Development and Validation of an Interpretable Machine Learning Model for Osteoporosis Prediction Using Multimodal Musculoskeletal Ultrasound and Clinical Features.

Objectives

This study aimed to develop and validate an interpretable machine learning model for osteoporosis prediction based on multimodal ultrasound and clinical features.

Methods

Participants with suspected osteoporosis were enrolled from 2 medical centers. Clinical variables and multimodal musculoskeletal ultrasound parameters, including muscle thickness, cross-sectional area, and shear wave elastography measurements, were collected. LASSO regression was used for feature selection. Eight machine learning algorithms, including logistic regression, decision tree, random forest, k-nearest neighbors, support vector machine, naïve Bayes, neural network, and XGBoost, were developed and compared. Model performance was evaluated in the training, internal validation, and external test cohorts using the area under curve (AUC), accuracy, and F1 score. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).

Results

The LASSO regression demonstrated that age, gastrocnemius medial head's muscle thickness (GmhMT), rectus femoris' cross-sectional area (RFCSA), gender, body mass index (BMI), and TibialisSWE were key features for osteoporosis. The XGBoost model demonstrated accuracy of 0.889, 0.825, and 0.773 in the training, validating, and testing sets. SHAP analysis revealed the importance ranking of factors as age, GmhMT, RFCSA, gender, BMI, and TibialisSWE. Personalized prediction explanations through SHAP values demonstrated the contribution of each feature to the final prediction, enhancing result interpretability.

Conclusion

An interpretable machine learning model based on multimodal musculoskeletal ultrasound and clinical features demonstrated promising performance for osteoporosis prediction. That may provide a practical and explainable tool for early osteoporosis risk stratification and individualized assessment.