ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation

2026-10-01

AI Summary

This study addresses the omission of critical information and the generation of unsupported content in SOAP note generation, alongside the absence of Thai clinical datasets. We propose a physician-inspired atomic fact salience reasoning framework that optimizes summarization by assigning clinical significance levels to atomic facts. The model is further refined through prompt engineering, synthetic data training, and Group Relative Policy Optimization (GRPO) reinforcement learning. As key contributions, we construct and release ThaiClinicBench, the first real-world Thai clinical summarization benchmark, together with a synthetic training dataset. Experimental results demonstrate that the proposed reasoning framework outperforms Chain-of-Thought prompting. Furthermore, when employed as a reward signal within GRPO, it enables smaller models to match Gemini in factual precision while surpassing it in completeness.

Abstract

Automatic SOAP note generation can ease the documentation burden on physicians, but existing reasoning methods often omit clinically important information and generate unsupported content. Progress in Thai is further hindered by the lack of publicly available datasets. We propose ASCRIBE, a physician-inspired reasoning framework that ascribes a clinical-significance level to each extracted atomic fact in the conversation before summarization, making a general-purpose LLM a more reliable scribe. We also release ThaiClinicBench, the first de-identified Thai clinical summarization benchmark of real encounters, together with a synthetic training corpus derived from real clinical notes. As a prompt, ASCRIBE outperforms chain-of-thought prompting on GPT-5.4 and Gemini 3.1 Pro across the physician-aligned LLM-judge metrics and improves on standard prompting by up to 10.3 points on the completeness LLM-judge metric. As a GRPO reward, it enables a Gemma-4-E4B model trained solely on synthetic data to match Gemini 3.1 Pro in factual precision and surpass it in completeness. Code and data can be found at https://github.com/loolootech/ascribe.

Problem

• SOAP note generation
• clinical reasoning
• Thai language
• dataset scarcity
• hallucination

Innovation

• Atomic Fact Reasoning
• Clinical Significance
• SOAP Note Generation
• GRPO Reward
• Thai Clinical Benchmark