GRU-D-Weibull: A novel real-time individualized endpoint prediction

Area of Science:

• Computational biology
• Health informatics
• Machine learning

Background:

• Electronic health record (EHR) data offers potential for chronic disease management but suffers from quality issues hindering traditional predictive modeling.
• Real-world implementation of predictive models is limited by data quality challenges in EHRs.

Purpose of the Study:

• To propose and evaluate a novel predictive modeling approach, GRU-D-Weibull, for real-time individualized endpoint prediction and risk management using EHR data.
• To assess the performance and real-world implementability of GRU-D-Weibull in patients with chronic kidney disease stage 4 (CKD4).

Main Methods:

• Developed GRU-D-Weibull, a model leveraging gated recurrent units with decay (GRU-D) to model Weibull distribution for endpoint prediction.
• Systematically evaluated GRU-D-Weibull against alternative survival models (AFT, XGB(AFT), RSF, Nnet-survival) using a CKD4 patient cohort (n=6879).
• Compared performance using metrics like C-index and L1-loss, and experimented with in-process and post-process calibrations.

Main Results:

• GRU-D-Weibull demonstrated strong predictive performance (C-index ~0.77 at 4.3 years follow-up) and superior accuracy (L1-loss ~0.45 years at 4 years follow-up) compared to other models.
• The model effectively handled missing EHR data, with prediction errors decreasing as more data became available post-index date.
• Post-training recalibration aligned predicted survival probabilities with observed outcomes across various prediction horizons.

Conclusions:

• GRU-D-Weibull offers advantages in handling missing EHR data and providing reliable probability and point estimates for diverse prediction horizons.
• The model shows potential for individualized endpoint risk management in chronic diseases, utilizing real-time clinical data.
• Further research on data quality impacts and clinical workflow integration is warranted.