Deep Mixture Weibull: An Explainable AI Approach for Survival Analysis with Principle Hidden Competing Risks

Survival analysis estimates the probability of survival times for subjects and has broad applications across various domains. Despite the wide-ranging applications of survival analysis across various domains, its successful implementation in critical sectors involving life safety remains challenging due to the lack of interpretability and transparency of existing advanced machine learning (ML) models. To address this, we propose Deep Mixture Weibull (DMW) model, which combines the flexibility of deep learning with the interpretability of Weibull distributions. In real-world scenarios where competing risks exist, the probability of event occurrence is naturally a combination of different underlying risk probabilities. Lever-aging this intuition, DMW models survival outcomes through multiple two-parameter Weibull components, each representing a principal hidden competing risk (PHCR). To tackle multimodal identifiability, we design a specialized loss function incorporating penalty terms for the shape and scale parameters of the Weibull model. Furthermore, expression weights for each component are determined by high-dimensional features and interpreted post hoc using Shapley values, enabling insight into how different PHCRs impact individual subjects. We validate the effectiveness of DMW through extensive experiments on multiple public benchmarks, demonstrating superior predictive accuracy and interpretability compared to state-of-the-art models.