Knowledge Injected Multimodal Irregular EHRs Model for Medical Prediction

The health conditions among patients in intensive care units (ICUs) are essential for providing optimal care and improving patient outcomes. Electronic health records (EHRs) serve as a valuable source of information in ICUs, capturing numerical time series data and lengthy clinical note sequences. However, these EHRs often present challenges due to the irregularity of data collection at varying time intervals. Additionally, existing works ignored to incorporate domain knowledge and integrate memory knowledge from existing multimodal EHRs to improve the care process. Our method first addresses the challenges by (1) transforming clinical note representations into multivariate irregular time series, where each note is associated with different time intervals; (2) injecting static domain knowledge based on patient clinical manifestations into patient representations; (3) storing knowledge from the samples in dynamic memory and learning to inject memory knowledge based on patient representations. The exceptional performance of our proposed methods in two medical prediction tasks consistently outperforms state-of-the-art (SOTA) baselines in both single modality and multimodal fusion scenarios.