EpiCure: an epoch-based generalized EM probabilistic inference for mixture cure model

We propose EpiCure, a novel framework that integrates the Expectation-Maximization (EM) algorithm with neural network training for mixture cure modeling. The key innovation is an epoch-based generalized EM procedure that updates latent cure statuses once per epoch, avoiding costly nested iterations while preserving probabilistic interpretability. EpiCure jointly models cure incidence and survival latency via shared representation learning, with the latency component flexibly specified as a mixture of Weibull and Log-normal primitives. Extensive experiments on simulated and real-world survival datasets demonstrate that EpiCure achieves superior predictive performance in terms of AUC and Integrated Brier Score (IBS), while maintaining computational efficiency comparable to state-of-the-art methods. Convergence analysis and clinical case studies further validate the stability and practical utility of the proposed framework.