Patients-like-me Framework for Explainable Clinical Prediction

The Patients-like-me (PLM) framework combines language models with graph neural networks via a Variational Expectation-Maximization algorithm to model inter-patient relationships. Tested on MIMIC-III and MIMIC-IV, it improves clinical prediction performance and provides reference-patient attribution with minimal computational overhead.