Abstract:
MEGRec addresses retrospective specialty or service label ranking from heterogeneous electronic health record inputs. Clinical text, demographic variables, and temporal-location features are encoded and fused through an attention-based representation, followed by a hierarchical extreme multi-label classifier over M=250 labels. Evaluation on the MIMIC-III benchmark reports P@10=90.7%, R@10=87.6%, F1@10=89.1%, NDCG@10=89.5%, and Macro-AUC =94.5%, with lower training and inference cost than XMC and MUFIN under the reported setup. An explanation module combines modality-level attention and medical concept mapping to generate case-level rationales. Qualitative examples and limited clinician feedback indicate that the generated explanations are clinically plausible, but the explanation evidence remains preliminary because the human evaluation is small and non-comparative. The study therefore presents a scalable proof-of-concept for explainable specialty label ranking from multi-source EHR data under a large output space.
Keywords:
Heterogeneous EHR Fusion, Extreme Multi-label Classification, Explainable AI, ClinicalBERT, Attention Fusion, Clinical Decision Support, Specialty Label Ranking, Hierarchical Classification