Temporal Visiting-Monitoring Feature Interaction Learning for Modelling Structured Electronic Health Records

Xiang Li, Haoxuan Li, Jun Wang, Mengyue Yang

Keywords: Structured EHR, Clinical Prediction, Temporal Feature Interaction

Abstract: Electronic health records (EHRs) contain patients’ longitudinal visit records, and modelling EHRs can be applied to various clinical prediction tasks. Previous works primarily focus on visit sequences and perform feature interaction on visit-level data to capture patient states. Nonetheless, incorporating finer-grained monitoring sequences simultaneously in structured EHRs, where each visit involves multiple monitoring sessions, can improve prediction performance. However, these studies have not accounted for the relationships between visit-level and monitoring-level data. To fill this gap, we propose an EHRs modelling method aimed at modelling the dynamic interaction between visit-level and monitoring-level data and capturing finer-grained health trends. We first capture the dynamic influence between medical data, and then perform a visiting-monitoring feature interaction on the relationships between visit data and monitoring data, to obtain the representation of patients' state for clinical prediction. We conducted extensive experiments on disease prediction and drug recommendation tasks, with MIMIC-III and MIMIC-IV datasets, demonstrating that our method outperforms state-of-the-art models significantly.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8702