Interpretable Machine Learning Approach for Osteoporosis Risk Predictive Assessment with Mult-strategy Learning.

Osteoporosis is a medical condition that makes bones weak, fragile, and easy to break. It is caused by estrogen deficiency, which leads to increased osteoclast differentiation and activation, accelerated bone resorption that outpaces formation, and rapid bone loss. With the growing global burden of osteoporosis and age-related fragility fractures, there is an increasing need for accessible and reliable approaches for early risk assessment. Osteoporosis affects hundreds of millions of individuals worldwide and is associated with substantial morbidity, reduced quality of life, and healthcare costs due to fracture-related complications [1]. However, diagnosis remains challenging because dual-energy X-ray absorptiometry (DEXA), the gold-standard method for assessing bone mineral density, requires specialized equipment and infrastructure that may not be widely available, particularly in resource-limited settings [2]. In countries such as Uganda where specialized diagnostics are limited and largely concentrated in urban centers(Kampala), it leads to delayed identification and management of individuals at risk creating substantial barriers to early diagnosis. Furthermore, the traditional osteoporosis risk assessment approaches may not fully capture complex interactions among demographic, clinical, and lifestyle factors. This study develops machine learning, deep learning, and a hybrid multi-strategy ensemble that combines gradient boosting with attention-based deep learning approaches for osteoporosis risk prediction using routinely collected demographic, clinical, and lifestyle characteristics of patients.The objective is to identify individuals at elevated risk of osteoporosis using routinely available patient information. Multiple machine learning , deep learning and hybrid multi-strategy ensemble models were developed and evaluated for osteoporosis risk prediction, including Logistic Regression, Random Forest, XGBoost, LightGBM, and TabNet. Model performance was assessed using stratified 5-fold cross-validation and an independent test set based on accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). LightGBM out performed other models yielding the highest cross-validation performance with an AUROC of 0.917, while XGBoost demonstrated the best generalization on the independent test set with an AUROC of 0.893. To improve predictive performance and robustness, a hybrid multi-strategy ensemble integrating the LightGBM and TabNet models was developed yielding an accuracy of 87.2%, precision of 95.1%, recall of 78.6%, F1-score of 86.0%, and AUROC of 0.893, demonstrating competitive performance through complementary learning strategies. Explainable AI techniques, including SHAP and LIME, were applied to assess feature representation and model interpretability. The explainability analysis identified age as the most influential predictor, followed by fracture history, hormonal status, body weight, lifestyle factors, and medication use, demonstrating that the proposed framework learns clinically meaningful patterns while providing interpretable decision support for osteoporosis risk assessment in resource-limited settings. [1] International Osteoporosis Foundation. Facts and Statistics: Epidemiology of Osteoporosis and Fragility Fractures. International Osteoporosis Foundation. https://www.osteoporosis.foundation/facts-statistics/epidemiology-of-osteoporosis-and-fragility-fractures [2] National Osteoporosis Foundation. Clinician’s Guide to Prevention and Treatment of Osteoporosis. Osteoporosis International. 2022.