Optimal Personalized and Dynamic Classification with Fixed Trees for Survival Outcomes

Abstract: Motivated by Alzheimer’s disease (AD) research need to predict risk of cognitive decline dynamically based on various biomarkers and risk factors in the presence of complicated interactions, I will present a work-in-progress statistical framework that utilizes Fixed Survival Trees (FSTs) for time-dynamic classification of survival outcomes.

FSTs provide a structured yet adaptable approach to classifying individuals into high- and low-risk categories using potentially longitudinally collected biomarkers. Unlike traditional tree-based survival models, which dynamically grow and prune based on data, FSTs maintain a fixed topology while optimizing time-dependent classification thresholds. Our method incorporates cumulative and incidence-based dynamic classification principles, extending time-dependent ROC and AUC concepts to obtain optimal decision thresholds at internal nodes. Additionally, we introduce a semi-parametric modeling framework that incorporates personalized risk factors and efficiently estimates classification thresholds using a maximum concordance estimator. By integrating survival analysis with tree-based classification, FSTs aim to provide personalized, interpretable, and computationally efficient risk stratification tools for AD progression.