Related Experiment Video
Updated: May 26, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Automated Monkeypox Identification from Electronic Medical Records Using Large Language Models - Shenzhen City,
Diyang Xue1, Ye Ye2, Yu Wu1
1Shenzhen Center for Disease Control and Prevention, Shenzhen City, Guangdong Province, China.
Introduction:
Monkeypox (mpox) is an emerging infectious disease with diverse clinical manifestations that complicate early case identification. Traditional diagnostic methods relying on laboratory confirmation may delay detection during the early stages of infection. Clinical free-text data in electronic medical records (EMRs), such as chief complaints and histories of present illness, contain valuable information for early identification yet remain underutilized in routine surveillance. This study explored an automated approach to mpox identification from EMRs using large language models (LLMs).
Methods:
We conducted a retrospective study of 239 individuals - 126 laboratory-confirmed mpox cases and 113 non-mpox controls - from multiple hospitals in Shenzhen, China. Free-text chief complaints and histories of present illness were processed using the DeepSeek-R1-14B large language model to extract structured clinical variables. Fourteen clinically relevant features were selected for model development. Naïve Bayes, logistic regression, and random forest classifiers were trained on these features, and a Jieba plus term frequency-inverse document frequency (TF-IDF) approach served as a baseline comparison. Qwen3-14B was also applied for feature extraction as a secondary comparison.
Results:
mpox cases most frequently presented at fever clinics (55.6%), with considerable variation in care-seeking patterns. Models using DeepSeek-extracted features outperformed those based on traditional TF-IDF representations and Qwen3-14B, with logistic regression achieving the best performance [area under the receiver operating characteristic curve (AUROC), 0.927; accuracy, 87.5%]. DeepSeek demonstrated high accuracy (96.1%) in extracting relevant clinical symptoms, including fever, rash, pustules, and HIV infection.
Conclusion:
LLM-based extraction of clinical features from routine text data represents a promising approach for automated mpox case identification and may support future intelligent surveillance and early warning systems.
