Objective: The multivariate clinical temporal series (MCTS) extracted from electronic health records (EHRs) can characterize the dynamic physiological processes. Previous deep patient representation models were proposed to address imputation values and irregular sampling in MCTS. However, the change in physiological status, particularly instantaneous velocity, has not received adequate attention.
Methods: To address this gap, we propose a Time- and Velocity-aware Gated Recurrent Unit (GRU-TV) model for patient representation learning. In the GRU-TV model, we apply the neural ordinary differential equation to describe the instantaneous velocity of the patient's physiological status. This instantaneous velocity is embedded in the hidden state updating process of the basic GRU model for the awareness of uneven time intervals. Besides, the forward propagation of the GRU-TV model also incorporates this instantaneous velocity to enable the perception of non-uniform changes in the patient's physiological status over time.
Results: The performance of the GRU-TV model is evaluated on multiple clinical concerns across two real-world datasets. The average AUC for the sub-tasks on the complete, 70% sampled, and 50% sampled PhysioNet2012 datasets are 0.89, 0.84, and 0.83, respectively. The average AUC for the acute care phenotype classification on the complete, 20% sampled, and 10% sampled MIMIC-III datasets are 0.84, 0.82, and 0.80, respectively. The mean absolute deviation of the length-of-stay regression task is 1.84 days.
Conclusion: The superior performance underscores the importance of instantaneous physiological changes in patient representation and clinical decision-making, particularly under challenging data conditions.
Keywords: Electronic health records; Irregularly sampled series; Ordinary differential equation; Patient representation learning.
Copyright © 2025. Published by Elsevier Inc.
