New perspectives on semi-parametric and pseudo-value approaches for modeling clinical time-to-event data

Abstract

Clinical decision-making often relies on quantitative measures derived from statistical time-to-event models, enabling risk assessment through the quantification of survival probabilities. A key goal of these modeling approaches in guiding clinical decisions is to provide accurate risk estimates while using as little patient information as possible. Standard time-to-event modeling techniques often rely on restrictive assumptions, and their violation bear the risk of biased estimates. Furthermore, these models may need to be tailored to specific population groups, such as children or elderly patients. The aim of this cumulative dissertation was to develop and evaluate new modeling approaches for clinical time-to-event outcomes, focusing on interpretability and applicability in clinical settings, minimal model assumptions, and the ability to filter out the most relevant patient information required for accurate risk estimation. To this end, this dissertation presents three modeling strategies that address the aforementioned goals. In the first work, a tool for the pre-interventional risk assessment of 30-day mortality in the population of elderly patients was developed. Translating the underlying semi-parametric Cox model to a simple scoring system, this tool is user-friendly and only involves three risk factors. The development process focused on interpretability and applicability in clinical settings, while relying on a selection of the most relevant patient information within the Cox model framework. To further reduce assumptions, the second and third works developed modeling approaches for risk assessment in terms of survival probabilities and the restricted mean survival time. These methods, which are based on pseudo-value regression and machine learning methods, demonstrate the reduction of assumptions compared to standard modeling techniques while being able to automatically select most relevant risk factors and interactions among them. The presented modeling approaches maintain interpretability and are able to quantify causal treatment effects, as illustrated in simulation studies and on clinical datasets. All research articles included in this dissertation have been published in international peer-reviewed journals.