panel of Fig. 4 illustrates that the RFF dimension (m) has a pronounced effect on model performance. At lower feature dimensions ( m ≤ 30), C-index values are comparatively reduced. Performance improves as m increases, plateauing at approximately m ≥ 80 (C-index ≈ 0.64). The highest mean C-index was achieved with the configuration m = 150 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 1.0 (C = 0.645 ± 0.009). Fig. 4Heatmaps showing cross-validated C-index (left) and training time (right) across RFF dimension (m) and regularization parameter (λ) on the METABRIC dataset Heatmaps showing cross-validated C-index (left) and training time (right) across RFF dimension (m) and regularization parameter (λ) on the METABRIC dataset At larger feature dimensions (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:=\:100-200$$\end{document}), a notable decline in both performance and stability is observed under low regularization levels (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:\:=\:0.001$$\end{document}), where standard deviations increase markedly (e.g., SD = 0.026 at m = 120 vs. SD = 0.005 at m = 30). This suggests that insufficient regularization at high model capacity may lead to overfitting. The right panel of Fig. 4 demonstrates that training time increases approximately linearly with the RFF dimension. While training times range from approximately 0.05–0.09 s for small feature dimensions, they rise to approximately 0.56–0.77 s for m = 200. Overall, these results indicate that m is the primary hyperparameter governing model capacity, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} plays a regularization role, particularly at larger feature dimensions. When balancing performance and computational cost, the combination of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:$$\end{document}= 100 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.01 or 0.1 offers a well-calibrated choice for this dataset. Bandwidth adaptivity and practical guidance The automatically selected \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} values diverge systematically between scenarios (Table 6): in the non-linear setting, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} ranges from 2.6 to 8.5 (localized kernels suited for complex structures), while in the linear setting, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} values are substantially larger (4.0–33.3), yielding smoother approximations that recover near-linear relationships. Within each scenario, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\:$$\end{document}exhibits a consistent inverse relationship with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}, suggesting a compensatory mechanism between regularization and kernel localization. Collectively, these analyses suggest: (i) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:\approx\:\:\lceil\:\sqrt{n}$$\end{document}⌉ is sufficient for discrimination, but \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:2-3\times\:$$\end{document} is advisable when reliable uncertainty quantification is a priority; as coverage improves with higher m; (ii) strong regularization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:\:$$\end{document}= 1.0) should be avoided for analytical CI generation; and (iii) moderate regularization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:\:\in\:\:\{{10}^{-3},{10}^{-2}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-1}$$\end{document}}) offers the most favorable balance between discrimination and coverage. Empirical validation of the subsampling and early stopping strategies for bandwidth selection on the SUPPORT2 dataset confirms that these approximations have negligible impact on both σ selection (< 9% variation) and discriminatory performance (< 0.0001 C-index change); full results are provided in Appendix B (Table S1). Model calibration and confidence interval analysis To validate the predictive accuracy of the proposed model, the calibration—i.e., the agreement between predicted survival probabilities and observed event rates—was analyzed [36]. For this assessment on the METABRIC test set, an Inverse Probability of Censoring Weighting (IPCW) weighted logistic regression approach was adopted to account for the influence of censoring [37]. Calibration was evaluated at 14 time horizons (12–240 months) for both RFF-Cox and CoxPH [38], enabling a direct comparison of how non-linear modelling affects predictive calibration. Calibration Slope and Integrated Calibration Error (ICI) metrics were employed to quantify systematic bias and overall model fit [39]. These results are summarized in Table 8. Table 8IPCW calibration metrics across