Time-aware attention-based deep representation learning for multi-source longitudinal data with structured missingness in electronic medical records

This study introduces a new framework to address structured missing data in multi-source longitudinal electronic medical records (EMR). We use a time-aware attention-based deep learning approach to handle the multi-source and longitudinal irregularity of EMR data. Traditional late fusion methods based on multi-modal data often fall short in capturing the correlations between different data sources. Our method includes a mask-guided self-attention module that captures missing patterns and dependencies over time within each data source. Additionally, we designed a time-aware cross-source attention module to fuse sequences from different sources. This module aligns them chronologically and learns global correlations and time dependencies. We enhance this approach with a contrastive loss method. This method reduces the distance between embeddings from different sources, promoting effective data fusion. Auxiliary tasks based on missing masks help reconstruct original sequences, aiding in learning effective representations between sources. We evaluated our framework on large-scale, multi-source EMR databases like MIMIC-IV and eICU-CRD. Our method performs better in predicting in-hospital mortality and length of stay (LoS) compared to state-of-the-art methods. This framework addresses EMR data complexities and improves clinical predictive models, ensuring more accurate and comprehensive patient health assessments.