Attention-Based Deep Recurrent Model for Survival Prediction

Zhaohong Sun,Wei Dong,Jinlong Shi,Kunlun He,Zhengxing Huang

DOI: https://doi.org/10.1145/3466782

2021-10-31

ACM Transactions on Computing for Healthcare

Abstract:Survival analysis exhibits profound effects on health service management. Traditional approaches for survival analysis have a pre-assumption on the time-to-event probability distribution and seldom consider sequential visits of patients on medical facilities. Although recent studies leverage the merits of deep learning techniques to capture non-linear features and long-term dependencies within multiple visits for survival analysis, the lack of interpretability prevents deep learning models from being applied to clinical practice. To address this challenge, this article proposes a novel attention-based deep recurrent model, named AttenSurv , for clinical survival analysis. Specifically, a global attention mechanism is proposed to extract essential/critical risk factors for interpretability improvement. Thereafter, Bi-directional Long Short-Term Memory is employed to capture the long-term dependency on data from a series of visits of patients. To further improve both the prediction performance and the interpretability of the proposed model, we propose another model, named GNNAttenSurv , by incorporating a graph neural network into AttenSurv, to extract the latent correlations between risk factors. We validated our solution on three public follow-up datasets and two electronic health record datasets. The results demonstrated that our proposed models yielded consistent improvement compared to the state-of-the-art baselines on survival analysis.