Analysis of prognostic factors for esophageal cancer after radical resection and the applica-tion value of machine learning prediction model
10.3760/cma.j.cn115610-20250819-00542
- VernacularTitle:食管癌根治术后预后影响因素分析及机器学习预测模型的应用价值
- Author:
Yue ZHAO
1
;
Sijie ZHANG
;
Haiming LI
;
Yijun MA
;
Zhan ZHANG
;
Zhenyi LI
;
Junjie LIU
;
Hui TIAN
;
Yu TIAN
Author Information
1. 山东大学齐鲁医院胸外科，济南 250012
- Publication Type:Journal Article
- Keywords:
Esophageal neoplasms;
Prognosis influencing factor;
Machine learning;
Prediction model;
Application
- From:
Chinese Journal of Digestive Surgery
2025;24(10):1305-1317
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the prognostic factors for esophageal cancer after radical resection and the application value of machine learning prediction model.Methods:The retrospective cohort study was conducted. The clinicopatholigical data of 406 esophageal cancer patients who were admitted to Qilu Hospital of Shandong University from January 2018 to March 2022 were collected. There were 357 males and 49 females, aged (64±8)years. All patients underwent radical resection of esophageal cancer. The 406 patients were randomly divided into a training set of 285 cases and a validation set of 121 cases at a 7∶3 ratio based on a random number table. The training set was used to construct prediction model, and the validation set was used to validate prediction model. Patients were divided into high-risk group and low-risk group based on risk scores. Observation indicators: (1) follow-up of patients and analysis of influencing factors for prognosis; (2) construction and validation of machine learning prediction models. Comparison of measurement data with normal distribution between groups was conducted using the independent sample t test. Comparison of measurement data with skewed distribution between groups was conducted using the Mann-Whitney U test. Comparison of count data between groups was conducted using the chi-square test. Comparison of ordinal data between groups was conducted using the rank sum test. The Kaplan-Meier method was used to calculate survival rate and plot survival curve, and the Log-rank test was used for survival analysis. The Cox proportional hazard regression model was used for univariate and multivariate analyses. Independent influencing factors were included, and data processing, machine learning model construction, and visualization were performed using R packages including random survival forest (RSF), gradient boosting machine (GBM), least absolute shrinkage and selection operator Cox regression (LASSO-Cox), Cox proportional hazards model boosting (CoxBoost), survival support vector machine (survivalsvm), extreme gradient boosting (XGBoost), supervised principal component analysis (SuperPC), and Cox partial least squares regression (plsRcox). Receiver operating characteristic (ROC) curves were drawn, and sensitivity, specificity, and area under the curve (AUC) were calculated. The Delong test was used to assess the differences in AUC among different models in the training set, and the time-dependent ROC was used to compare the predictive performance of different models. Calibration curves were used to evaluate model accuracy, and decision curve analysis (DCA) was used to evaluate overall net benefit. Results:(1) Follow-up of patients and analysis of influencing factors for prognosis. All 406 patients were followed up postoperatively for 28(range, 6-36)months, with 1- and 3-year overall survival rate of 86.5% and 40.9%, respectively. The 285 patients in the training set were followed up postoperatively for 30(range, 6-36)months, with 1- and 3-year overall survival rate of 85.1% and 35.5%, respectively. The 121 patients in the validation set were followed up postoperatively for 25(range, 6-36)months, with 1- and 3-year overall survival rate of 87.0% and 43.2%, respectively. There was no significant difference in postoperative overall survival rate between the training set and the validation set ( χ2=3.20, P>0.05). Results of multivariate analysis showed that left thoracic surgical approach, preopera-tive neutrophil count, vascular invasion, perineural invasion, pathological T2-4 stage, pathological N2-3 stage, and postoperative pneumonia were independent risk factors affecting postoperative survival of 285 patients in the training set ( hazard ratio=1.466, 1.037, 1.482, 1.549, 5.268, 7.727, 22.202, 2.539, 2.686, 1.425, 95% confidence interval as 1.026-2.096, 1.003-1.073, 1.008-2.179, 1.105-2.170, 1.201-23.099, 1.833-32.576, 4.734-104.128, 1.577-4.087, 1.631-4.422, 1.018-1.994, P<0.05). (2) Construction and validation of machine learning prediction models. Independent risk factors affecting postoperative survival were included to construct RSF, GBM, LASSO-Cox, CoxBoost, survivalsvm, XGBoost, SuperPC, and plsRcox machine learning prediction models. Results of Delong test showed that there were significant differences in the AUC of RSF and GBM from the other six models ( P<0.05). Results of time-dependent ROC curve showed that all 8 machine learning predic-tion models had good discriminative ability in the training cohort, among which the RSF machine learning prediction model had the best predictive performance. Results of calibration curve showed that the RSF machine learning prediction model fitted well for predicting postoperative 1-, 2-, and 3-year overall survival in the training cohort, with high consistency with actual results. Results of decision curve analysis showed that within a threshold range of 0-0.80, the RSF machine learning prediction model provided a better overall net benefit. Further analysis showed that in the validation set, the AUC of RSF machine learning prediction model for postoperative 1-, 2-, and 3-year survival prediction were 0.786 (95% confidence interval as 0.609-0.962), 0.774 (95% confidence interval as 0.676-0.873), and 0.750 (95% confidence interval as 0.652-0.848), respectively. Results of calibration curve showed that the RSF machine learning prediction model fitted well for predicting postopera-tive 1-, 2-, and 3-year overall survival in the validation set, with high consistency with actual results. In the training set, the optimal cutoff value of the RSF machine learning prediction model risk score was 11.7. Patients with risk score ≥11.7 were classified as the high-risk group, and those with risk score <11.7 as the low-risk group. The median survival times of the two groups were 18.0 months and >36.0 months, respectively, showing a significant difference between them ( χ2=73.30, P<0.05). In the validation set, the optimal cutoff value of the RSF machine learning prediction model risk score was 11.7. Patients with risk score ≥11.7 were classified as the high-risk group, and those with risk score<11.7 as the low-risk group. The median survival times of the two groups were 17.0 months and>36.0 months for the high-risk and low-risk groups, respectively, showing a significant difference between them ( χ2=35.20, P<0.05). Conclusions:Left thoracic surgical approach, preoperative neutrophil count, vascular invasion, perineural invasion, pathological T2-4 stage, pathological N2-3 stage, and postoperative pneumonia are independent risk factors affecting survival of esophageal cancer patients after radical resection. The RSF machine learning prediction model constructed based on these factors can effectively distinguish the survival prognosis of high-risk and low-risk patients.
