MiGHT-EHR graph transformer outperforms baselines on 4 clinical prediction tasks

Electronic health records are messy: they stitch together patients, visits, diagnoses, prescriptions, and procedures across time, and each prediction task—mortality, readmission, length-of-stay, drug recommendation—shares statistical structure. Most existing models capture only part of that picture. A new paper from Anirudh Rayas, Yuan Wang, and Pavan Turaga introduces MiGHT-EHR (Multi-task Graph Transformer for Heterogeneous Temporal EHRs), a unified representation learning method that jointly models heterogeneous clinical entities, longitudinal patient trajectories, and interdependencies across tasks. The model constructs a heterogeneous graph where nodes are clinical entities and edges connect statistically associated entities identified via normalized point-wise mutual information, then applies graph transformer layers to learn temporal and relational patterns.

On MIMIC-III and MIMIC-IV, MiGHT-EHR outperforms state-of-the-art baselines across all four evaluated tasks, with particularly strong improvements in mortality and readmission prediction. The authors also show the learned representations are clinically interpretable: patient neighborhoods cluster by outcomes, salient medical concepts emerge as linear directions in embedding space, and the model's probability estimates are well calibrated. That mix of performance and interpretability makes MiGHT-EHR a promising backbone for downstream clinical decision support, where understanding why a model flags a high-risk patient matters as much as the prediction itself.