FrailNet
Deep survival modeling with temporal adaptation and latent heterogeneity
FrailNet is a deep survival modeling framework for time-to-event prediction in settings where the standard assumptions of Cox regression break down: when hazards evolve nonmonotonically in response to past events, when subjects differ in ways that observed covariates cannot fully capture, and when events are rare.
The core idea is to preserve the Cox partial likelihood structure while replacing its static log-risk term with a richer representation. Recurrent layers (LSTM/GRU) condition current risk on each subject’s accumulated history, a reconstruction module handles denoising and imputation under structured missingness, and a variational frailty term captures persistent unobserved heterogeneity between subjects. The framework also supports recurrent events natively, where multiple onsets within the same subject spell are modeled as repeated occurrences rather than terminal failures.
These properties make FrailNet applicable across any domain where agents or systems adapt to experience, missingness is structured rather than random, and events of interest are infrequent, including epidemiology, criminology, political science, and clinical research.
We evaluate FrailNet on two centuries of dyadic interstate conflict data from the Correlates of War project, spanning 108,797 dyad-years across 21,777 country pairs. The domain is deliberately chosen as a demanding stress test: states adapt strategically to prior interactions, latent conflict propensity varies widely across dyads, and events range from rare (MID onset, 5.1%) to extremely rare (interstate war, 0.4%). Against Cox PH, DeepSurv, SurvNet, and DeepHit baselines, FrailNet’s advantage grows with event rarity and is most pronounced at war onset (C = 0.970 vs. 0.919 for the next best model).
Materials
Note: This reflects the current draft of an active research manuscript submitted for peer review.