国际肿瘤学杂志 ›› 2026, Vol. 53 ›› Issue (8): 476-484.doi: 10.3760/cma.j.cn371439-20250823-00077

• 论著 • 上一篇    下一篇

基于机器学习算法构建高龄结直肠癌患者术后生存预测模型

沈才路, 吴锐荣, 吴优, 葛晓松()   

  1. 江南大学附属医院肿瘤内科无锡 214100
  • 收稿日期:2025-08-23 出版日期:2026-08-08 发布日期:2026-07-21
  • 通讯作者: 葛晓松,Email: gexiaosong@qq.com
  • 作者简介:

    沈才路:研究过程实施、统计学分析、论文撰写;吴锐荣、吴优:研究过程实施、统计学分析、论文修改;葛晓松:论文修改、经费支持

  • 基金资助:
    国家自然科学基金(81502042);江苏省自然科学基金(BK20140171);江苏省卫生健康委员会指导性项目(Z2023020)

Construction of a postoperative survival prediction model for elderly patients with colorectal cancer based on machine learning algorithms

Shen Cailu, Wu Ruirong, Wu You, Ge Xiaosong()   

  1. Department of Medical OncologyAffiliated Hospital of Jiangnan UniversityWuxi 214100, China
  • Received:2025-08-23 Online:2026-08-08 Published:2026-07-21
  • Contact: Ge Xiaosong, Email: gexiaosong@qq.com
  • Supported by:
    National Natural Science Foundation of China(81502042);Natural Science Foundation of Jiangsu Province of China(BK20140171);Guiding Project of Jiangsu Provincial Health Commission(Z2023020)

摘要:

目的 基于机器学习算法构建适用于高龄结直肠癌(CRC)患者的术后生存预测模型,识别影响患者预后的关键因素。方法 利用美国监测、流行病学与最终结果数据库(SEER),提取2000—2019年3 576例接受手术治疗的高龄CRC患者临床资料。采用随机抽样方法按7∶3比例分为训练集(n=2 503)和验证集(n=1 073)。通过最小绝对值收敛和选择算子(LASSO)回归筛选影响肿瘤特异性生存期(CSS)的关键变量,分别构建LASSO-Cox模型和DeepSurv深度学习模型。采用一致性指数(C-index)、受试者操作特征(ROC)曲线、Brier评分等指标评估模型预测效能。运用沙普利加性解释(SHAP)方法解析DeepSurv模型的可解释性。绘制Kaplan-Meier生存曲线并行log-rank检验,评估模型的风险分层能力。通过决策曲线分析(DCA)评估模型临床实用性。结果 训练集、验证集患者基线特征差异均无统计学意义(均P>0.05)。LASSO回归筛选出8个影响老年CRC患者CSS的关键预后因子:年龄、T分期、N分期、组织学分级、神经周围浸润、阳性淋巴结比率、术前血清癌胚抗原(CEA)和术后辅助化疗。多因素分析显示,N分期(N1HR=2.43,95%CI为1.86~3.17,P<0.001;N2HR=3.49,95%CI为2.32~5.27,P<0.001)、未接受术后辅助化疗(HR=1.94,95%CI为1.51~2.50,P<0.001)、年龄≥85岁(HR=1.49,95%CI为1.22~1.82,P<0.001)、神经周围浸润阳性(HR=1.61,95%CI为1.26~2.04,P<0.001)、术前血清CEA阳性(HR=1.44,95%CI为1.19~1.75,P<0.001)、组织学分级Ⅲ~Ⅳ级(HR=1.49,95%CI为1.20~1.85,P<0.001)、阳性淋巴结比率≥0.09(HR=1.48,95%CI为1.09~2.01,P=0.012)和T3-4期(HR=1.47,95%CI为1.09~1.99,P=0.012)均为影响高龄CRC患者CSS的独立危险因素。在训练集中,DeepSurv模型预测患者1、2、3年CSS率的曲线下面积(AUC)分别为0.727(95%CI为0.681~0.769)、0.753(95%CI为0.720~0.791)和0.744(95%CI为0.717~0.773),LASSO-Cox模型的相应AUC分别为0.720(95%CI为0.675~0.764)、0.748(95%CI为0.714~0.782)和0.741(95%CI为0.712~0.771);在验证集中,DeepSurv模型预测患者1、2、3年CSS率的AUC分别为0.758(95%CI为0.699~0.818)、0.786(95%CI为0.743~0.830)和0.778(95%CI为0.737~0.819),LASSO-Cox模型的相应AUC分别为0.757(95%CI为0.701~0.812)、0.783(95%CI为0.736~0.829)和0.768(95%CI为0.724~0.810)。LASSO-Cox模型在训练集中预测1、2、3年CSS率的Brier评分分别为0.064、0.105、0.143,在验证集中相应得分分别为0.058、0.097、0.135;DeepSurv模型在训练集中的相应评分分别为0.060、0.093、0.127,在验证集中相应得分分别为0.056、0.087、0.120。SHAP可解释性分析表明,N分期、年龄和术后辅助化疗是影响DeepSurv模型预测效能的前3个关键变量。生存分析显示,LASSO-Cox模型预测训练集和验证集低、高风险组3年CSS率分别为92.49%和76.73%(χ2=132.21,P<0.001)、94.31%和76.88%(χ2=69.07,P<0.001);DeepSurv模型预测训练集和验证集低、高风险组3年CSS率分别为92.60%和76.40%(χ2=142.34,P<0.001)、94.60%和76.20%(χ2=76.70,P<0.001)。DCA曲线显示两种模型均具有良好的临床实用性。结论 N分期、年龄、术后辅助化疗、阳性淋巴结比率、术前血清CEA、组织学分级、神经周围浸润及T分期均为高龄CRC患者的独立预后因素。基于上述变量构建的LASSO-Cox模型和DeepSurv深度学习模型均具有良好的预测性能和风险分层能力。

关键词: 机器学习, 结直肠肿瘤, 高龄患者, 药物疗法, 预后

Abstract:

Objective To construct a postoperative survival prediction model suitable for elderly patients with colorectal cancer (CRC) based on machine learning algorithms, as well as to identify key factors influencing the prognosis of patients. Methods Clinical data of 3 576 elderly CRC patients who underwent surgical treatment between 2000 and 2019 were obtained from the United States Surveillance, Epidemiology, and End Results (SEER) database. The dataset was randomly divided into the training cohort (n=2 503) and the validation cohort (n=1 073) at a 7∶3 ratio. Least absolute shrinkage and selection operator (LASSO) regression was applied to identify key predictors of cancer-specific survival (CSS), followed by the construction of LASSO-Cox and DeepSurv deep learning models. Model performance was evaluated using the concordance index (C-index), receiver operator characteristic (ROC) curves, and Brier score. To analyze the interpretability of the DeepSurv model, the Shapley additive explanation (SHAP) method was applied. The Kaplan-Meier survival curve was plotted and the log-rank test was performed to evaluate the risk stratification ability of the model. The decision curve analysis (DCA) was used to verify the clinical practicability. Results There were no statistically significant differences in baseline characteristics between the training cohort and validation cohort (all P>0.05). Eight key prognostic factors influencing CSS in elderly CRC patients were screened out using LASSO regression: age, T stage, N stage, histological grade, perineural invasion, positive lymph node ratio, preoperative serum carcinoembryonic antigen (CEA), and postoperative adjuvant chemotherapy. Multivariate analysis confirmed that N stage (N1HR=2.43, 95%CI: 1.86-3.17, P<0.001; N2HR=3.49, 95%CI: 2.32-5.27, P<0.001), absence of postoperative adjuvant chemotherapy (HR=1.94, 95%CI: 1.51-2.50, P<0.001) , age≥85 years (HR=1.49, 95%CI: 1.22-1.82, P<0.001), positive perineural invasion (HR=1.61, 95%CI: 1.26-2.04, P<0.001), positive preoperative serum CEA (HR=1.44, 95%CI: 1.19-1.75, P<0.001), histological grade Ⅲ-Ⅳ (HR=1.49, 95%CI: 1.20-1.85, P<0.001), positive lymph node ratio≥0.09 (HR=1.48, 95%CI: 1.09-2.01, P=0.012), and T3-4 stage (HR=1.47, 95%CI: 1.09-1.99, P=0.012) were independent risk factors for CSS in elderly patients with CRC. In the training cohort, the DeepSurv model yielded areas under the curve (AUCs) of 0.727 (95%CI: 0.681-0.769), 0.753 (95%CI: 0.720-0.791), and 0.744 (95%CI: 0.717-0.773) for predicting 1-, 2-, and 3-year CSS rate, respectively. The LASSO-Cox model yielded corresponding AUCs of 0.720 (95%CI: 0.675-0.764), 0.748 (95%CI: 0.714-0.782), and 0.741 (95%CI: 0.712-0.771). In the validation cohort, the DeepSurv model achieved AUCs of 0.758 (95%CI: 0.699-0.818), 0.786 (95%CI: 0.743-0.830), and 0.778 (95%CI: 0.737-0.819) for predicting 1-, 2-, and 3-year CSS rate, respectively. The LASSO-Cox model demonstrated similar performance, with AUCs of 0.757 (95%CI: 0.701-0.812), 0.783 (95%CI: 0.736-0.829), and 0.768 (95%CI: 0.724-0.810). The LASSO-Cox model achieved Brier scores of 0.064, 0.105, and 0.143 for predicting 1-, 2-, and 3-year CSS rate in the training cohort, and 0.058, 0.097, and 0.135 in the validation cohort. The DeepSurv model achieved Brier scores of 0.060, 0.093, and 0.127 in the training cohort, and 0.056, 0.087, and 0.120 in the validation cohort. SHAP analysis identified N stage, age, and postoperative adjuvant chemotherapy as the three most influential variables affecting the predictive performance of the DeepSurv model. Survival analysis revealed that the 3-year CSS rates for the low- and high-risk groups predicted by the LASSO-Cox model were 92.49% and 76.73% in the training cohort (χ2=132.21, P<0.001), and 94.31% and 76.88% in the validation cohort (χ2=69.07, P<0.001). The corresponding rates predicted by the DeepSurv model were 92.60% and 76.40% in the training cohort (χ2=142.34, P<0.001), and 94.60% and 76.20% in the validation cohort (χ2=76.70, P<0.001). The DCA curve showed that both models had good clinical practicability. Conclusions N stage, age, postoperative adjuvant chemotherapy, positive lymph node ratio, preoperative serum CEA, histological grade, perineural invasion, and T stage are identified as independent prognostic factors in elderly patients with CRC. The LASSO-Cox model and the DeepSurv deep learning model developed based on these variables demonstrate favorable predictive performance and robust risk stratification ability.

Key words: Machine learning, Colorectal neoplasms, Elderly patients, Drug therapy, Prognosis