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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Natural Language Processing

Background:

• Automated glaucoma subtype classification from clinical notes is currently limited by a lack of clinically actionable, subspecialty-aligned explanations.
• Existing AI models often lack transparency, hindering their practical application in healthcare settings.

Purpose of the Study:

• To develop an interpretable AI system for automated glaucoma subtype classification from clinical notes.
• To enhance the clinical utility of AI in ophthalmology by providing subspecialty-aligned explanations.

Main Methods:

• Extended the Ci-SSGAN architecture with a GPT-5.2-to-Qwen3-8B teacher-distilled reasoning module.
• Fine-tuned the Qwen3-8B model on 2,660 de-identified ophthalmology notes with expert-reviewed rationales.
• Evaluated the model's performance on 294 notes using ROUGE-L and BERTScore metrics.

Main Results:

• The fine-tuned Qwen3-8B model achieved a ROUGE-L score of 0.792 ± 0.013 and a BERTScore F1 of 0.955 ± 0.004.
• Outperformed eight zero-shot comparators, including advanced models like GPT-4o and GPT-4.1.
• Demonstrated the efficacy of privacy-preserving distillation for creating interpretable AI.

Conclusions:

• Privacy-preserving distillation is a viable method for developing interpretable AI in medical domains.
• The developed model offers a path towards clinically actionable AI for glaucoma subtype classification.
• Enhanced AI interpretability can significantly improve clinician adoption and trust in AI-driven healthcare solutions.