time horizons (METABRIC)Time (Months)RFF-Cox SlopeCoxPH SlopeRFF-Cox ICICoxPH ICIRFFCox CI Widthn at risk120.9410.5160.00050.00310.0122559240.6410.4310.01210.01470.0433529360.9690.7620.00440.01110.0722492480.8030.6220.01530.02480.1048461600.7640.6190.02050.02980.1251429720.7900.6720.02030.02880.1388399840.7990.6850.02260.03170.1515367960.7410.6650.03120.03710.16583321080.8390.7980.02760.03240.17582981200.9300.8820.02240.02700.18542641500.9350.8700.01060.01900.19952011801.1971.1840.02520.02740.20591422101.1611.2040.02670.03300.2059912401.3311.3210.04010.03970.196945Slope values are derived from the IPCW-weighted logistic calibration model (ideal values: Intercept = 0, Slope = 1). The Integrated Calibration Index (ICI) quantifies calibration error, where lower values indicate better calibration. RFF-Cox CI Width denotes the mean width of the 95% confidence intervals calculated using the proposed analytical method. CoxPH does not provide analytical individual-level CIs under the standard implementation. n_at_risk denotes the number of subjects at risk at each time point Table 8IPCW calibration metrics across time horizons (METABRIC)Time (Months)RFF-Cox SlopeCoxPH SlopeRFF-Cox ICICoxPH ICIRFFCox CI Widthn at risk120.9410.5160.00050.00310.0122559240.6410.4310.01210.01470.0433529360.9690.7620.00440.01110.0722492480.8030.6220.01530.02480.1048461600.7640.6190.02050.02980.1251429720.7900.6720.02030.02880.1388399840.7990.6850.02260.03170.1515367960.7410.6650.03120.03710.16583321080.8390.7980.02760.03240.17582981200.9300.8820.02240.02700.18542641500.9350.8700.01060.01900.19952011801.1971.1840.02520.02740.20591422101.1611.2040.02670.03300.2059912401.3311.3210.04010.03970.196945Slope values are derived from the IPCW-weighted logistic calibration model (ideal values: Intercept = 0, Slope = 1). The Integrated Calibration Index (ICI) quantifies calibration error, where lower values indicate better calibration. RFF-Cox CI Width denotes the mean width of the 95% confidence intervals calculated using the proposed analytical method. CoxPH does not provide analytical individual-level CIs under the standard implementation. n_at_risk denotes the number of subjects at risk at each time point IPCW calibration metrics across time horizons (METABRIC) Slope values are derived from the IPCW-weighted logistic calibration model (ideal values: Intercept = 0, Slope = 1). The Integrated Calibration Index (ICI) quantifies calibration error, where lower values indicate better calibration. RFF-Cox CI Width denotes the mean width of the 95% confidence intervals calculated using the proposed analytical method. CoxPH does not provide analytical individual-level CIs under the standard implementation. n_at_risk denotes the number of subjects at risk at each time point An examination of Table 8 reveals that ICI values remain low across all time points, suggesting that the model predictions generally exhibit reasonable agreement with observed survival probabilities. Over the full follow-up period, RFF-Cox exhibits a mean absolute slope deviation of 0.181 compared to 0.299 for CoxPH, representing a 39% relative improvement in calibration slope accuracy. Similarly, RFF-Cox achieves a lower mean ICI (0.020 vs. 0.026). This advantage is most pronounced at early-to-intermediate time horizons (12–96 months). However, deviations of the calibration slope from unity at certain time points indicate potential shifts in the prediction scale, particularly over longer follow-up durations. At later time points (180–240 months), both models exhibit slope values exceeding 1.0, indicating overdispersion in predicted probabilities as the risk set contracts. The slope deviations at intermediate horizons (24–96 months) arise from the interaction between the ridge penalty’s implicit shrinkage of predicted risk scores toward the mean and the Breslow estimator’s finite-sample behavior under non-linear risk functions. Appendix C confirms this pattern is dataset-consistent (SUPPORT2: mean |Slope − 1| = 0.091), supporting that the deviations reflect regularization-induced calibration compression rather than RFF approximation error. Furthermore, the width of the confidence intervals for predicted survival probabilities is observed to increase over time. This trend is consistent with the growing uncertainty in estimates due to increased censoring and a diminishing risk set at later time horizons. To assess sensitivity to the ridge regularization parameter, Table 9 presents calibration metrics across four values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document}. Table 9Calibration sensitivity to ridge parameter on the METABRIC test setλMean |Slope − 1|Mean ICI0.0010.6690.0720.0100.1940.0210.1000.1810.0201.0000.1860.019 