Penalised spline estimation of covariate-specific time-dependent ROC curves

María Xosé Rodríguez-Álvarez, Vanda Inácio · 2025

Read Penalised spline estimation of covariate-specific time-dependent ROC curves by María Xosé Rodríguez-Álvarez, Vanda Inácio (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-13604.txt.

Introduction

Maria Xose Rodriguez-Alvarez and Vanda Inacio build a smoother way to judge how well a biomarker forecasts an event over time, in work from 2025. Time dependent ROC curves are the standard tool for such checks. A marker's accuracy, though, may shift with patient traits like age or sex, and older estimators handle that poorly.

Their estimator uses penalized splines, curves kept smooth by a roughness cost. Flexible models cover both the event risk and the marker's own spread. Bent risk patterns and non proportional hazards fit naturally. Simulations test the method against earlier ones. The goal is accuracy scores that respect how patients differ.

Catalog record

CREATORMaría Xosé Rodríguez-Álvarez, Vanda Inácio
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE68 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-13604

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Penalised spline estimation of covariate-specific time-dependent ROC curves María Xosé Rodríguez-Álvarez, Vanda Inácio Abstract The identification of biomarkers with high predictive accuracy is a crucial task in medical research, as it can aid clinicians in making early decisions, thereby reducing morbidity and mortality in high-risk populations. Time-dependent receiver operating characteristic (ROC) curves are the main tool used to assess the accuracy of prognostic biomarkers for outcomes that evolve over time. Recognising the need to account for patient heterogeneity when evaluating the accuracy of a prognostic biomarker, we introduce a novel penalised-based estimator of the time-dependent ROC curve that accommodates a possible modifying effect of covariates. We consider flexible models for both the hazard function of the event time given the covariates and biomarker and for the location-scale regression model of the biomarker given covariates, enabling the accommodation of non-proportional hazards and nonlinear effects through penalised splines, thus overcoming limitations of earlier methods. The simulation study demonstrates that our approach successfully recovers the true functional form of the covariate-specific time-dependent ROC curve and the corresponding area under the curve across a variety of scenarios. Comparisons with existing methods further show that our approach performs favourably in multiple settings. Our approach is applied to evaluating the Global Registry of Acute Coronary Events risk score’s ability to predict mortality after discharge in patients who have suffered an acute coronary syndrome and how this ability may vary with the left ventricular ejection fraction. An R package, CondTimeROC , implementing the proposed method is provided. † † footnotetext: María Xosé Rodríguez-Álvarez, CITMaga and Departamento de Estatística e Investigación Operativa, Universidade de Vigo, Vigo, Spain ( [email protected] ). Vanda Inácio, School of Mathematics, University of Edinburgh, Scotland, UK ( [email protected] ) key words: Biomarker, location-scale regression model, piecewise exponential additive model, penalised splines, predictive accuracy, survival analysis 1 Introduction Prognostic markers that accurately predict disease onset, progression, or mortality are essential in clinical practice, and it is critical to rigorously evaluate their predictive accuracy before adopting them. Prospective studies are often conducted to assess the prognostic performance of a biomarker, with the marker measured at baseline and patients followed over time to observe the occurrence of the event of interest (e.g., disease onset, recurrence, or death). In this context, the disease outcome is clearly time-dependent: subjects are initially event-free but may experience the event during the follow-up period. The receiver operating characteristic (ROC) curve is arguably the most popular graphical tool for evaluating the discriminatory ability of a continuous biomarker for a binary disease outcome. In classical ROC analysis, within the context of a diagnostic study, a subject’s disease status is treated as a fixed characteristic, with the biomarker measured concurrently with the binary disease status. To account for the time lag between marker measurement and event occurrence in prognostic studies, time-dependent extensions of sensitivity, specificity, and consequently the ROC curve, have been proposed. The seminal paper is due to Heagerty et al., 2000 , with Heagerty and Zheng, 2005 