Interpretable machine learning classifiers for the reliable prediction of fall induced hip fracture risk

Classical computing methods are costly and require advanced skills, limiting their clinical use. A data-driven framework offers an effective alternative for disease diagnosis and prediction. This study aims to apply and evaluate machine learning (ML) classifiers to predict hip fracture risk using a binary classification based on the fracture risk index (FRI) from Quantitative Computed Tomography-based Finite Element Analysis (QCT-based FEA). This study concentrated on comparing the performance of different ML models such as logistic regression, Support Vector Classifier (SVC), Decision Tree (DT), Categorical Boosting Model (CatBoost), Extreme Gradient Boost Model (XGBM), and Random Forest (RF) for the prediction of hip fracture probability. The models were trained with a dataset comprises clinical parameters, bone anatomy, and loading directions mimicking sideways fall postures. All the ML models were compared based on the performance metrics—precision, recall, F1 score, accuracy, and Area Under Receiver Operating Curve (AUROC). Both logistic regression and SVC exhibited the highest performance in assessing fracture risk with 82% accuracy and 87% AUROC. These models are also interpretable, as we used SHapley Additive exPlanations (SHAP) to identify the most important features and their impacts on the prediction process. Despite being trained on a limited dataset, this study demonstrates the viability of machine learning models in predicting hip fracture risk. To the best of our knowledge, these models are the first interpretable ML-based predictors of hip fracture risk.