Prediction of repeat fragility fractures: Application of machine learning

Abstract

Abstract Background Despite the exceptional performance of machine learning (ML) in modern medical research, it is rarely used for repeated-fracture prediction in osteoporotic patients. In this study, we aim to evaluate the predictive capability of various ML models and introduce features that are more relevant to repeated fragility fracture in osteoporotic patients. Methods Fragility fracture patients who were referred to our Fracture Liaison Service were divided into the index fragility fracture (n = 905) and repeated fragility fracture groups (n = 195). Twenty-seven features were used for model training in males and females separately. The ML models included random forest, XGBoost, CatBoost, logistic regression, LightGBM, AdaBoost, multi-layer perceptron, and support vector machine. A 10-fold cross-validation approach was used to assess the performance of the models. Results In almost all the feature sets, CatBoost had the best performance with a maximum area under the curve and accuracy of 0.951 and 87% for the female group and 0.990 and 93.4% for the male group, respectively. Age, CRP, vitamin D3, creatinine, blood urea nitrogen (BUN), parathyroid hormone (PTH), femoral neck Z-score, menopause age, number of pregnancies, phosphorus, calcium, and body mass index had the highest contribution in the female group and CRP, femoral neck T-score, PTH, Hip T-score, BMI, BUN, creatinine, alkaline phosphatase, and spinal Z-score had the highest contribution in the male group. Conclusion ML models, particularly CatBoost, are promising tools for the prediction of repeat fragility fracture in osteoporotic patients. These models can help clinicians to implement personalized strategies to prevent repeat fragility fractures in the future.