Bridging the Version Gap: Multi-version Training Improves ICD Code Prediction, Especially for Rare Codes

In three linesA label-wise attention model trained on combined ICD-9 and ICD-10 data improves rare medical code prediction by 27% micro F1 (18K rare codes) and macro metrics on frequent codes, despite version mismatch. Version-independent approach to automate clinical coding.

Summary generated by Claude — human-verified