A framework for structuring Thai medical documents using systematized nomenclature of medicine clinical terms and large language models
5 affiliations
- 1Department of Clinical Epidemiology and Biostatistics, Faculty of Medicine Ramathibodi Hospital, Mahidol University, Bangkok, Thailand.
- 2Thai Health Information Standards Development Center, Health Systems Research Institute, Ministry of Public Health, Bangkok, Thailand.
- 3Faculty of Medicine Ramathibodi Hospital, Chakri Naruebodindra Medical Institute, Mahidol University, Bangkok, Thailand.
- 4Centre for Public Health, Queen's University Belfast, Belfast, United Kingdom.
- 5Department of Emergency and Critical Care Medicine, Chiba University Graduate School of Medicine, Chiba, Japan.
Abstract
Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT) enables semantically interoperable health data, but implementation in Thailand is hindered by language and workflow challenges. We evaluated whether large language models (LLMs) can automate SNOMED CT concept extraction from Thai clinical documents. The objectives of this study are to develop a proof-of-concept framework for automated clinical named entity recognition and SNOMED CT concept linking from Thai medical documents, and to compare the performance and resource requirements of closed-source and open-source LLMs on a human-annotated dataset. The dataset was manually annotated from 220 Thai medical documents (6,390 entities) from three outpatient departments. Our retrieval-augmented framework was evaluated using four LLMs (Gemini 2.5 Pro, GPT-4.1, Qwen 3, and MedGemma) as experimental based models which were comparatively analyzed for entity recognition using precision, recall, F1 score and concept linking using character-level metric, Intersection over Union (IoU), as well as resource consumptions. For entity recognition, Gemini achieved the highest performance (loose F1 83.31%). For SNOMED CT concept extraction ranked most similar (Top-1) by encoder models for retrieval and re-ranking, IoU scores were 39.42% (Gemini), 34.44% (GPT), 29.86% (MedGemma), and 28.62% (Qwen). Average processing time for end-to-end extraction was 5–6 minutes per document for Gemini and GPT, with cost estimates in United states dollars (USD) of USD 0.035 and USD 0.019 per document, respectively; local open-source inference required up to 10 minutes per document. This study established a proof-of-concept framework for automated clinical entity recognition and SNOMED CT concept linking from Thai medical documents, demonstrating practical LLM performance with trade-offs between accuracy, cost, and efficiency across closed-source and open-source models. Author summary: We developed and tested a way to turn Thai clinical notes into a standard structured health data. This work addresses a problem in healthcare data landscape where the most useful clinical information is recorded in the messy doctor notes making it difficult to use, analyze, and share for research and decision-making. This challenge is particularly more pronounced in non-English speaking settings like Thailand where fewer tools and resources are available. With the advancement of language model technologies, we used large language models to automatically identify medical concepts in Thai medical documents and link them to standardized international medical terminology. The results show promise, but performance depends on how clearly information is written and how well local terms match standard terms. We found that the language technologies can capture clinical concepts in low-resourced language settings. In long term, this may support better secondary use of health data which is the foundational requirement for data-driven decision-making, patient-centered healthcare delivery, and value-based health system.
Related
- Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature
- Genetic diversity and phylogeography of swine influenza A virus in 2013-2022 in Europe using a harmonized genotyping nomenclature
- Notes on the nomenclature of Artemisia (I): A. subg. Ponticae (Asteraceae)
- Performance of Pseudonymization Methods Applied to Hospital Documents in the Context of Health Data Warehouses
- Biological sex and disease architecture in systemic sclerosis: a multilayer translational framework for precision medicine
BibTeX
@article{tangchitnob2026framework,
title = {{A framework for structuring Thai medical documents using systematized nomenclature of medicine clinical terms and large language models}},
author = {Nat Tangchitnob and Wanchana Ponthongmak and Boonchai Kijsanayotin and Oraluck Pattanaprateep and Sithakom Phusanti and Pongsakorn Atiksawedparit and Kamonporn Suwanthaweemeesuk and Gareth J McKay and Takehiko Oami and Ammarin Thakkinstian},
journal = {PLOS Digital Health},
year = {2026},
month = oct,
doi = {10.1371/journal.pdig.0001791},
url = {https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001791},
note = {CC BY},
}
RIS
TY - JOUR
TI - A framework for structuring Thai medical documents using systematized nomenclature of medicine clinical terms and large language models
AU - Nat Tangchitnob
AU - Wanchana Ponthongmak
AU - Boonchai Kijsanayotin
AU - Oraluck Pattanaprateep
AU - Sithakom Phusanti
AU - Pongsakorn Atiksawedparit
AU - Kamonporn Suwanthaweemeesuk
AU - Gareth J McKay
AU - Takehiko Oami
AU - Ammarin Thakkinstian
PY - 2026
DA - 2026/10/01
JO - PLOS Digital Health
DO - 10.1371/journal.pdig.0001791
UR - https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001791
AB - Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT) enables semantically interoperable health data, but implementation in Thailand is hindered by language and workflow challenges. We evaluated whether large language models (LLMs) can automate SNOMED CT concept extraction from Thai clinical documents. The objectives of this study are to develop a proof-of-concept framework for automated clinical named entity recognition and SNOMED CT concept linking from Thai medical documents, and to compare the performance and resource requirements of closed-source and open-source LLMs on a human-annotated dataset. The dataset was manually annotated from 220 Thai medical documents (6,390 entities) from three outpatient departments. Our retrieval-augmented framework was evaluated using four LLMs (Gemini 2.5 Pro, GPT-4.1, Qwen 3, and MedGemma) as experimental based models which were comparatively analyzed for entity recognition using precision, recall, F1 score and concept linking using character-level metric, Intersection over Union (IoU), as well as resource consumptions. For entity recognition, Gemini achieved the highest performance (loose F1 83.31%). For SNOMED CT concept extraction ranked most similar (Top-1) by encoder models for retrieval and re-ranking, IoU scores were 39.42% (Gemini), 34.44% (GPT), 29.86% (MedGemma), and 28.62% (Qwen). Average processing time for end-to-end extraction was 5–6 minutes per document for Gemini and GPT, with cost estimates in United states dollars (USD) of USD 0.035 and USD 0.019 per document, respectively; local open-source inference required up to 10 minutes per document. This study established a proof-of-concept framework for automated clinical entity recognition and SNOMED CT concept linking from Thai medical documents, demonstrating practical LLM performance with trade-offs between accuracy, cost, and efficiency across closed-source and open-source models. Author summary: We developed and tested a way to turn Thai clinical notes into a standard structured health data. This work addresses a problem in healthcare data landscape where the most useful clinical information is recorded in the messy doctor notes making it difficult to use, analyze, and share for research and decision-making. This challenge is particularly more pronounced in non-English speaking settings like Thailand where fewer tools and resources are available. With the advancement of language model technologies, we used large language models to automatically identify medical concepts in Thai medical documents and link them to standardized international medical terminology. The results show promise, but performance depends on how clearly information is written and how well local terms match standard terms. We found that the language technologies can capture clinical concepts in low-resourced language settings. In long term, this may support better secondary use of health data which is the foundational requirement for data-driven decision-making, patient-centered healthcare delivery, and value-based health system.
SN - 2767-3170
ER -