LATCN model outperforms state-of-the-art with 28% recall boost in medical coding
Nuzumlalı et al.'s label-wise attention TCN slashes missed codes by nearly a third
Medical coding—assigning diagnosis and procedure codes from hospital notes—is notoriously difficult, even for experts, because relevant information is scattered across long documents and varies per code. To tackle this, the team models the task as multi-label text classification and introduces a novel architecture: a multi-layer temporal convolutional network (TCN) followed by label-specific attention. The TCN learns global document structure over very long sequences, while the attention mechanism lets the model focus on different sections of the same note for each code—something previous models struggled to do.
The results are striking: compared to the 2019 state-of-the-art, the model boosts F1 by 9% and recall by 28%. The researchers emphasize recall as the more critical metric in clinical settings, where missing a diagnosis code could have serious consequences. The work, originally done in 2019 but posted as a record, demonstrates that combining global sequence learning with label-wise attention can significantly improve automated medical coding, paving the way for more reliable clinical decision support systems.
- Model uses multi-layer Temporal Convolutional Networks (TCN) to capture long-range dependencies in clinical notes.
- Label-wise attention mechanism enables focusing on different text sections for each diagnosis or procedure code.
- Achieves 9% F1 improvement and 28% recall increase over 2019 baselines on multi-label medical coding.
Why It Matters
Automated, high-recall medical coding reduces human error and speeds up reimbursement, improving patient care workflow.