Prediction of coronary heart disease risk based on multimodal EHRs

Clinical diagnostic information for cardiovascular-related diseases is often stored in various forms within electronic health records (EHRs). The utilization of EHRs for accurate early prediction and diagnosis of patients facilitates timely interventions, ultimately leading to a reduction in mortality rates. However, in the context of coronary heart disease risk prediction tasks, relying on a single type of data or model restricts the amount of information that can be captured during data mining, thereby resulting in suboptimal prediction accuracy. To address these challenges, this paper proposes the coronary heart disease prediction model, CHD-TSBAL, which leverages the combination of Bio-Bert, cross-attention, and Bi-LSTM. Firstly, EHRs are embedded using Bio-Bert. Subsequently, Bi-LSTM is employed to capture temporal data, incorporating long-term information to extract temporal features. Construction of textual and time-series analysis models for clinical diagnosis, disease risk and health status, respectively, using cross-attention. Finally, these models are fused to complete the overall model construction process. Experimental evaluations conducted on the MIMIC-III dataset show that CHD-TSBAL model significantly enhances the prediction of coronary heart disease risk. Specifically, it achieves 4.25% increase in ROC-AUC compared to Bi-LSTM and 2.98% increase in accuracy compared to DFATS.