GRU-TV: Time- and Velocity-aware Gated Recurrent Unit for patient representation.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Multivariate clinical temporal series (MCTS) from electronic health records (EHRs) are crucial for understanding dynamic physiological processes.
• Existing deep learning models for MCTS often struggle with data imputation and irregular sampling.
• The instantaneous velocity of physiological status changes has been largely overlooked in patient representation.

Purpose of the Study:

• To address the gap in representing instantaneous physiological changes in MCTS.
• To develop a novel deep learning model for patient representation learning that accounts for temporal dynamics and velocity.
• To improve the accuracy of patient state characterization from EHR data.

Main Methods:

• Proposed a Time- and Velocity-aware Gated Recurrent Unit (GRU-TV) model.
• Utilized neural ordinary differential equations to model instantaneous physiological velocity.
• Integrated instantaneous velocity into the hidden state updates and forward propagation of the GRU model to handle uneven time intervals and non-uniform changes.

Main Results:

• The GRU-TV model demonstrated strong performance across multiple clinical tasks on real-world datasets (PhysioNet2012 and MIMIC-III).
• Achieved high average AUC scores for sub-tasks (e.g., 0.89 on complete PhysioNet2012 data) and phenotype classification (e.g., 0.84 on complete MIMIC-III data).
• Showcased robustness even with significant data sampling (e.g., 50% on PhysioNet2012, 10% on MIMIC-III) and achieved a mean absolute deviation of 1.84 days for length-of-stay prediction.

Conclusions:

• The findings highlight the critical role of instantaneous physiological changes in patient representation.
• The GRU-TV model offers a significant advancement in clinical decision-making, especially with incomplete or irregularly sampled EHR data.
• Incorporating velocity awareness enhances the ability to model complex physiological dynamics for better patient care insights.