Scholarlink

Osteoporotic fractures pose a major healthcare burden, particularly among aging populations. Traditional tools like FRAX estimate 10-year fracture risk but miss Imminent Fracture Risk (IFR), defined as fractures within two years. We trained machine learning models, including ensemble methods, on data from 2,949 patients in Ontario’s Fracture Screening and Prevention Program. Two-year performance was modest (PR-AUC \(\approx \) 0.45, representing an 80% relative lift over the 0.25 prevalence baseline), but remained constrained by class imbalance and missing predictors. Extending prediction to five years increased PR-AUC to \(\approx \) 0.60, an 18% lift over the 0.51 baseline; post-split SMOTE and downsampling yielded minor improvement. SHapley Additive exPlanations (SHAP) analysis identified fall history, prior fractures, and age as the most influential yet context-dependent features. Although not yet ready for clinical deployment, this work establishes a foundation for future IFR prediction in healthcare, highlighting the need for richer data collection with scheduled follow-up to capture censoring information.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Advancing Imminent Fracture Risk Prediction: Integrating Machine Learning with Enhanced Feature Engineering

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohammad Maghsoudimehrabani,
  • Edward R. Sykes

摘要

Osteoporotic fractures pose a major healthcare burden, particularly among aging populations. Traditional tools like FRAX estimate 10-year fracture risk but miss Imminent Fracture Risk (IFR), defined as fractures within two years. We trained machine learning models, including ensemble methods, on data from 2,949 patients in Ontario’s Fracture Screening and Prevention Program. Two-year performance was modest (PR-AUC \(\approx \) 0.45, representing an 80% relative lift over the 0.25 prevalence baseline), but remained constrained by class imbalance and missing predictors. Extending prediction to five years increased PR-AUC to \(\approx \) 0.60, an 18% lift over the 0.51 baseline; post-split SMOTE and downsampling yielded minor improvement. SHapley Additive exPlanations (SHAP) analysis identified fall history, prior fractures, and age as the most influential yet context-dependent features. Although not yet ready for clinical deployment, this work establishes a foundation for future IFR prediction in healthcare, highlighting the need for richer data collection with scheduled follow-up to capture censoring information.