Personalized Diabetes Treatment Support Using Large Language Models Fine-Tuned on Electronic Health Records: Development and Evaluation Study
JMIR Formative Research · 2026
DOI: 10.2196/71541 ↗Kurz gefasst
Fachartikel in der Zeitschrift JMIR Formative Research, 2026. Verfasst von Shenyang He und 2 weiteren Autor:innen. Beteiligte Einrichtung: Department of Clinical Laboratory Medicine, Jinniu Maternity and Child Health Hospital of Chengdu, Chengdu, 610041, China, 86 18227387870. Hauptthemen laut MEDLINE: Large Language Models, Electronic Health Records, Precision Medicine, Diabetes Mellitus. Die Arbeit wurde bisher 5-mal zitiert (Crossref) und nennt 25 Literaturangaben. Der Volltext ist frei in PubMed Central verfügbar.
Themen: Diabetes mellitus
Abstract (Originalsprache)
Background
Effective diabetes management requires individualized treatment strategies tailored to patients’ clinical characteristics. With recent advances in artificial intelligence, large language models (LLMs) offer new opportunities to enhance clinical decision support, particularly in generating personalized recommendations.
Objective
This study aimed to develop and evaluate an LLM-based outpatient treatment support system for diabetes and examine its potential value in routine clinical decision-making.
Methods
Three compact LLMs (Llama 3.1-8B, Qwen3-8B, and GLM4-9B) were fine-tuned on deidentified outpatient electronic health records using a parameter-efficient low-rank adaptation approach. The optimized models were embedded into a prototype hospital information system via a retrieval-augmented generation framework to generate individualized treatment recommendations, laboratory test suggestions, and medication prompts based on demographic and clinical data.
Results
Among the models evaluated, the fine-tuned GLM4-9B demonstrated the strongest performance, producing clinically reasonable treatment plans and appropriate laboratory test recommendations and medication suggestions. It achieved a mean Bilingual Evaluation Understudy for 4-grams score of 67.93 (SD 2.74) and mean scores of 44.30 (SD 3.91) for Recall-Oriented Understudy for Gisting Evaluation for overlap of unigrams, 27.34 (SD 1.85) for Recall-Oriented Understudy for Gisting Evaluation for overlap of bigrams, and 37.67 (SD 2.88) for Recall-Oriented Understudy for Gisting Evaluation for Longest Common Subsequence.
Conclusions
The fine-tuned GLM4-9B shows strong potential as a clinical decision support tool for personalized diabetes care. It can provide reference recommendations that may improve clinician efficiency and support decision quality. Future work should focus on enhancing medication guidance, expanding data sources, and improving adaptability in cases involving complex comorbidities.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Autor:innen und Institutionen
- Department of Clinical Laboratory Medicine, Jinniu Maternity and Child Health Hospital of Chengdu, Chengdu, 610041, China, 86 18227387870.
- Information Center, West China Hospital of Sichuan University, Chengdu, Sichuan, China.
ORCID-Profile: Shenyang He, Yu Zhang, Jiaxi Li
Zitieren
APA
He, S., Zhang, Y., & Li, J. (2026). Personalized Diabetes Treatment Support Using Large Language Models Fine-Tuned on Electronic Health Records: Development and Evaluation Study. JMIR Formative Research. https://doi.org/10.2196/71541
Vancouver
He S, Zhang Y, Li J. Personalized Diabetes Treatment Support Using Large Language Models Fine-Tuned on Electronic Health Records: Development and Evaluation Study. JMIR Formative Research. 2026. doi:10.2196/71541
BibTeX
@article{He2026,
author = {He, Shenyang and Zhang, Yu and Li, Jiaxi},
title = {Personalized Diabetes Treatment Support Using Large Language Models Fine-Tuned on Electronic Health Records: Development and Evaluation Study},
journal = {JMIR Formative Research},
year = {2026},
doi = {10.2196/71541}
}
EUVIMED