Patients-like-me: A Variational LM-GNN Framework for Explainable Clinical Prediction

A novel framework called Patients-like-me (PLM) has been developed, merging language models (LMs) with graph neural networks (GNNs) to enhance clinical predictions derived from electronic health records (EHRs). This method fuses local patient semantics from LMs with the broader cohort structure provided by GNNs, facilitating reference-patient attribution for better explainability. To optimize PLM's training, a Variational Expectation-Maximization algorithm alternates updates between LMs and GNNs under a supervised variational objective. Tests conducted on MIMIC-III and MIMIC-IV indicate that PLM surpasses leading techniques across both encoder-only and decoder-only LM frameworks, incurring only slight additional computational costs. The research can be found on arXiv with the identifier 2608.04193.