Modern triage requires simultaneous evaluation ofmultiple organ systems; single-label classifiers fail to capture comorbidity.We address clinical routing from basic lab parametersand a minimal adaptive interview triggered when a value is out ofreference range. We posit that a BERT-based model pre-trainedin the medical domain (Masked Language Modelling, MLM)can route patients to eight classes (primary care, specialists,emergency SOR). We use a hybrid data approach: MIMIC-III/IV(specialists and SOR cases) and Synthea (chronic trajectories andprimary care). Labels come from ICD-to-specialist mapping; labresults are quantised for triage. We focus on methodology: MLMpre-training on the hybrid corpus, multi-label fine-tuning with aclinically cost-sensitive loss, and “honest” patient-level validation.Results include ROC curves, calibration (ECE), Precision-Recallcurves, confusion matrices, and attention visualisation for interpretability.High ROC AUC for SOR and low ECE supporthuman-in-the-loop deployment with emphasis on safety.
Banasik et al. (Tue,) studied this question.