Development of a deep learning model for survival prediction in heart failure: competing risk and frailty model

Abstract

This study presents a novel deep learning (DL) framework, the Deep Neural Frailty Competing Risks (DNFCR) model, which simultaneously integrates frailty and competing risks (CR) for mortality prediction in heart failure (HF). While existing models like Neural Frailty Models (NFM) address frailty and DeepHit handles CR, DNFCR is the first DL approach to combine both components, offering improved handling of censored data in healthcare applications. We evaluated DNFCR against established models (DeepSurv, CoxPH) using real-world HF data, assessing its ability to capture unobserved heterogeneity through proportional (DNFCR_PF) and non-proportional (DNFCR_NF) frailty structures. In this retrospective cohort study, we analyzed 435 HF patients (enrolled March-September 2018; 5-year follow-up until July 2023) with 57 demographic/clinical features, categorized by cause-specific mortality. The models' performance was evaluated using the C-index, IBS, and INBLL. Our results in mortality from heart failure demonstrated marginal but consistent improvements in predictive accuracy when incorporating frailty (C-index: ~0.66). Differences in predicting other-cause mortality were minimal across the models. DNFCR-PF also performed better regarding the IBS (0.17 ± 0.01) and INBLL (0.02 ± 0.53). However, the clinical relevance of the cause-specific mortality requires further validation in both outcome categories. Comparative analysis revealed that DNFCR and traditional models achieved comparable accuracy in survival prediction, highlighting that DL's superiority is context-dependent and influenced by data and unmeasured confounders. Strengths include DNFCR's potential for personalized risk stratification and adaptability to other diseases with CR. The DNFCR model showed DL potential in HF survival prediction but requires clinical validation compared to traditional approaches.

Ähnliche Arbeiten