MiGHT-EHR Improves Clinical Prediction with Graph Transformers

Electronic Health Records (EHRs) contain a wealth of information that could revolutionize clinical prediction, but their inherent complexity makes effective learning challenging. EHRs are characterized by heterogeneous clinical entities (like patients, diagnoses, prescriptions), their intricate interactions, longitudinal patient data across visits, and shared statistical patterns across various prediction tasks. Existing machine learning methods often only capture a subset of these crucial properties, limiting their predictive power. To overcome these limitations, researchers developed MiGHT-EHR, a Multi-task Graph transformer for Heterogeneous Temporal EHRs. This innovative framework constructs a heterogeneous graph from EHR data, where nodes represent different clinical entities and edges signify statistically associated entities. By modeling these complex relationships within a unified representation learning method, MiGHT-EHR can jointly process all three challenging aspects of EHR data. Evaluated on the MIMIC-III and MIMIC-IV datasets, MiGHT-EHR demonstrated superior performance compared to current state-of-the-art methods across four critical tasks: drug recommendation, prediction of length-of-stay, mortality, and readmission. The improvements were particularly significant for mortality and readmission predictions. Furthermore, analysis of the learned representations revealed clinically interpretable structures, such as patient neighborhoods organized by outcomes and salient medical concepts recoverable as linear directions, indicating its potential for both accurate prediction and clinical insight.