文献速递头颈部肿瘤放疗放疗物理与技术

机器学习/人工智能模型预测头颈癌放射性毒性的实际效能:系统综述与荟萃分析

The actual performance of ML/AI models in predicting radiation-induced toxicity in head and neck cancer: a systematic review and meta-analysis.
PubMed 文献 · Radiother Oncol · 2025-12-18

该系统综述与荟萃分析共纳入67项研究、568个模型,评估机器学习/人工智能模型预测头颈癌患者放射性毒性的准确性及方法学质量。汇总AUROC为0.76(95%CI 0.73-0.78),提示中等判别能力,但研究间异质性大,影像生物标志物可显著提升模型表现。回顾性设计为主和预测因子选择差异可能引入偏倚,影响模型可靠性与泛化性。

要点

  • ML/AI模型预测头颈癌放射性毒性汇总AUROC为0.76,判别能力中等。
  • 纳入影像生物标志物可显著改善模型预测表现。
  • 回顾性设计为主和预测因子选择差异可能引入偏倚,限制泛化性。

核心结论

  • ML/AI模型对头颈癌患者放射性毒性具有中等判别能力,汇总AUROC为0.76(95%CI 0.73-0.78)。
  • 方法学限制影响模型可靠性与泛化性,需标准化透明报告模型开发与验证过程。

研究设计与样本

  • 系统综述与荟萃分析,检索PubMed、EMBASE、Web of Science和Cochrane Library。
  • 纳入67项研究、共568个模型,采用PROBAST工具评估偏倚风险,行分层多水平荟萃分析估计汇总AUROC。

主要发现

  • ML/AI模型汇总AUROC为0.76(95%CI 0.73-0.78),研究间存在显著异质性。
  • 纳入影像生物标志物显著改善模型表现;前瞻性与内部验证表现相当,外部验证反映真实泛化性。

关键数值与推荐

  • 汇总AUROC=0.76(95%CI 0.73-0.78),对应头颈癌放射性毒性预测模型。
  • 纳入影像生物标志物可显著改善模型性能;前瞻性与内部验证表现相当,外部验证体现真实泛化性。

临床怎么用

  • ML/AI模型可用于头颈癌放射性毒性风险预测的辅助评估,但需警惕方法学异质性带来的不确定性。
  • 临床应优先关注经外部验证的模型,并推动模型开发与验证过程的标准化透明报告。

局限性

  • 回顾性设计为主和预测因子选择差异可能引入偏倚,影响模型可靠性与泛化性。
  • 研究间存在显著异质性,方法学限制制约模型的临床应用。

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

An increasing number of Artificial intelligence (AI) and machine learning (ML) models are being developed to predict radiation-induced toxicities (RITs) in patients with head and neck cancer (HNC). But their performance and reliability remain uncertain. This systematic review and meta-analysis evaluated the predictive accuracy and methodological quality of these models. We comprehensively searched PubMed, EMBASE, Web of Science, and the Cochrane Library to identify studies reporting on ML/AI models for predicting RITs in HNC patients. Eligible studies were assessed for bias risk using the PROBAST tool, and key performance metrics, including the area under the receiver operating curve (AUROC), were extracted. A hierarchical multilevel meta-analysis was performed to estimate pooled AUROC values, and subgroup analyses explored the influence of study characteristics on model performance. A total of 67 studies with a total of 568 models were included, showing moderate discriminatory power of ML/AI models, with a pooled AUROC = 0.76; 95 % CI: 0.73-0.78. Nonetheless, substantial heterogeneity was observed across studies. Incorporating imaging biomarkers significantly improved model performance. Prospective and internal validation showed comparable performance; external validation shows true generalizability. The predominance of retrospective designs and variability in predictor selection may have introduced bias, affecting model reliability and generalisability. ML/AI models hold promise for predicting RITs in HNC patients, but methodological constraints limit their applicability. Standardised and transparent reporting of model development and validation processes is vital for improving comparability among studies. Future research should explore hybrid modelling methods and the integration of clinical, dosimetric, radiomic, and genomic data to boost predictive accuracy.

来源:Radiother Oncol,查看原文。版权归原作者所有。

引用

  1. Ugwu GC, Jalali F, Liu G, et al. The actual performance of ML/AI models in predicting radiation-induced toxicity in head and neck cancer: a systematic review and meta-analysis. Radiother Oncol. 2025. PMID: 41421773. doi:10.1016/j.radonc.2025.111350

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。