Related Experiment Videos
Multimodal machine learning for menopause status prediction using LLM-extracted ultrasound features
Weiwei Yin1, Zhengyuan Shen2, Chun Feng3
1Department of Obstetrics and Gynecology, Hangzhou Red Cross Hospital, Hangzhou, Zhejiang, China.
Frontiers in Endocrinology
|July 17, 2026
Summary
Large language models (LLMs) extract ultrasound features to predict menopausal status, offering a valuable alternative when hormone tests are unavailable. This AI-driven approach enhances women
Area of Science:
- Reproductive endocrinology and AI-driven healthcare analytics.
- Development of multimodal predictive models for women's health.
Background:
- Menopausal status assessment is vital for women's health management and chronic disease risk stratification.
- Traditional menopause diagnosis relies on serum hormone tests (e.g., FSH, AMH), considered the clinical gold standard.
- Ultrasound is a routine gynecological exam, but its unstructured report data is underutilized for predictive modeling.
Purpose of the Study:
- To investigate the utility of large language models (LLMs) in extracting morphological features from unstructured ultrasound reports for menopausal status prediction.
- To develop and evaluate a multimodal prediction model integrating ultrasound-derived features with anthropometric data.
Main Methods:
- Utilized a training set of 713 Chinese women and an external validation set of 284 women.
- Employed LLMs to automatically extract three key morphological features: ovarian atrophy (OA), endometrial atrophy (EA), and uterine atrophy (UA) from ultrasound reports.
- Constructed a multimodal prediction model by fusing extracted ultrasound features with anthropometric data and trained eight machine learning models.
Main Results:
- The highest Area Under the Curve (AUC) of 0.984 was achieved by integrating anthropometric and hormone features in the validation set.
- A model combining anthropometric features and ultrasound morphological features yielded an AUC of 0.935.
- The qwen-plus LLM demonstrated superior feature extraction performance, closely aligning with expert annotations.
Conclusions:
- LLMs can effectively extract structured morphological information from unstructured ultrasound reports.
- The fusion of ultrasound morphological features and anthropometric data offers a supplementary method for assessing menopausal status.
- This approach is particularly beneficial in clinical settings where hormone data is temporarily unavailable.