Medical coding is a fundamental but difficult task in modern healthcare: it maps clinical documents to standardized ICD codes, demanding familiarity with an enormous code space. A new preprint on arXiv proposes a span-centric learning framework for LLM-based explainable medical coding, shifting attention from whole documents to the specific spans of text that justify each code.
The authors argue that identifying these supporting spans can make the coding process more transparent. Rather than treating a document as an opaque unit, the model is trained to ground its predictions in concrete evidence, which is especially valuable in clinical settings where accountability matters.
Because the abstract is brief, the paper's full methodology and experimental results are not detailed here. The contribution, as stated, is a span-centric design for explainable coding, addressing a known limitation of black-box document-level classifiers.