Chain-of-Thought Supervision Enables Explainable and Efficient Medical Summarization with Small Language Models

Abstract

Medical report summarization is vital for enhancing clinical decision-making, reducing cognitive load, and standardizing healthcare documentation. However, current dependence on large language models (LLMs) faces challenges due to their computational intensity, high inference costs, and cloud dependency, conflicting with strict clinical privacy and latency needs. To address this, we introduce a new framework for reasoningbased summarization using Small Language Models (SLMs) fine-tuned with structured Chain-of-Thought (CoT) supervision. We constructed a synthetic dataset of 110 histopathology subtopics, annotated with Single-CoT and Diverse-CoT rationales. Utilizing LoRA-based fine-tuning, our compact Qwen2.5-0.5B model, Med16Reason-0.5B (fine-tuned on Diverse-CoT), achieved a ROUGE-L F1 of 43.00 and BERT F1 of 76.06. This performance surpassed that of larger biomedical models such as BioMistral-7B, AlpaCare-7B, and BioMedLM-2.7B, and demonstrated strong human alignment through GPT-4.1 evaluations (e.g., 8.99 summary score, 43 blind picks). The explainability of our model was assessed through qualitative analysis of its reasoning chains, and its clinical utility was validated via both automated metrics and a pilot human evaluation with clinical experts. Structured feedback from practicing pathologists confirmed that the model's transparent reasoning enables error detection and supports, rather than replaces, expert clinical judgment. Our proposed model is a viable alternative to larger parameter scales for clinical summarization, offering a pathway toward interpretable, private, and resource-efficient AI in healthcare.

Ähnliche Arbeiten