Hybrid Machine Learning Models and FRAX Tool for Comprehensive Osteoporosis and Fragility Fracture Risk Prediction

Abstract

Osteoporosis is a global bone disease characterized by decreased bone mineral density and impaired bone microarchitecture, significantly increasing risk of fragility fractures. Although the standard-imaging technique remains Dual-energy X-ray-Absorptiometry (DXA), there are issue that hinder its widespread adoption. In order to solve this issue, this work combined machine learning classification techniques with a Fracture Risk Assessment Tool (FRAX)-like risk scoring framework with long-term goal of providing an end-to-end and scalable prediction solution of present osteoporosis status along with future fracture risk. The used Data is 1,958 evaluated patient records. The used methodology included preprocessing, which consisted of normalization, one-hot coding, and SMOTE-based class balancing. Three machine learning (ML) models—decision tree (DT), random forest (RF), and multilayer perceptron (MLP)—were trained to predict osteoporosis. Using clinical risk predictors as variables and risk ratios derived from scientific studies as weights, a rule-based model was constructed to predict fracture risk. Cases were divided into low, intermediate, and high-risk categories based on the resulting risk scores. DT had the highest performance, with 88.4% accuracy, precision 1.00, recall 0.767, F1-score 0.868, and Area Under the Receiver Operating characteristic Curve (AUROC, ROC-AUC or AUC-ROC) 0.889. The model motivated by FRAX estimated 51.5%, 38.9%, and 9.7% of subjects as being in the Low, Medium, and High-risk categories, respectively. The prevalence of osteoporosis in each category was near 35%, 58%, and near 100%, respectively the latter with very good agreement of estimated risk with diagnosed condition. Combining predictive power of ML with the interpretability of a FRAX-like risk tool results in a powerful decision support tool for prevalent osteoporosis and high future fragility fracture risk.

Ähnliche Arbeiten