Time-series deep survival prediction for hemodialysis patients using an attention-based Bi-GRU network

Abstract

Background and objective: The number of end-stage renal disease (ESRD) patients treated with hemodialysis (HD) has significantly increased, but the prognosis remains poor. Time-series features have been included in only a few studies to predict HD patient survival, and how to utilize such features effectively remains unclear. This article aims to develop a more accurate, interpretable, and clinically practical personalized survival prediction model for HD patients.

Methods: This study proposed and evaluated an attention-based Bi-GRU network using time-series features for survival prediction. A distance-based loss function was proposed to improve performance. We used data from 1232 ESRD patients who received regular hemodialysis treatment for ≥ 3 months from 2007 to 2016 at the First Affiliated Hospital of Zhejiang University. The proposed model was compared with representative sequence modeling deep learning architectures and existing survival analysis methods in terms of the C-index and IBS value. Post hoc tests were used to test statistical significance. The attention map was used to assess feature importance over time. The impact of time-series changes on survival was investigated after controlling initial values (using BMI as an example).

Results: The proposed method outperformed other sequence modeling architectures and the state-of-the-art survival analysis approaches in terms of the C-index and the integrated Brier score (IBS) value. Our method achieved a C-index of 0.7680 (95% confidence intervals [CI]: 0.7645, 0.7716) and an IBS of 0.1302 (95% confidence intervals [CI]: 0.1292, 0.1313), showing an improvement of up to 5.4% in terms of the C-index and a decrease of 3.2% in terms of the IBS value. The addition of the distance-based loss function improved the performance. The predicted risk and actual risk levels closely agreed. This study also found that even after controlling the initial body mass index (BMI) values, different 3-month BMI trends could produce different survival outcomes.

Conclusions: This study proposed a more effective and interpretable method to use time-series information in survival analysis. The proposed method may help promote personalized medicine and improve patient prognosis.

Keywords: Deep learning; Hemodialysis (HD); Loss function; Survival analysis; Time-series.