Ascribe improves Thai clinical note accuracy with significance-based reasoning
ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation
Computation and Language
Summary
Writing detailed medical notes takes a lot of time for doctors, and computers can help but often miss important details or add wrong ones. The authors created ASCRIBE, a system that decides how important pieces of information are before making summaries, helping large language models write better medical notes. They also made a new dataset of real Thai medical conversations to help improve and test these tools. Their method made summaries more complete and precise compared to earlier techniques.
What this means in practice
- •For hospital data teams: Create more reliable automatic summaries of Thai clinical visits by using a significance-aware reasoning method for large language models.
- •For medical software developers: Build clinical documentation tools that leverage ASCRIBE prompts to improve medical note completeness and factual accuracy in Thai-language healthcare settings.$Commercial implications: It enables new Thai-language clinical note generation software with higher accuracy and completeness, valuable for healthcare providers and software vendors.
Authors
Tarm Kalavantavanich, Teerawut Ponarchar, Pattaramanee Arsomngern, Jenta Wonglertsakul, Watcharakorn Chuthong, Chiraphat Boonnag, Knot Pipatsrisawat, Titipat Achakulvisut
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.