A1624
Title: DeepMsurv: M-spline based DNN framework for time-to-event data
Authors: Vu Tuan Anh - Pukyong National University (Korea, South) [presenting]
Takeshi Emura - Hiroshima University (Japan)
Il Do Ha - Pukyong National University (Korea, South)
Abstract: In survival analysis, deep neural networks (DNNs) have been widely used to capture complex nonlinear relationships in time-to-event data. However, standard DNN-based models such as DeepSurv primarily optimize the Cox partial likelihood and do not directly estimate the baseline hazard function during training. To address this limitation, a novel M-spline-based DNN survival framework for time-to-event data, referred to as DeepMsurv, is proposed, which is efficiently optimized via a loss function based on the full log-likelihood under a five-parameter cubic M-spline-based hazard. By integrating the M-splines into the network architecture to flexibly model the baseline hazard, the approach enables the direct construction of interpretable hazard plots. Through comprehensive simulation studies and practical real-data applications, the proposed method demonstrates competitive predictive performance compared with state-of-the-art models, while substantially improving computational efficiency, interpretability, and clinical utility. The DeepMsurv method was implemented using PyTorch.