Table 9Calibration sensitivity to ridge parameter on the METABRIC test setλMean |Slope − 1|Mean ICI0.0010.6690.0720.0100.1940.0210.1000.1810.0201.0000.1860.019 Calibration sensitivity to ridge parameter on the METABRIC test set Insufficient regularization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.001) leads to substantial calibration degradation, while performance stabilizes at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} ≥ 0.01. The optimal value (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} = 0.1) is consistent with the grid search results in Sect. Hyperparameter sensitivity and coverage analysis. These findings indicate that the slope deviations observed in Table 8 are not attributable to the RFF approximation itself, but rather reflect the interaction between regularization strength and the finite-sample behavior of the Breslow estimator under non-linear risk scores. Figure 5 displays (a) the IPCW-weighted calibration curve obtained at 36 months for the METABRIC test set, contextualized by the findings in Table 8. In the plot, the blue line represents the model predictions, the black dashed line indicates ideal calibration (y = x), and the light blue shaded region corresponds to the analytical 95% confidence band computed via the log–log transformation and the Delta method. Fig. 5a IPCW calibration curve at 36 months with 95% confidence band. b Temporal calibration drift (|slope − 1|) a IPCW calibration curve at 36 months with 95% confidence band. b Temporal calibration drift (|slope − 1|) The curve generally follows the ideal diagonal line. This visual agreement is further supported by the calibration metrics (Slope = 0.969, ICI = 0.0044). However, the confidence band is observed to widen noticeably in the lower prediction range of 0–0.60. This pattern is associated with lower observation density in this range and the increased variance induced by IPCW weights. The confidence band displayed in Fig. 5 reflects the uncertainty surrounding the calibration curve itself. In contrast, the RFF-Cox CI Width reported in Table 8 represents the width of the confidence intervals for the survival probabilities predicted by the model for individual subjects. Since these two metrics represent distinct sources of uncertainty, they are not directly comparable. Nonetheless, the general tendency of the calibration curve to remain close to the ideal line indicates that the model exhibits reasonable calibration at the 36-month horizon. In panel (b) (Fig. 5), RFF-Cox maintains consistently lower slope deviation than CoxPH across all time horizons, with the largest separation at 12–72 months. Both models converge at later horizons (180–240 months), where the contracting risk set limits the discriminating power of either approach. The upward drift observed for both models beyond 150 months is consistent with the increasing CI widths reported in Table 8, reflecting cumulative baseline hazard uncertainty at extended follow-up. Individual survival predictions and uncertainty analysis In addition to the model’s overall performance, the reliability of predictions at the individual level is a critical factor. Therefore, three patients representing low (ID 260), medium (ID 523), and high (ID 283) risk profiles were selected from the test set based on their ranked risk scores for a detailed analysis (Table 10). The confidence intervals reported here incorporate uncertainty from both the linear predictor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:\left({x}\right)$$\end{document} and the baseline cumulative hazard \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H_0\left(t\right)$$\end{document} via the Greenwood variance estimator (Sect. Variance propagation via the multivariate Delta method). Table 10Risk metrics and 36-month predictions for representative patientsPatient IDRisk GroupHR (CoxPH)HR (RFF-Cox) [95% CI]Log-Risk ± SES(36mo) RFF-Cox [95% CI]260Low Risk0.600.51 [0.40, 0.64]−0.67 ± 0.120.954 [0.939, 0.965]523Intermediate Risk1.341.06 [0.72, 1.55]0.05 ± 0.190.906 [0.862, 0.937]283High Risk2.532.46 [2.00, 3.02]0.90 ± 0.110.795 [0.743, 0.839]HR Hazard Ratio, SE Standard Error, CI Confidence Interval; S(36 mo): Predicted survival probability at 36 months. The RFF-Cox model generates confidence intervals using the analytical covariance matrix and the Delta method, incorporating both parameter uncertainty (Var(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document})) and baseline hazard uncertainty (Var(H0)). Table 10Risk metrics and 36-month predictions for representative patientsPatient IDRisk GroupHR (CoxPH)HR (RFF-Cox) [95% CI]Log-Risk ± SES(36mo) RFF-Cox [95% CI]260Low