Thai hospitals generate thousands of clinical documents every day, yet much of that text remains locked inside free-form notes that computers cannot reliably interpret. Researchers have now built and tested a proof-of-concept framework that uses large language models to pull clinical terms out of Thai medical records and link them to SNOMED CT, the international vocabulary that lets health systems share and compare data. The best-performing model recognized clinical entities with a loose F1 score of 83.31 percent, while concept linking remained the harder task, with the strongest model reaching an Intersection over Union score of 39.42 percent.
The work addresses a practical barrier. SNOMED CT enables semantically interoperable health data, but its adoption in Thailand has been slowed by language and workflow challenges. To test whether automation could help, the team assembled a manually annotated dataset from 220 Thai medical documents drawn from three outpatient departments, containing 6,390 annotated entities. Human annotators marked where clinical concepts appeared in the text, creating a ground-truth reference against which the models could be measured.
The framework follows a retrieval-augmented design. Instead of asking a language model to guess the correct SNOMED CT code from memory, the system first retrieves candidate concepts using encoder models and then re-ranks them, presenting the model with the most similar options. The researchers evaluated four large language models in this pipeline: Gemini 2.5 Pro and GPT-4.1 as closed-source systems, and Qwen 3 and MedGemma as open-source alternatives. Performance was scored on two separate tasks, entity recognition using precision, recall and F1, and concept linking using a character-level metric called Intersection over Union, alongside measurements of time and cost.
On entity recognition, Gemini 2.5 Pro achieved
On entity recognition, Gemini 2.5 Pro achieved the highest performance, with a loose F1 score of 83.31 percent. This task asks the model to find the spans of text that correspond to clinical concepts, such as a diagnosis or a medication, without necessarily assigning them a code. The open-source models trailed behind, and the gap illustrates a recurring tension in clinical natural language processing: the most capable systems are often proprietary, while locally deployable models sacrifice some accuracy for control and data privacy.
Concept linking proved substantially harder for every model. Measured by Intersection over Union, Gemini scored 39.42 percent, GPT-4.1 scored 34.44 percent, MedGemma scored 29.86 percent and Qwen scored 28.62 percent. These figures indicate that even the strongest configuration got the character-level overlap right in roughly two out of five cases, leaving considerable room for improvement before automated coding could stand in for trained human coders.
The resource costs of the pipeline were also quantified. Average end-to-end processing time was five to six minutes per document for Gemini and GPT-4.1, with estimated costs of 0.035 and 0.019 United States dollars per document respectively. Local open-source inference required up to ten minutes per document. For a hospital processing large volumes of records, these per-document figures scale quickly, and the trade-off between accuracy, cost and speed becomes a concrete planning decision rather than an abstract one.
The authors frame the results as a proof
The authors frame the results as a proof of concept rather than a deployable system. The dataset covered outpatient documents from three departments, so performance on other document types remains untested. What the study establishes is that automated clinical named entity recognition and SNOMED CT concept linking from Thai medical documents is feasible with current language models, and that the performance and resource requirements of closed-source and open-source options can now be compared on a human-annotated benchmark. The approach offers a template for adapting the terminology to other low-resource linguistic settings where SNOMED CT was not designed around the local language.
Subject of Research: Automated extraction of SNOMED CT concepts from Thai medical documents using large language models
Article Title: A framework for structuring Thai medical documents using systematized nomenclature of medicine clinical terms and large language models
Article References: Tangchitnob, N., Ponthongmak, W., Kijsanayotin, B.,
Article References: Tangchitnob, N., Ponthongmak, W., Kijsanayotin, B., Pattanaprateep, O., Phusanti, S., Atiksawedparit, P., Suwanthaweemeesuk, K., McKay, G. J., Oami, T., & Thakkinstian, A. (2026). A framework for structuring Thai medical documents using systematized nomenclature of medicine clinical terms and large language models. PLOS Digital Health, 5(10), e0001791. https://doi.org/10.1371/journal.pdig.0001791
DOI:10.1371/journal.pdig.0001791
Keywords: SNOMED CT, large language models, Thai medical documents, clinical named entity recognition, concept linking, health data interoperability, retrieval-augmented framework, natural language processing, digital health, clinical terminology, framework, structuring
News. AI reads Thai medical records and maps them to a global clinical vocabulary. Scienmag
