An explainable predictive machine learning model of osteopenia for perimenopausal women based on clinical data: a retrospective single-center study.

Background: Osteoporosis is increasingly prevalent, yet early detection remains difficult. This study aimed to develop a machine learning (ML)-based model for identifying individuals at risk of osteopenia using clinical data. Methods: Female participants who were aged 45 years and above with complete femoral neck BMD data were included. A total of 1,108 participants were divided into training (70%) and test (30%) datasets. Various ML-based algorithms (random forest (RF), least absolute shrinkage and selection operator (LASSO), gradient boosting decision tree (GBDT) were used to identify key predictors of osteopenia, including clinical and biochemical markers. Model performance was evaluated based on area under the curve (AUC), accuracy, sensitivity, specificity, precision, and F1 score, with SHAP analysis used for feature importance. Results: Totally, 17 predictors were identified, with RF demonstrating the best performance (AUC: 0.978 in training, 0.933 in validation datasets). Key predictors of osteopenia included menopause, age, procollagen type I N-terminal propeptide (PINP), beta-crosslaps (β-CTX), height, and estimated glomerular filtration rate (eGFR). RF outperformed other models in both identify accuracy and clinical utility, with an AUC of 0.90 and AUPR of 0.93 for the top six features. A web-based calculator was developed for clinical use. Conclusion: The RF model can effectively identify osteopenia and could improve early detection of osteoporosis. This model, with its integration into clinical practice, has the potential to enhance patient outcomes and reduce osteoporosis-related risks.