Risk0.600.51 [0.40, 0.64]−0.67 ± 0.120.954 [0.939, 0.965]523Intermediate Risk1.341.06 [0.72, 1.55]0.05 ± 0.190.906 [0.862, 0.937]283High Risk2.532.46 [2.00, 3.02]0.90 ± 0.110.795 [0.743, 0.839]HR Hazard Ratio, SE Standard Error, CI Confidence Interval; S(36 mo): Predicted survival probability at 36 months. The RFF-Cox model generates confidence intervals using the analytical covariance matrix and the Delta method, incorporating both parameter uncertainty (Var(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document})) and baseline hazard uncertainty (Var(H0)). Risk metrics and 36-month predictions for representative patients HR Hazard Ratio, SE Standard Error, CI Confidence Interval; S(36 mo): Predicted survival probability at 36 months. The RFF-Cox model generates confidence intervals using the analytical covariance matrix and the Delta method, incorporating both parameter uncertainty (Var(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document})) and baseline hazard uncertainty (Var(H0)). Table 10 demonstrates the expected separation between risk groups at the 36-month cross-section. For the low-risk patient (ID 260), the hazard ratio (HR) is 0.51, with a survival probability of 95.4%. For the high-risk patient (ID 283), the HR rises to 2.46, while the survival probability declines to 79.5%; the medium-risk patient (ID 523) lies between these two extremes. An examination of uncertainty metrics reveals that log-risk standard errors (SE) remain within the ~ 0.11–0.12 range for low- and high-risk patients, whereas this value increases to 0.19 for the medium-risk patient. This finding indicates higher uncertainty in predictions for medium-risk profiles, consistent with the wider confidence bands observed in Fig. 6. Fig. 6Individual survival predictions for representative low-, medium-, and high-risk patients: A comparison of RFF-Cox and CoxPH Individual survival predictions for representative low-, medium-, and high-risk patients: A comparison of RFF-Cox and CoxPH In Fig. 6, solid lines represent RFF-Cox predictions, dashed lines represent CoxPH predictions, and shaded areas represent the 95% survival confidence bands for the respective patients. The curves for the three patients are separated according to their risk profiles: ID 260 maintains high survival, while ID 283 shows a decline. When comparing the models, the divergence between the RFF-Cox and CoxPH curves is more pronounced for the medium-risk patient (ID 523), likely arising from RFF-Cox capturing non-linear relationships with different flexibility. Confidence bands are narrow for low- and high-risk patients, whereas they are significantly wider for the medium-risk patient, directly corresponding to the higher SE value (0.19) in Table 10. The RFF-Cox model produces an accurate risk ranking and provides a quantitative measure of uncertainty for each patient. The wide confidence bands in the medium-risk group show that the model transparently reflects uncertainty in predictions. This can provide concrete guidance to researchers regarding the reliability of predictions in clinical decision-making processes. Discussion In this study, the Cox proportional hazards model was extended into a scalable framework capable of modeling non-linear risk structures using Random Fourier Features (RFF) representation. The findings demonstrate that the method offers a more flexible prediction structure compared to other survival models in both simulation experiments and real-world applications. In simulations, the U-shaped risk function constructed for the BMI variable clearly revealed this flexibility (Fig. 1). The synthetic covariate distributions employed, particularly the log-normal structures, were selected to reflect the characteristic features of real-world biomedical data [40]. While the classical Cox model failed to capture this curve due to its linearity constraint, RFF-Cox produced a smooth prediction surface closely approximating the true risk profile. This capability stems from the RFF transformation reducing the infinite-dimensional kernel space to a finite-dimensional feature map, thereby enabling the resolution of non-linear relationships via linear optimization tools. Backward compatibility and covariate-level inference An important property of the proposed framework is its backward compatibility with the classical Cox model. In the linear simulation scenario (Table 3b), the equivalent coefficients derived via the first-order Taylor expansion recover the classical Cox PH estimates with negligible differences, confirming that the RFF approximation introduces minimal penalty when the true hazard is linear. In the non-linear scenario (Table 3a), RFF-Cox