A multiparameter model (OsteoSC-M3) for early detection and risk stratification of osteoporosis in women: a multicentre cohort study in China.

Background: Osteoporosis is a major global health issue, but current early screening tools lack accuracy, and the gold-standard diagnostic method, dual-energy X-ray absorptiometry (DXA), is not widely accessible. This limits large-scale early screening efforts, as evidenced by China's low screening rate of 3.7% among adults aged over 50. To address these limitations, we developed OsteoSC-M3, a new risk assessment model for women using a stacked ensemble machine learning algorithm, designed to offer a precise and accessible tool for early risk identification. Methods: A model was developed based on a cohort of 5745 females (aged 50-90 years) from Shanghai Sixth People's Hospital (2013-2023); sex was self-reported. Participants were randomly partitioned into a Training Set (70%, n = 4022) and an Internal Validation Set (30%, n = 1723). First, missing values in the Training Set were imputed using the missForest method, and key predictive variables were selected through VIF, LASSO, and multivariable logistic regression. Nine base models and an ensemble model with XGBoost as the meta-learner were then constructed using the Training Set. The model was further validated on the Internal Validation Set and two independent multicentre external cohorts (Shanghai 2024 cohort, n = 1839; Jiangsu 2025 cohort, n = 305). Additionally, its prognostic value was assessed through a prospective study, which enrolled 405 female participants with low bone mass (T-score between -1.0 and -2.5) from the Health Examination Centre of Shanghai Sixth People's Hospital, with a follow-up period exceeding five years, where outcomes were evaluated using Kaplan-Meier survival analysis and Cox regression. This study is registered with the Chinese Clinical Trial Registry (part of the WHO International Clinical Trials Registry Platform), number ChiCTR2500097213. Findings: Eleven key variables, including age, height, weight, and eight serum biomarkers, were identified. All nine base models performed well without overfitting. The OsteoSC-M3 model achieved excellent diagnostic performance with a Training Set AUC of 0.973 (95% CI 0.969-0.976). It maintained excellent performance in the Internal Validation Set (AUC 0.943, 95% CI 0.933-0.954) and two external cohorts (Cohort 1: AUC 0.958, 95% CI 0.949-0.966; Cohort 2: AUC 0.917, 95% CI 0.897-0.928), demonstrating strong generalisability. Calibration curves showed good agreement, and decision curve analysis confirmed clinical utility. Prospectively, the model agreed well with follow-up results (Kappa = 0.82) and identified high-risk individuals 29 ± 9.31 months earlier (HR = 7.63, Interpretation: The OsteoSC-M3 model integrates demographic features with eight simple laboratory parameters to enable early, non-invasive prediction of osteoporosis risk in women, providing a 29-month lead time before clinical diagnosis. Superior to OSTA in risk stratification, it effectively identifies high-risk individuals requiring DXA confirmation and demonstrates strong utility in resource-limited settings without DXA equipment. This innovative screening approach creates a critical window for early intervention that could significantly improve participants' quality of life while reducing the economic burden on healthcare systems. Future research should focus on validating the model in more diverse ethnic populations and assessing its long-term impact on clinical outcomes through large-scale prospective trials. Funding: This study was funded by the Shanghai Pudong New Area Health Commission Joint Research Project, the Hospital Contractual Research Project, the Pudong New Area Science and Technology Development Fund, the Clinical Research Programme of Shanghai Sixth People's Hospital, and the Hospital's Clinical Backbone Team Cultivation Programme.