PGCASurv: A Prior-Guided Cross-Attention Framework for Dynamic Survival Model with Longitudinal Data.

Abstract

This study proposes a dynamic survival analysis method to characterize how a patient's risk of adverse outcomes evolves over time during long-term follow-up. Traditional approaches usually provide a one-time risk score at baseline, which makes it difficult to capture improvement or deterioration in a patient's condition during follow-up. We first treat a patient's static baseline information as a "prior" risk profile, which represents the patient's baseline risk structure relative to the overall population. A deep neural network is then used to learn how repeatedly measured test results and other follow-up measurements modify this prior over the time axis, thereby yielding a dynamic risk trajectory that is consistent with clinical intuition. The model further introduces a cross-attention mechanism to update longitudinal information under the reference frame defined by the static prior. The proposed method is trained and evaluated on five clinical datasets covering different diseases and is compared with commonly used statistical models and state-of-the-art neural-network-based methods. The results show that, across all datasets, the proposed method is more accurate and reliable, or at least statistically competitive, in discriminating between high- and low-risk individuals and in providing survival probabilities. Furthermore, through inter-pretability analysis, the important risk factors identified by the model are consistent with existing medical evidence. Overall, considering dynamic risk as a combination of "population-level prior" and "individual follow-up adjustment" shows promising modeling potential and may help clinicians identify high-risk patients earlier and formulate more individualized follow-up and treatment plans.