文献速递头颈部肿瘤放疗放疗物理与技术

基于SHAP可解释机器学习的纵向delta影像组学预测头颈癌放疗后口干症

SHAP-based interpretable machine learning with longitudinal delta-radiomics across seven weeks of treatment for xerostomia prediction in head-and-neck cancer.
PubMed 文献 · Radiat Oncol · 2026-04-23

本研究回顾性分析131例接受VMAT放化疗的头颈癌患者,利用每周CBCT提取腮腺delta影像组学特征,结合7种可解释机器学习分类器预测放疗后口干症(RIXM)。结果显示第4周特征与6个月终点预测最强,SVM的AUC为0.79,NTCP模型AUC为0.82,SHAP识别出合并症、delta影像组学特征及DmeanRT_PG为显著预测因子。

要点

  • 每周CBCT衍生的腮腺delta影像组学可动态捕捉治疗中变化。
  • 第4周特征对6个月RIXM预测性能最强,支持早期干预。
  • SHAP识别的显著预测因子包括合并症、delta影像组学特征和DmeanRT_PG。

核心结论

  • Delta影像组学联合可解释分类器可准确预测头颈癌患者放疗后3、6、12个月的口干症。
  • 第4周和6个月的特征预测能力最强,支持早期、个体化、基于剂量信息的放疗干预。

研究设计与样本

  • 回顾性队列研究,纳入2020至2023年接受VMAT放化疗(60-70 Gy,6-7周)的131例头颈癌患者(男101例,女30例;年龄23-87岁)。
  • 按患者层面分层拆分训练/验证集(n=91,69.5%)和测试集(n=40,30.5%)。
  • 从每周CBCT腮腺分割中按IBSI指南提取影像组学特征(每例6-7次),计算delta影像组学,采用单变量过滤与交叉验证递归特征消除混合方法筛选特征,并应用SMOTE。
  • 开发7种机器学习分类器和剂量学NTCP模型,使用SHAP评估模型可解释性,配对t检验、PPV和NPV评估特征显著性与预测性能。

主要发现

  • 3个月、6个月和12个月的口干症发生率分别为90.8%、91.6%和77.1%。
  • SVM预测性能最高(AUC: 0.79 ± 0.01, 95% CI: 0.65-0.83; PR-AUC: 0.98; 敏感度: 0.92, CI: 0.90-0.92; 特异度: 0.69, CI: 0.67-0.70; F1-score: 0.90 ± 0.01; BAcc: 0.81; MCC: 0.51; Brier score: 0.11 ± 0.01),PPV为0.81 ± 0.04,NPV为0.64 ± 0.02。
  • NTCP模型结果相似(AUC: 0.82 ± 0.01, 95% CI: 0.67-0.89; PR-AUC: 0.97; 敏感度: 1.00, CI: 0.96-1.00; 特异度: 0.72, CI: 0.72-0.74; F1-score: 0.92 ± 0.01; BAcc: 0.86; MCC: 0.48; Brier score: 0.08 ± 0.01; PPV: 0.83 ± 0.01; NPV: 0.67 ± 0.02)。
  • SHAP分析识别出的显著预测因子包括合并症(mean: 1.204; p: 0.007)、delta影像组学特征strength-(IBSI:1 X9X)(mean: 1.179; p: 0.041)、histogram gradient-(IBSI: RHQZ)(mean: 1.021; p: 0.043)、intensity kurtosis-(IBSI: C317)(mean: 0.592; p: 0.045)和DmeanRT_PG(mean: 0.447; p: 0.045)。

临床怎么用

  • 可利用治疗第4周的CBCT delta影像组学特征早期识别6个月时口干症高风险患者。
  • SHAP识别的显著预测因子(合并症、delta影像组学特征、DmeanRT_PG)可辅助临床解释和个体化放疗决策。

