上海口腔医学

上海口腔医学 ›› 2025, Vol. 34 ›› Issue (4): 362-368.doi: 10.19439/j.sjos.2025.04.004

• 论著 • 上一篇    下一篇

注意力机制对口腔癌竞争事件生存分析的探索与应用

金路1,2, 张睿1, 姒蜜思1, 高尚2,*, 陈谦明1,*   

  1. 1.浙江大学医学院附属口腔医院·浙江大学口腔医学院, 浙江省口腔疾病临床医学研究中心, 浙江省口腔生物医学研究重点实验室, 浙江大学癌症研究院, 口腔生物材料与器械浙江省工程研究中心, 浙江 杭州 310000;
    2.江苏科技大学计算机学院, 江苏 镇江 212100
  • 收稿日期:2024-03-01 修回日期:2024-04-16 出版日期:2025-08-25 发布日期:2025-08-26
  • 通讯作者: 高尚,E-mail: gaoshang@sohu.com;陈谦明,E-mail: qmchen@zju.edu.cn。*共同通信作者
  • 作者简介:金路（1992-）,男,工程师,在读硕士研究生,E-mail: jinlu@zjkq.com.cn
  • 基金资助:
    浙江省卫生健康重大科技计划项目(WKJ-ZJ-2212)

Exploration and application of attention mechanism in survival analysis of competitive events in oral cancer

Jin Lu1,2, Zhang Rui2, Si Misi2, Gao Shang1, Chen Qianming2   

  1. 1. The Stomatology Hospital, Zhejiang University School of Medicine, Zhejiang Provincial Clinical Research Center for Oral Diseases, Key Laboratory of Oral Biomedical Research of Zhejiang Province, Cancer Center of Zhejiang University, Engineering Research Center of Oral Biomaterials, Devices of Zhejiang Province. Hangzhou 310000, Zhejiang Province;
    2. School of Computer, Jiangsu University of Science, Technology. Zhenjiang 212100, Jiangsu Province, China
  • Received:2024-03-01 Revised:2024-04-16 Online:2025-08-25 Published:2025-08-26

PDF (PC)

3

摘要： 目的：采用深度学习中的注意力机制构建口腔癌竞争事件生存分析（oral survival analysis with attention,OSAA）模型,并与其他多个基准模型进行比较,初步探索其对口腔癌辅助诊疗的应用价值。方法：选取美国国立癌症研究所监测、流行病学及结局项目（Surveillance, Epidemiology, and End Results Program, SEER）数据库中符合条件的口腔癌患者数据作为研究对象,建立、训练Cox比例风险模型、基于深度学习的生存分析模型（DeepSurv、DeepHit等）及OSAA模型并进行预测,通过一致性指数(concordance index, C-index)和综合布尔里分数(integrated Brier score, IBS)测试各模型的预测效果,结合Kaplan-Meier生存曲线及时间依赖的受试者操作特征 (receiver operating characteristic, ROC) 曲线等对模型区分度进行检验。结果：OSAA在不同的数据集上都具有较高的C-index和较低的IBS,生存曲线及ROC曲线较其他模型具有更高的区分度。结论：OSAA模型具备优于其他模型的预测性能,在不同数据集和任务下有更好的稳健性和泛化能力,对建立以口腔癌为代表的口腔疾病辅助诊疗模型具有一定价值。

关键词: 注意力机制, 口腔癌, 竞争事件, 生存分析, 辅助诊疗

Abstract: PURPOSE: This study constructed a model of OSAA (oral survival analysis with attention) for survival analysis of competitive events in oral cancer based on attention mechanism, and explored its application value in oral auxiliary diagnosis and treatment of oral cancer. METHODS: Eligible data of oral cancer patient from Surveillance, Epidemiology, and End Results Program(SEER) database were selected as research subjects. Cox proportional hazards models, deep learning-based survival analysis models (such as DeepSurv, DeepHit), and OSAA models were established and trained for prediction. The predictive performance of each model was tested through concordance index (C-index) and integrated Brier score (IBS) test. The model's discriminative ability was evaluated using the Kaplan-Meier survival curve and the time-dependent receiver operating characteristic (ROC) curve. RESULTS: OSAA demonstrated ahigher C-index and a lower IBS on different datasets, with more distinct survival and ROC curves compared to other models. CONCLUSIONS: The OSAA model exhibits superior predictive performance compared to other models, with better robustness and generalization ability under different datasets and tasks. It has a certain value for establishing auxiliary diagnosis and treatment models for oral diseases represented by oral cancer.

Key words: Attention mechanism, Oral cancer, Competing events, Survival analysis, Auxiliary diagnosis and treatment

中图分类号: