An explainable machine learning approach to predict fragility fractures and the identification of important features
Sayem Borhan1,2,3, Alexandra Papaioannou4,5,6, Jonathan Adachi5
1Department of Health Research Methods, Evidence, and Impact, McMaster University, Martha Wing, H321, 50 Charlton Avenue East, Hamilton, ON, L8N 1Y3, Canada. borhana@mcmaster.ca.
None:
In this study, we developed ML algorithms to predict fragility fractures, considering the occurrence of fractures at different skeletal sites, using the data from the Canadian Multicentre Osteoporosis Study (CaMos) with participants aged 50 years or older. We considered 73 baseline features, and the outcome was the first incidence of fracture at any of the following sites: hip, spine, pelvis, ribs, shoulder, and forearm. The ML algorithms were evaluated in terms of the ROC_AUC. SHapley Additive exPlanations (SHAP) analysis was performed to identify the important features and to investigate the interaction among these features. In total, 7753 subjects were included in the study. Approximately 72% were female, and the average age was 67 years. We found that the XGBoost algorithm had a slightly better ROC_AUC (0.70; 95% CI 0.67, 0.73). From the SHAP analysis, we found that BMD was the most important feature and the total hip BMD interacted the most with femoral neck BMD. This study demonstrated that XGBoost was a marginally superior ML algorithm for predicting fragility fractures. In addition, we identified important features that contribute to the prediction of fragility fractures. Intervention focusing on these features will help to prevent the incidence of these fractures.