RESurv: A Deep Survival Analysis Model to Reveal Population Heterogeneity by Individual Risk.

Qiguang Zheng,Qifan Shen,Xin Su,Kuo Yang,Zixin Shu,Xuezhong Zhou

DOI: https://doi.org/10.1109/bibm55620.2022.9995313

2022-01-01

Abstract:Surviva1 analysis is a widely used statistical approach to model and analysis time-to-event data, which consists of event occurrence and duration before it occurs. Traditional statistical methods have good interpretability but performance is limited by strong assumptions. In recent years, methods based on machine learning, especially deep learning, have achieved success, but the lack of interpretability limits their real-world applications. This paper proposes a novel interpretable deep learning survival analysis method, RESurv, which predicts the hazard directly without any priori assumptions. Experiments on multiple public datasets validate that RESurv outperforms or approaches state-of-the-art methods. We further estimate the inflence of exposure factors on prognosis at individual level by using counterfactual approaches. Patients are then divided into subgroups by hierarchical clustering. Significant differences in risk response patterns and prognosis suggest heterogeneity in their pathological mechanisms. We explore the risk factors between subpopulations identified by different characteristics and find distinct risk patterns among them, some of which are consistent with medical research findings. The results suggest the potential of RESurv to estimate individual risk, and thereby revealing the heterogeneity of risk response patterns across the populations.