不确定性量化融合集成模型用于食管癌患者放疗所致重度免疫抑制的个性化预测
Uncertainty Quantification-Incorporated Ensemble Model for Personalized Prediction of Severe Radiotherapy-Induced Immunosuppression Among Esophageal Cancer Patients
- The University of Texas MD Anderson Cancer Center（德克萨斯大学安德森癌症中心）
- University of Texas Health Science Center（德克萨斯大学健康科学中心）
- Amsterdam UMC（阿姆斯特丹大学医学中心）
- Massachusetts General Hospital（麻省总医院）
- University of Washington（华盛顿大学）
- Institute for Data Science in Oncology（肿瘤数据科学研究所）
- Harvard Medical School（哈佛医学院）
- Telperian
- Texas A&M University（德克萨斯农工大学）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本研究构建融合不确定性的集成机器学习模型，利用临床和剂量学变量预测食管癌患者放疗后ALC最低点及G4RIL风险，并通过CRCP量化不确定性，实现风险感知的个性化决策。
AI中文摘要：
重度放射性淋巴细胞减少症（RIL），特别是4级RIL（G4RIL），是接受同步放化疗的食管癌患者中常见的免疫毒性反应，也是预后不良的强预测因子。本研究开发了一种融合不确定性的集成机器学习模型，利用基线和剂量学变量预测绝对淋巴细胞计数（ALC）最低点，以支持质子治疗选择及剂量分布优化等治疗决策。研究分析了2001年7月至2022年4月期间接受光子或质子治疗的1,491例食管癌患者。利用临床特征和剂量学变量（包括由辐射剂量分布导出的复合剂量学评分（CDS））构建了加权集成模型。为量化个体水平的预测不确定性，开发了交叉残差共形预测（CRCP）方法，并与标准共形预测方法进行了对比评估。质子治疗患者的平均ALC最低点显著高于光子治疗患者（0.33对0.25 K/uL）。集成模型对ALC最低点预测的平均绝对误差为0.0926 K/uL，G4RIL分类的ROC-AUC为0.78。基线ALC、计划靶区体积和CDS是最具影响力的预测因子。CRCP在具有数学保证的覆盖率下实现了具有竞争力的预测区间宽度，在区间宽度与覆盖准确性之间取得了良好平衡。该框架能够支持放射肿瘤学中具有风险意识的个性化临床决策。
英文摘要：
Severe radiation-induced lymphopenia (RIL), specifically Grade 4 RIL (G4RIL), is a frequent immune toxicity and a strong predictor of poor survival for esophageal cancer patients undergoing concurrent chemoradiation therapy. This study developed an uncertainty-incorporated ensemble machine learning model to predict Absolute Lymphocyte Count (ALC) nadir using baseline clinical and dosimetric variables, supporting treatment decisions such as proton therapy selection and dose distribution optimization. A cohort of 1,491 esophageal cancer patients treated with photon or proton therapy between July 2001 and April 2022 was analyzed. A weighted ensemble model was constructed using clinical features and dosimetric variables, including the composite dosimetric score (CDS) derived from radiation dose distributions. To quantify individual-level predictive uncertainty, Cross-Residual Conformal Prediction (CRCP) was developed and evaluated against standard conformal prediction methods. Proton therapy patients maintained a significantly higher mean ALC nadir than photon patients (0.33 vs. 0.25 K/uL). The ensemble model achieved a mean absolute error of 0.0926 K/uL for ALC nadir prediction and an ROC-AUC of 0.78 for G4RIL classification. Baseline ALC, Planning Target Volume, and CDS were the most influential predictors. CRCP achieved competitive prediction interval widths with mathematically guaranteed coverage, offering a favorable balance between interval width and coverage accuracy. This framework enables risk-aware, personalized clinical decision-making in radiation oncology.