局限性

  • 本研究为单中心回顾性队列,样本量有限(131例),需外部验证。
  • 口干症评估基于CTCAE v5.0分级,可能存在主观性。
  • Delta影像组学特征提取依赖CBCT图像质量和分割精度。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

BACKGROUND: Radiation-induced xerostomia (RIXM) remains a common toxicity in head-and-neck cancer (HNC) patients receiving radiotherapy, with currently used static dosimetric models exhibiting limited predictive accuracy. This study evaluates delta-radiomics from longitudinal VMAT-CBCT imaging, combined with explainable machine learning classifiers (MLCs), to capture treatment-related changes in the parotid gland across seven weeks and improve RIXM prediction. METHODS: A retrospective cohort of 131 HNC patients (101 males, 30 females; aged 23-87) treated with VMAT-based chemoradiotherapy (60-70 Gy over 6-7 weeks) from 2020 to 2023 was analyzed. RIXM was graded using CTCAE v5.0 at 3, 6, and 12-months post-treatment, representing early, intermediate, and late endpoints. Radiomic features were extracted from weekly CBCT-derived parotid segmentations (6-7 per patient) following IBSI guidelines, and delta-radiomics was computed across multiple preprocessing pipelines. The dataset was split into training/validation (n = 91; 69.5%) and testing (n = 40; 30.5%) sets using stratified patient-level splitting. A hybrid feature selection method combining univariate filtering and recursive feature elimination with cross-validation was applied, followed by SMOTE. Seven MLCs and a dosimetric NTCP model were developed using Python (v7.0.8, scikit-learn). Model interpretability was assessed with SHAP; feature significance and predictive performance were evaluated using paired t-tests, PPV, and NPV. RESULTS: Early, intermediate, and late RIXM occurred in 90.8% (3-months), 91.6% (6-months), and 77.1% (12-months) of patients, respectively. The strongest predictions were found for week-4 and 6-month endpoint. SVM achieved the highest predictive performance (AUC: 0.79 ± 0.01, 95% CI: 0.65-0.83; PR-AUC: 0.98; sensitivity: 0.92, CI: 0.90-0.92; specificity: 0.69, CI: 0.67-0.70; F1-score: 0.90 ± 0.01; BAcc: 0.81; MCC: 0.51; Brier score: 0.11 ± 0.01), with balanced PPV (0.81 ± 0.04) and NPV (0.64 ± 0.02). Similar results were observed with the NTCP model (AUC: 0.82 ± 0.01, 95% CI: 0.67-0.89; PR-AUC: 0.97; sensitivity: 1.00, CI: 0.96-1.00; specificity: 0.72, CI: 0.72-0.74; F1-score: 0.92 ± 0.01) and calibration (BAcc: 0.86; MCC: 0.48; Brier score: 0.08 ± 0.01; PPV: 0.83 ± 0.01; NPV: 0.67 ± 0.02). SHAP analysis identified comorbidities (mean: 1.204; p: 0.007), delta-radiomic features (strength-(IBSI:1 X9X) (mean: 1.179; p: 0.041), histogram gradient-(IBSI: RHQZ) (mean: 1.021; p: 0.043), intensity kurtosis-(IBSI: C317) (mean: 0.592; p: 0.045)), and DmeanRT_PG (mean: 0.447; p: 0.045), as significant predictors of RIXM. CONCLUSIONS: Delta-radiomics combined with explainable classifiers accurately predicted RIXM at 3, 6, and 12-months post-treatment. Week-4 and 6-month features were most predictive, supporting early, personalized, dose-informed radiotherapy interventions.

来源:Radiat Oncol,查看原文。版权归原作者所有。

引用

  1. Samson DO, Ismail M, Hanifa MAM, et al. SHAP-based interpretable machine learning with longitudinal delta-radiomics across seven weeks of treatment for xerostomia prediction in head-and-neck cancer. Radiat Oncol. 2026. PMID: 42026673. doi:10.1186/s13014-026-02846-5

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。