Interpretable EHR Disease Prediction System Based on Disease Experts and Patient Similarity Graph (DE-PSG)

Deep learning models, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks, have gained significant traction in leveraging the data collected from electronic health records (EHRs) to predict future events or patient outcomes in the healthcare industry. Though these models already proficiently capture sequence data and provide invaluable insights and treatment solutions for patients, it would be desirable to further enhance their versatility, interpretability, and ability to handle sparse patient data. In this paper, we introduce an interpretable EHR disease prediction system founded on disease experts and a patient similarity graph (DE-PSG). The model leverages the Multi-gate Mixture-of-Experts (MMoE) model as its foundation framework and integrates multiple expert Gated Recurrent Unit (GRU) and Transformer models to enhance the predictive capabilities of the base model. Addressing the challenge of sparse disease data, this study constructs data based on a patient similarity graph. To boost interpretability, a multi-expert network is introduced to emulate expertise from various medical domains. Through the auxiliary expert loss function, the proficiency of experts in predicting specific diseases is improved. Experimental evaluations of our newly proposed DE-PSG model using the MIMIC-IV dataset reveal its superior disease prediction performance. In comparison to the leading benchmark model, Residual Mixture-of-Experts (R-MoE), the DE-PSG model enhances the AUROC by 2.3%.