introducing a taxonomy for time-dependent measures of accuracy (i.e., sensitivity and specificity). As detailed in Section 2.1 , we focus on the so-called cumulative/dynamic ROC curve, which allows to evaluate how well a marker measured at baseline can distinguish between subjects who experience the event of interest and those who do not, within a given follow-up interval of the form [0,t] . This has been a prolific area of research over the past two decades (for a review, see Blanche et al., 2013 ; Kamarudin et al., 2017 , and references therein) . It is now widely acknowledged that patient characteristics can influence the discriminatory ability of a biomarker, and that failing to account for them may lead to misleading conclusions about the biomarker’s performance. Understanding how covariates affect a biomarker’s performance is therefore vital for identifying the subpopulations in which the biomarker can be reliably used. Although extensive literature exists on adjusting for covariates in the ROC framework for diagnostic studies, spanning parametric, semiparametric, and nonparametric methods (see Pardo-Fernández et al., 2014 ; Inácio et al., 2021 , for a review) , comparatively less attention has been given to covariate adjustment in time-dependent ROC curves for prognostic settings. Most existing approaches fall within the semiparametric framework and primarily rely on proportional hazards regression models ( Song and Zhou, 2008 ; Zheng et al., 2010 ; Zheng et al., 2012 , see, e.g.,) . While these methods provide a structured modelling framework, they have two key limitations: (i) proportional hazards are assumed for both covariates and biomarker effects, and (ii) continuous covariates effects are modelled linearly. Both (i) and (ii) may not always be appropriate in practice. From a fully nonparametric perspective, the only available method is that of Rodríguez-Álvarez et al., 2016 , which employs kernel smoothing with inverse probability of censoring weights. However, this approach is restricted to a single continuous covariate, limiting its applicability, and it exhibits high variance, particularly in small samples, which are not uncommon in prognostic studies. Recently, Sewak et al., 2025 developed an alternative modelling strategy, employing semiparametric marginal models for the biomarker and event time distributions, linked via a Gaussian copula to capture their joint dependence. This approach, while not relying on any form of proportional hazards model, still assumes linear effects of continuous covariates. This work is motivated by the clinically relevant application in Rodríguez-Álvarez et al., 2016 , which examined the impact of left ventricular ejection fraction (LVEF) on the prognostic value of the Global Registry of Acute Coronary Events (GRACE) risk score for predicting post-discharge mortality in patients with acute coronary syndrome. Although LVEF is a well-established predictor of mortality, it is not included in the GRACE score, leading to risk estimates that may be confounded by LVEF and underscoring the need for appropriate covariate adjustment. Discrepancies between the results of the nonparametric method of Rodríguez-Álvarez et al., 2016 and those obtained using the semiparametric approach of Song and Zhou, 2008 underscore the sensitivity of covariate-adjusted prognostic performance to modelling assumptions. To address this gap in the literature, we develop a flexible semiparametric approach for the covariate-specific time-dependent ROC curve, modelling the hazard function of event times given the biomarker and covariates using a piecewise exponential additive model ( Friedman, 1982 ; Bender et al., 2018 ) . This model accommodates both continuous and categorical variables and allows for different functional forms of the biomarker and covariates effects. Importantly, and unlike Song and Zhou, 2008 , we incorporate (i) smooth nonlinear effects of continuous covariates and of the biomarker, and (ii) time-varying (potentially nonlinear) effects, thereby addressing non-proportional hazards. As in Song and Zhou, 2008 , our estimators of sensitivity and specificity (and thus the ROC curve) require estimation of the biomarker distribution given the covariates, which we achieve using a location-scale model that incorporates smooth nonlinear effects in both the mean and variance functions. In all cases, smooth nonlinear effects are modelled using penalised splines ( Eilers and Marx, 1996 ; Wood, 2017 ) . Our approach therefore offers a compromise between the competing methods proposed by Song and Zhou, 2008 and Rodríguez-Álvarez et al., 2016 . Notably, the variance of the estimators produced by our method can be substantially lower than that of

Read the complete text