Scholarlink

<p>To develop and validate predictive models for osteoradionecrosis (ORN) after head and neck radiation therapy (RT) using time-to-event data with death as the competing risk, and to quantify the degree of risk overestimation when the competing risk is ignored. In this prognostic study of patients who underwent curative RT between 2011 and 2018, with ongoing follow-up, sociodemographic, clinical, and dosimetric data were collected. The binary ORN outcome was defined by the ClinRad system (grade ≥ 1); all-cause mortality was the competing event. Fine-Gray regression (FGR), Random Survival Forests (RSF) with Gray’s test splitting rule, and DeepHit were implemented using repeated nested stratified cross-validation. Feature selection and interpretation were guided by SHapley Additive exPlanations (SHAP). For comparison, non-competing risk models such as Cox proportional hazards (Cox PH) and standard RSF (S-RSF) with log-rank splitting rule were also trained. Of 2,466 patients, 183 developed ORN during follow-up, and 714 died. Three versions of each model were developed using 20, 10, and 5 features. The 10- and 5-feature RSF models performed best. Considering simplicity, the 5-feature model, which included tumor site, D10cc, smoking pack-years, periodontal condition, and dental insurance, was selected for production. At 60 months, Brier Score was 0.061 (95% CI: 0.060–0.063), Integrated Brier Score 0.038 (95% CI: 0.037–0.040), time-dependent AUC 0.776 (95% CI: 0.762–0.789), and C-index 0.772 (95% CI: 0.757–0.787). FGR closely followed, whereas DeepHit underperformed. Non-competing models, including the S-RSF, overestimated ORN risk, predicting an average 60-month cumulative incidence of 8.7% versus 6.8% with the 5-feature RSF. A parsimonious RSF model reliably estimated individualized ORN risk while avoiding overestimation from ignored competing risks. An interactive web application was developed to support clinical implementation.</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

Machine Learning Models for Individualized Osteoradionecrosis Risk Prediction in Head and Neck Cancer

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Mohammad Moharrami,
  • Erin Watson,
  • Shao Hui Huang,
  • Sreenath Madathil,
  • John Kim,
  • Andrew McPartlin,
  • Nauman H. Malik,
  • Sonica Singhal,
  • Ezra Hahn,
  • John Waldron,
  • Scott Bratman,
  • John de Almeida,
  • Christopher Yao,
  • Andrew Hope,
  • Carlos Quinonez,
  • Michael Glogauer,
  • Ali Hosni

摘要

To develop and validate predictive models for osteoradionecrosis (ORN) after head and neck radiation therapy (RT) using time-to-event data with death as the competing risk, and to quantify the degree of risk overestimation when the competing risk is ignored. In this prognostic study of patients who underwent curative RT between 2011 and 2018, with ongoing follow-up, sociodemographic, clinical, and dosimetric data were collected. The binary ORN outcome was defined by the ClinRad system (grade ≥ 1); all-cause mortality was the competing event. Fine-Gray regression (FGR), Random Survival Forests (RSF) with Gray’s test splitting rule, and DeepHit were implemented using repeated nested stratified cross-validation. Feature selection and interpretation were guided by SHapley Additive exPlanations (SHAP). For comparison, non-competing risk models such as Cox proportional hazards (Cox PH) and standard RSF (S-RSF) with log-rank splitting rule were also trained. Of 2,466 patients, 183 developed ORN during follow-up, and 714 died. Three versions of each model were developed using 20, 10, and 5 features. The 10- and 5-feature RSF models performed best. Considering simplicity, the 5-feature model, which included tumor site, D10cc, smoking pack-years, periodontal condition, and dental insurance, was selected for production. At 60 months, Brier Score was 0.061 (95% CI: 0.060–0.063), Integrated Brier Score 0.038 (95% CI: 0.037–0.040), time-dependent AUC 0.776 (95% CI: 0.762–0.789), and C-index 0.772 (95% CI: 0.757–0.787). FGR closely followed, whereas DeepHit underperformed. Non-competing models, including the S-RSF, overestimated ORN risk, predicting an average 60-month cumulative incidence of 8.7% versus 6.8% with the 5-feature RSF. A parsimonious RSF model reliably estimated individualized ORN risk while avoiding overestimation from ignored competing risks. An interactive web application was developed to support clinical implementation.