Objective: To develop an efficient and domain-adapted system to process colonoscopy and pathology reports using knowledge distillation techniques. Patients and Methods:We implemented a knowledge distillation framework to create a smaller, domain-specific LLM-based NLP model for summarizing and extracting key information from clinical reports.The model was trained on a dataset consisting of 5,500 colonoscopy reports and 7,000 pathology reports taken from January 1, 2024, through June 30, 2024.Performance was evaluated against ground truth polyp categories derived from pathology report diagnoses. Results:The distilled model demonstrated high domain-specific performance, achieving 95.2% accuracy (95% CI, 93.9%-96.5%),0.95 precision, and a 31.5% improvement in inference speed relative to the teacher model.Despite being substantially smaller than the teacher model, it maintained strong capability in polyp category identification from key clinical factors including polyp number, size, histology, and location across colonoscopy and pathology reports.Domain clinicians reported high agreement with model outputs across all six evaluated clinical questions, confirming its reliability for supporting follow-up recommendation workflows. Conclusion:This work presents a step toward making domain-specific NLP models for gastroenterology more efficient and scalable.By leveraging knowledge distillation, we demonstrate the potential for creating more practical domain-specific models that can assist in interpreting complex clinical documentation.Future work will focus on real-world validation and expanding the model to other procedural report types.
Mau et al. (Fri,) studied this question.