Scalable nonlinear Cox modeling via random Fourier features with analytic uncertainty

Description

Fahrettin Kaya published a 2026 methods paper making nonlinear survival modelling fast enough for large medical cohorts. The workhorse Cox model assumes risk rises in a straight line with each factor, yet real biomarker risks often bend into U shapes. Flexible kernel versions capture such curves but slow to a crawl as sample size grows, and most machine-learning alternatives cannot put error bars on individual predictions.

Kaya's approach, called RFF-Cox, projects the data through random Fourier features, a trick that imitates a kernel with a fixed set of computed features, so cost grows far more gently. A key smoothing parameter is chosen by an approximation to the marginal likelihood. Uncertainty comes from the Fisher information matrix, propagated to survival predictions, and a Taylor-based framework yields hazard ratios and significance tests per factor. The package aims to give big-cohort studies both flexibility and honest error bars.

Catalog record

Source: https://europepmc.org/article/MED/PMC13543420