A machine learning model combining large language model–extracted ultrasound features with anthropometric data distinguished between premenopausal and postmenopausal status and may provide a supplementary assessment approach when reproductive hormone data are unavailable, according to a retrospective, multicenter study published in Frontiers in Endocrinology. 

Researchers developed and externally validated a multimodal machine learning framework using clinical data from 997 Chinese women treated at two centers between January 2020 and April 2026. The study initially included 713 women from Hangzhou Red Cross Hospital and 284 women from the Second Affiliated Hospital of Zhejiang University School of Medicine. To focus model development on the menopausal transition period, analyses were restricted to women aged 45 to 55 years, resulting in a development cohort of 305 women and an external validation cohort of 110 women.

The researchers used the Qwen3.5-plus large language model (LLM) to automatically extract three menopause-related morphological features from free-text gynecologic ultrasound reports: ovarian atrophy, endometrial atrophy, and uterine atrophy. They evaluated these ultrasound-derived features alone and in combination with anthropometric variables and reproductive hormone measurements using eight machine learning algorithms: XGBoost, Random Forest, Support Vector Machine, K-Nearest Neighbors, LightGBM, Artificial Neural Network, CatBoost, and Naive Bayes.

Model performance was evaluated in an independent external validation cohort using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score. SHapley Additive exPlanations (SHAP) analysis was used to quantify each feature’s contribution to model predictions. 

Clinical and ultrasound features differ by menopausal status

Reproductive hormone profiles differed substantially between premenopausal and postmenopausal women. Follicle-stimulating hormone and luteinizing hormone concentrations were substantially higher in postmenopausal women, whereas estradiol concentrations were markedly lower. Anti-Müllerian hormone showed the largest relative decrease, falling by more than 99%, although the researchers noted that this marker had a high rate of missing data because it was selectively used for fertility assessment.

The LLM also identified statistically significant differences in ultrasound-derived ovarian, endometrial, and uterine atrophy between premenopausal and postmenopausal women. According to the researchers, these morphological features complemented reproductive hormone measurements and were incorporated into the multimodal prediction model.

Multimodal model improves prediction of menopausal status

Using anthropometric features alone produced an AUC of 0.839 in the external validation cohort. Using hormone features alone achieved an AUC of 0.980, while ultrasound morphological features alone achieved an AUC of 0.907. Combining anthropometric and ultrasound features increased the AUC to 0.935. The highest performance was achieved by combining anthropometric and hormone features (AUC, 0.984), while incorporating all available features produced an AUC of 0.985. The researchers concluded that ultrasound morphology supplemented hormone-based prediction and may provide a supplementary assessment approach when hormone data were unavailable.

When the eight machine learning algorithms were compared using anthropometric and ultrasound morphological features, CatBoost achieved the best performance in the external validation cohort, with an AUC of 0.935, accuracy of 87%, precision of 78%, recall of 90%, and an F1 score of 0.833. Tree-based ensemble models generally outperformed traditional machine learning approaches. 

To validate the reliability of LLM-extracted morphological features, 100 ultrasound reports were randomly selected from the training set and independently interpreted by two attending gynecologists. The LLM-derived features were compared with expert consensus. Overall extraction performance improved across the evaluated Qwen models, with qwen-plus achieving the highest overall macro F1 score of 0.879. 

SHAP analysis showed that follicle-stimulating hormone was the most influential predictor when hormone data were available, followed by estradiol, age, luteinizing hormone, body mass index, and anti-Müllerian hormone. When hormone measurements were excluded, age became the strongest predictor, while ovarian, endometrial, and uterine atrophy all contributed positively to menopause prediction. Under these conditions, the model maintained an AUC of 0.935, supporting the use of ultrasound morphology combined with anthropometric features when hormone testing was unavailable.

The researchers concluded that “ultrasound morphological features can serve as a supplement to hormone features for predicting menopause.”

The authors reported no commercial or financial conflicts of interest and received no financial support for the research or its publication.