Page 78 - 南京医科大学自然版

南京医科大学学报（自然科学版） 第46卷第9期
·1348 · Journal of Nanjing Medical University（Natural Sciences） 2026年9月

·临床研究·

基于临床及剂量学特征构建模型预测鼻咽癌放疗后颞叶损伤的

研究

万丽娟 ，李金凯 ，苏国义 ，胡 昊 ，唐艳春 ，吴飞云 2*
2
1
2
1
1
南京医科大学第一附属医院放疗科，放射科，江苏 南京 210029
1 2
［摘 要］ 目的：基于鼻咽癌患者的临床及剂量学特征构建模型，并评估其对放疗后出现放射性颞叶损伤（temporal lobe inju⁃
ry，TLI）的预测价值。方法：回顾性分析2018年6月—2023年11月365例（包括TLI组42例和无TLI组323例）鼻咽癌患者的临
床资料及剂量学特征，使用单因素及多因素Logistic 回归分析筛选出与放射性TLI相关的独立预测因素。利用随机森林机器
学习算法构建临床特征模型及临床+剂量学特征联合模型。绘制受试者操作特征（receiver operating characteristic，ROC）曲线，
计算曲线下面积（area under curve，AUC）、灵敏度、特异度等参数，绘制校准曲线及决策曲线分析评估模型的预测效能。结果：
通过单因素、多因素Logistic回归分析得出临床特征中的T分期及剂量学特征中的 D 3 、V70 Gy这3个特征的差异有统计学意义
1 cm
（P < 0.05）。联合模型和临床模型的 AUC 值、灵敏度、特异度和准确率分别为 0.853 和 0.635、66.67%和 85.71%、86.38%和
39.94%、84.00%和45.00%。DeLong检验显示，联合模型与临床模型的预测效能差异具有统计学意义（P < 0.05）。校准曲线和
决策曲线分析结果均表明联合模型具有更高的校准度和临床净获益。结论：T分期、 D 3 、V70 Gy是放射性TLI的独立影响因素，
1 cm
基于三者的联合模型对鼻咽癌放疗后TLI具有较高的预测能力。
［关键词］ 鼻咽癌；放射性颞叶损伤；机器学习；预测模型
［中图分类号］ R739.62 ［文献标志码］ A ［文章编号］ 1007⁃4368（2026）09⁃1348⁃08
doi：10.7655/NYDXBNSN260364

Research on constructing a model based on clinical and dosimetric characteristics to
predict temporal lobe injury after radiotherapy for nasopharyngeal carcinoma
2
1
2
1
1
WAN Lijuan ，LI Jinkai ，SU Guoyi ，HU Hao ，TANG Yanchun ，WU Feiyun 2*
Department of Radiation Oncology，Department of Radiology，the First Affiliated Hospital of Nanjing Medical
1 2
University，Nanjing 210029，China
［Abstract］ Objective：To establish a prediction model based on the clinical and dosimetric characteristics of patients with
nasopharyngeal carcinoma（NPC），and assess its predictive efficacy for radiation ⁃ induced temporal lobe injury（TLI）following
radiotherapy. Methods：We performed a retrospective analysis of the clinical data from 365 NPC patients（42 patients in the TLI group
and 323 patients in the non⁃TLI group）who underwent radiotherapy⁃based comprehensive treatment in the radiotherapy department of
our institution from June 2018 to November 2023. Clinical and dosimetric characteristics were collected，and univariate and
multivariate logistic regression analyses were employed to identify independent risk factors associated with radiation⁃induced TLI. The
random forest（RF）machine learning algorithm was used to construct a clinical feature model and a combined clinical and dosimetric
feature model. Receiver operating characteristic（ROC）curves were plotted，and parameters such as the area under the ROC curve
（AUC），sensitivity，and specificity were calculated. Calibration curves and decision curve analysis（DCA）were also conducted to
evaluate the predictive performance of the models. Results：Univariate and multivariate logistic regression analyses revealed
statistically significant differences in three features：T ⁃ stage（clinical characteristic），as well as D 3 and V70Gy （dosimetric
1 cm
characteristics）（all P < 0.05）. The AUC values，sensitivity，specificity，and accuracy of the combined model and clinical model were
0.853 and 0.635，66.67% and 85.71%，86.38% and 39.94%，and 84.00% and 45.00%，respectively. The DeLong test indicated a

［基金项目］ 国家重点研发计划（2022YFC2401604）
通信作者（Corresponding author），E⁃mail：wfy_njmu@163.com（ORCID：0000⁃0002⁃0343⁃0458）
∗