基于机器学习的多种声门下喉癌预后预测模型

文献信息

中文摘要

背景： 声门下喉癌是喉癌中相对罕见的亚型，其病例数量有限，使得预后预测具有挑战性。本研究旨在比较多种声门下喉癌预后模型，并确定纳入其他喉亚部位的数据是否能提高对这一罕见亚型的预测性能。
方法： 从监测、流行病学和最终结果（SEER）数据库中提取了4,701例符合条件的喉癌病例。根据肿瘤位置，病例被分类为声门下喉（n=75）、声门喉（n=2,691）、声门上喉（n=1,514）和其他喉（n=421）。进一步创建了三个组合子集：子集1（声门下喉+声门喉）、子集2（声门下喉+声门上喉）和子集3（声门下喉+其他喉）。开发了四种预后模型，包括逐步Cox回归、最小绝对收缩和选择算子（LASSO）Cox回归、随机生存森林（RSF）和深度生存（DeepSurv）。使用一致性指数（C-index）以及5年和10年生存的时间依赖性曲线下面积（AUC）评估声门下病例的模型性能。
结果： 对于在完整队列上训练的模型，声门下病例的5年/10年AUC分别为：逐步Cox为0.748/0.676，LASSO Cox为0.666/0.677，RSF为0.716/0.665；DeepSurv模型的C-index为0.690。在子集1中，相应值分别为0.739/0.680、0.663/0.671和0.716/0.626，DeepSurv的C-index为0.624。在子集2中，相应值分别为0.679/0.620、0.698/0.756和0.713/0.792，DeepSurv的C-index为0.701。在子集3中，相应值分别为0.731/0.648、0.671/0.694和0.725/0.722，DeepSurv的C-index为0.672。总体而言，子集2在LASSO Cox和DeepSurv中表现最佳，而完整队列在逐步Cox中表现最佳，子集3在RSF中表现最佳。
结论： 多种预后模型在预测声门下喉癌结局方面显示出潜在效用。将声门上病例纳入训练队列可能会改善对这一罕见亚型的预后预测。在评估的方法中，逐步Cox模型在本研究条件下表现出最稳健和稳定的整体性能。

英文摘要

BACKGROUND: Subglottic laryngeal cancer is a relatively rare subtype of laryngeal cancer, and its limited case number makes prognostic prediction challenging. This study aimed to compare multiple prognostic models for subglottic laryngeal cancer and to determine whether incorporating data from other laryngeal subsites could improve predictive performance for this rare subtype.
METHODS: A total of 4,701 eligible laryngeal cancer cases were extracted from the Surveillance, Epidemiology, and End Results (SEER) database. According to tumor location, cases were classified into LarynxSubglottic (n=75), LarynxGlottic (n=2,691), LarynxSupraglottic (n=1,514), and LarynxOther (n=421). Three combined subsets were further created: subset 1 (LarynxSubglottic + LarynxGlottic), subset 2 (LarynxSubglottic + LarynxSupraglottic), and subset 3 (LarynxSubglottic + LarynxOther). Four prognostic models were developed, including stepwise Cox regression, least absolute shrinkage and selection operator (LASSO) Cox regression, random survival forest (RSF), and deep survival (DeepSurv). Model performance for subglottic cases was evaluated using the concordance index (C-index) and time-dependent areas under the curve (AUCs) for 5-year and 10-year survival.
RESULTS: For models trained on the complete cohort, the 5-year/10-year AUCs for subglottic cases were 0.748/0.676 for stepwise Cox, 0.666/0.677 for LASSO Cox, and 0.716/0.665 for RSF; the DeepSurv model achieved a C-index of 0.690. In subset 1, the corresponding values were 0.739/0.680, 0.663/0.671, and 0.716/0.626, with a DeepSurv C-index of 0.624. In subset 2, the values were 0.679/0.620, 0.698/0.756, and 0.713/0.792, with a DeepSurv C-index of 0.701. In subset 3, the values were 0.731/0.648, 0.671/0.694, and 0.725/0.722, with a DeepSurv C-index of 0.672. Overall, subset 2 showed the most favorable performance in LASSO Cox and DeepSurv, whereas the complete cohort performed best in stepwise Cox and subset 3 performed best in RSF.
CONCLUSIONS: Multiple prognostic models showed potential utility for predicting outcomes in subglottic laryngeal cancer. Incorporating supraglottic cases into the training cohort may improve prognostic prediction for this rare subtype. Among the evaluated approaches, the stepwise Cox model demonstrated the most robust and stable overall performance under the present study conditions.