outperforms CoxPH, recovering the saturated Age effect and avoiding a spurious CRP association produced by CoxPH. The Taylor-expansion-based framework (Sect. Clinical interpretation: equivalent coefficients and interaction inference) serves as a bridge between the classical linear Cox model and the non-linear RFF-Cox model. The first-order expansion yields equivalent linear coefficients (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{equiv}$$\end{document}) for covariate-level hazard ratio estimation and Wald testing. The second-order expansion enables formal hypothesis testing for covariate interactions via the Delta method with Bonferroni correction. In the simulation study, the true SBP × Smoking interaction was identified, with no surviving false positives after correction. This dual-derivative capability offers an advantage over tree-based and gradient boosting methods, where interaction detection typically relies on heuristic measures [10]. Uncertainty quantification and coverage A further key feature of the model is the analytical calculability of prediction uncertainty. While many machine learning-based survival methods produce only point estimates, this study directly derives the parameter covariance matrix via the Fisher information matrix. The covariance matrix was computed via eigen decomposition, preserving the correlation structure among parameters. This enabled the generation of analytical confidence intervals for individual survival curves. The multivariate Delta method propagates uncertainty from both parameter estimation and baseline hazard estimation, yielding more conservative confidence intervals than those obtained from single-source approximation. Analytical confidence intervals were validated through simulation-based coverage assessment (Sect. Hyperparameter sensitivity and coverage analysis). In the linear scenario, coverage approaches the nominal level. In the non-linear scenario, coverage stabilizes at a level closely matching the bootstrap reference. The gap between non-linear and linear coverage is attributable to the cumulative effect of the Breslow-type Hessian approximation and the chain-rule amplification of variance through the survival function. This pattern is consistent with the sensitivity of partial likelihood inference to functional form misspecification [41]. Additionally, strong regularization adversely affects coverage, suggesting that moderate regularization provides a favorable balance between discrimination and coverage. Calibration and temporal drift From a calibration perspective, IPCW-weighted analyses confirm that the predicted probabilities are in agreement with observed event rates. For instance, the proximity of the calibration curve obtained for the 36-month horizon to the ideal diagonal line, along with the low Integrated Calibration Error (ICI), indicates that the model does not produce systematic bias. RFF-Cox predictions are in better agreement with observed event rates than CoxPH, representing a relative improvement in mean absolute slope deviation. Ridge parameter sensitivity analysis (Table 9) indicates that slope deviations observed at certain time points are unlikely to be primarily attributable to the RFF approximation but rather reflect the interaction between regularization strength and finite-sample behavior of the Breslow estimator. At later horizons, both models exhibit slope values exceeding 1.0, indicating overdispersion as the risk set contracts. Cross-dataset validation on SUPPORT2 (Appendix C) supports these findings, suggesting that the observed slope deviations reflect regularization-induced calibration compression rather than dataset-specific effects. Computational efficiency The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left({n}^{3}\right)$$\end{document} Gram matrix cost inherent in traditional kernel methods is reduced to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(n{m}^{2}\right)$$\end{document} with the RFF approach (Table 2). Training times remain at the scale of seconds for large datasets, whereas competing methods require substantially longer. Practical guidance In real-world applications, hyperparameter analyses highlight the balance between the RFF dimension ( m) and performance (Fig. 2). Performance reaches a saturation point at moderate m values; larger values increase the computational burden without providing marginal improvement. Consequently, the optimal hyperparameter map of the model can be effectively extracted via Grid Search. Grid search analyses across simulated and real-world datasets (Sect. Hyperparameter sensitivity and coverage analysis) yield three guidelines: (i) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$m\approx\lceil\sqrt{n}\rceil$$\end{document} suffices for discrimination, but two to three times this value is advisable for reliable uncertainty quantification; (ii) strong regularization should be avoided for analytical CI generation; (iii) moderate regularization offers the best balance. Supplementary analyses (Appendix B) confirm that subsampling for bandwidth selection and early stopping have negligible impact on both σ selection and discriminatory performance. Limitations The study has certain limitations. (i) The model is limited to the RBF kernel; the impact of different kernel structures could be investigated in the future. (ii) The method focuses on a single event type; its extension to competing risks or multi-event scenarios represents an area for future research. (iii) In non-linear scenarios, analytical coverage falls short of the nominal level, reflecting cumulative approximation error. (iv) Bandwidth selection on a subsample may be suboptimal for highly heterogeneous datasets. (v) Detailed calibration analysis was conducted on one primary dataset, with supporting validation on a secondary cohort; additional datasets would further strengthen generalizability. Although the RFF transformation renders interpretability somewhat more abstract compared to classical coefficients, variable effects can be elucidated by employing partial dependence plots alongside model-agnostic explainability tools such as DALEX and SHAP. Future directions Future directions include: (i) extension to competing risks and multi-state models; (ii) investigation of alternative kernels (Matérn, polynomial) within the RFF framework, including connections to Gaussian process survival models [42]; (iii) broader benchmarking against recent deep learning survival models, such as DeepHit, Cox-Time, and DRSA, particularly in settings where the proportional hazards assumption may be violated; (iv) theoretical comparison of the Taylor-expansion-based inference with penalized spline [43] and GAM approaches; and (v) integration of model-agnostic explainability tools (DALEX, SHAP) for additional interpretability support. Conclusion The integration of the Cox model and Random Fourier Features offers a survival analysis method that is scalable, statistically transparent, and capable of modeling non-linear risk relationships. The proposed approach combines the flexibility of kernel methods with the computational efficiency of linear models. The method degenerates to the classical Cox model under linearity while providing substantial gains under non-linearity, with no risk of degradation in either regime. Findings from simulations and real-world data reveal that the method produces more accurate predictions than the classical Cox model in scenarios where the linearity assumption does not hold, while operating significantly faster than traditional machine learning algorithms on large datasets. The fundamental contribution of the method lies in its analytical uncertainty estimation based on the Fisher information matrix and the multivariate Delta method. Thanks to this approach, the model not only performs a risk ranking but also provides confidence intervals that quantitatively assess the reliability of the prediction for each individual patient. The Taylor-expansion-based inference framework bridges the interpretability gap between the classical linear Cox model and the non-linear RFF-Cox setting, enabling covariate-level hazard ratio estimation, Wald-type significance testing, and formal interaction detection — capabilities not available in tree-based or gradient boosting survival methods. In datasets where non-linear effects are weak, RFF-Cox matches classical performance without penalty. The critical advantage in such settings is not discriminatory improvement but the availability of formal inference tools — covariate-level Wald tests, interaction detection, and analytical individual-level confidence intervals — none of which are provided by standard CoxPH or tree-based implementations. The obtained low calibration errors and realistic uncertainty bands indicate that RFF-Cox can be utilized as a reliable tool in clinical risk prediction and biomedical data analysis. Practical guidelines for hyperparameter selection are provided to facilitate adoption. Future studies can advance this potential by extending the method to the context of competing risks, integrating different kernel functions, and conducting broader benchmarking against deep learning survival models. Supplementary Information Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2. Supplementary Material 3. Supplementary Material 3. Supplementary Material 4. Supplementary Material 4.