Scholarlink

Summary <p>This study used explainable AI to improve the Danish FREM model for predicting one-year risk of major osteoporotic fractures in over 2.4 million individuals aged ≥ 45. A DART boosting algorithm improved performance (AUC 0.77), with explainable outputs aiding clinical interpretation and guiding referrals for fracture risk assessment.</p> Purpose <p>This study aimed to use explainable artificial intelligence to improve the Fracture Risk Evaluation Model (FREM), in the prediction of imminent (one-year) risk of major osteoporotic fractures (MOFs).</p> Methods <p>FREM<sub>ML</sub> was trained and validated using complete registry data extracted for the Danish population ≥ 45&#xa0;years without previous osteoporosis diagnoses or treatment (N = 2,438,140). A Dropouts meet multiple Additive Regression Tree (DART) boosting algorithm was used. Predictors of MOFs (2022), automatically extracted for the 15-year lookback period (2007–2021), included hospital diagnoses, filled medication prescriptions, days since the last redemption of medications specific to fall and osteoporosis risk, as well as markers of polypharmacy and multi-morbidity. Stratified analyses were carried out, and model outputs were evaluated in the context of explainable artificial intelligence (AI).</p> Results <p>FREM<sub>ML</sub> displayed an overall area under the curve (95% confidence interval) of 0.77 (0.76, 0.77) – making it superior to previous versions of FREM. While age and sex were the most relevant predictors of MOF events, advanced feature engineering, including temporal information, contributed to model performance. Importantly, stratified analyses highlighted changing model performance across age groups and poorer prediction performance in males. Shapley Additive exPlanations values, a feature importance metric in explainable AI, facilitated clinical interpretation of relative MOF risk.</p> Conclusion <p>The publicly available FREM<sub>ML</sub> boosting model, combined with explainable AI, may be an effective decision support approach in a physician’s referral of individuals at high imminent risk of fractures to dual-energy X-ray absorptiometry.</p>

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

Introducing FREMML: a decision-support approach for automated identification of individuals at high imminent fracture risk

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月5日

查看全文
  • Marlene Rietz,
  • Jan C. Brønd,
  • Sören Möller,
  • Jens Søndergaard,
  • Bo Abrahamsen,
  • Katrine Hass Rubin

摘要

Summary

This study used explainable AI to improve the Danish FREM model for predicting one-year risk of major osteoporotic fractures in over 2.4 million individuals aged ≥ 45. A DART boosting algorithm improved performance (AUC 0.77), with explainable outputs aiding clinical interpretation and guiding referrals for fracture risk assessment.

Purpose

This study aimed to use explainable artificial intelligence to improve the Fracture Risk Evaluation Model (FREM), in the prediction of imminent (one-year) risk of major osteoporotic fractures (MOFs).

Methods

FREMML was trained and validated using complete registry data extracted for the Danish population ≥ 45 years without previous osteoporosis diagnoses or treatment (N = 2,438,140). A Dropouts meet multiple Additive Regression Tree (DART) boosting algorithm was used. Predictors of MOFs (2022), automatically extracted for the 15-year lookback period (2007–2021), included hospital diagnoses, filled medication prescriptions, days since the last redemption of medications specific to fall and osteoporosis risk, as well as markers of polypharmacy and multi-morbidity. Stratified analyses were carried out, and model outputs were evaluated in the context of explainable artificial intelligence (AI).

Results

FREMML displayed an overall area under the curve (95% confidence interval) of 0.77 (0.76, 0.77) – making it superior to previous versions of FREM. While age and sex were the most relevant predictors of MOF events, advanced feature engineering, including temporal information, contributed to model performance. Importantly, stratified analyses highlighted changing model performance across age groups and poorer prediction performance in males. Shapley Additive exPlanations values, a feature importance metric in explainable AI, facilitated clinical interpretation of relative MOF risk.

Conclusion

The publicly available FREMML boosting model, combined with explainable AI, may be an effective decision support approach in a physician’s referral of individuals at high imminent risk of fractures to dual-energy X-ray absorptiometry.