Multimodal machine learning for menopause status prediction using LLM-extracted ultrasound features.

Background: The assessment of menopausal status is crucial for individualized health management and risk stratification of chronic diseases in women. The traditional assessment of menopause mainly relies on serum reproductive hormone tests, such as FSH, AMH, etc., whose test results are stable and reliable and are the gold standard for clinical diagnosis. Meanwhile, ultrasound is a routine gynecological examination method. But the unstructured information in its report has not been systematically used for predictive modeling. Methods: This study included 713 Chinese women from Hangzhou Red Cross Hospital as the training set and another 284 from the Second Affiliated Hospital of Zhejiang University as the independent external validation set. Three morphological features of ovarian atrophy (OA), endometrial atrophy (EA) and uterine atrophy (UA) were automatically extracted from unstructured ultrasound report text by using large language model (LLM), and fused with anthropometric features to construct a multimodal prediction model. Eight machine learning models were trained and their predictive performances were compared. Results: The highest AUC of 0.984 was obtained by integrating anthropometric and hormone features in the validation set. The AUC is 0.935, which combines anthropometric features and ultrasound morphological features. The qwen-plus with the largest number of LLM parameters has the best feature extraction performance and is closest to the labeling results of human experts. Conclusion: LLM can effectively extract structured morphological features from ultrasound report text. The fusion of ultrasound morphological features and anthropometric features can provide supplementary evaluation methods for assessing menopausal status. This method is particularly suitable for clinical scenarios where hormone data is temporarily missing or unavailable.