Efficient GPU-accelerated fitting of observational health-scaled stratified and time-varying Cox models

Abstract: The Cox proportional hazards model stands as a widely-used semi-parametric approach for survival analysis in medical research and many other fields. Numerous extensions of the Cox model have further expanded its versatility. Statistical computing challenges arise, however, when applying many of these extensions with the increasing complexity and volume of modern observational health datasets. To address these challenges, we demonstrate how to employ massive parallelization through graphics processing units (GPU) to enhance the scalability of the stratified Cox model, the Cox model with time-varying covariates, and the Cox model with time-varying coefficients. First we establish how the Cox model with time-varying coefficients can be transformed into the Cox model with time-varying covariates when using discrete time-to-event data. We then demonstrate how to recast both of these into a stratified Cox model and identify their shared computational bottleneck that results when evaluating the now segmented partial likelihood and its gradient with respect to regression coefficients at scale. These computations mirror a highly transformed segmented scan operation. While this bottleneck is not an immediately obvious target for multi-core parallelization, we convert it into an un-segmented operation to leverage the efficient many-core parallel scan algorithm. Our massively parallel implementation significantly accelerates model fitting on large-scale and high-dimensional Cox models with stratification or time-varying effect, delivering an order of magnitude speedup over traditional central processing unit-based implementations.