A primary care-oriented framework for osteoporosis prediction: integrating multi-algorithm feature selection and machine learning model benchmarking

Summary Large-scale screening and primary care diagnosis of osteoporosis are challenging due to the limited availability of dual-energy X-ray absorptiometry (DXA). Therefore, predictive models for osteoporosis that do not rely on bone mineral density measurements are essential. Using routine health examination data and applying a consensus feature selection strategy, we constructed machine learning (ML)-based models and compared their performance across multiple dimensions. Purpose To develop a robust, low-cost, and interpretable framework for osteoporosis prediction based on readily available information, which can be easily integrated into electronic health systems to inform DXA referral in resource-limited settings. Methods Variables covered by the basic health examination package were obtained from participants who underwent health examinations between January 2024 and January 2025. We implemented a strategy in which only features selected by at least four out of the seven feature selection methods were retained. Six ML algorithms were employed to develop models, and their prediction performances were compared across discrimination, goodness-of-fit, and clinical utility. Results A total of 2,260 participants were included, with an osteoporosis prevalence of 7.1%. Eight features: age, sex, body mass index, alkaline phosphatase, hematocrit, high-density lipoprotein, urea, and platelet count, were selected for model development. In the validation set, the random forest and artificial neural network (ANN) models demonstrated strongest screening performance, with AUCs of 0.895 (95% CI 0.867–0.923) and 0.890 (95% CI 0.862–0.919), respectively. The ANN model additionally exhibited superior calibration. Shapley additive explanations were applied to interpret the ANN model for osteoporosis prediction. Conclusion This study presents an easily integrable model for osteoporosis prediction based on routine health examination data. The ANN model provides robust prediction, offering a potential tool for pre-DXA triage and early identification of high-risk individuals in primary care and resource-limited settings.