AutoMap: Automatic Medical Code Mapping for Clinical Prediction Model Deployment

Given a deep learning model trained on data from source hospital, how to deploy the target hospital automatically? How accommodate heterogeneous medical coding systems across different hospitals? Standard approaches rely existing code mapping tools, which have several practical limitations. To tackle this problem, we propose AutoMap automatically map codes EHR in coarse-to-fine manner: (1) Ontology-level Alignment: We leverage ontology structure learn coarse alignment between and systems; (2) Code-level Refinement: refine at fine-grained level for downstream tasks using teacher-student framework. evaluate models with two real-world datasets: eICU MIMIC-III. Results show that achieves relative improvements up 3.9% (AUC-ROC) 8.7% (AUC-PR) mortality prediction, 4.7% 3.7% (F1) length-of-stay estimation. Further, can provide accurate systems. Lastly, demonstrate adapt challenging scenarios: completely hospitals.