Back

Clinically aligned rationale generation for glaucoma subtype classification via a knowledge-distilled language model

Moradi, M.; Fujita, A.; Bineshfar, N.; Vu, D. M.; Aziz, K.; Liebman, D.; Wang, M.; Elze, T.; Eslami, M.; Zebardast, N.

2026-06-26 ophthalmology
10.64898/2026.06.15.26355522 medRxiv
Show abstract

Automated glaucoma subtype classification from clinical notes remains clinically unactionable without subspecialty-aligned explanations supporting clinician-facing deployment. We extended our Ci-SSGAN with a GPT-5.2-to-Qwen3-8B teacher-distilled reasoning module, fine-tuning Qwen3-8B on 2,660 de-identified ophthalmology notes using expert-reviewed rationales. On 294 notes, the fine-tuned model achieved ROUGE-L 0.792 and BERTScore F1 0.955, surpassing eight zero-shot comparators including GPT-4o and GPT-4.1, establishing privacy-preserving distillation as a path to interpretable AI.

Matching journals

The top 1 journal accounts for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.