Predicting Postoperative Recurrence Using a Support Vector Machine for Patients With Esophageal Squamous Cell Carcinoma: Machine Learning Modeling Development and Validation Study.

Background: While numerous models have been developed to predict overall survival in postoperative patients with esophageal squamous cell carcinoma (ESCC), few have specifically focused on predicting postoperative recurrence. Objective: This study aimed to develop and validate a support vector machine (SVM)-based predictive model for evaluating recurrence risk and identifying associated factors in ESCC patients following surgery. Methods: We retrospectively analyzed clinical data from 311 ESCC patients who underwent surgery at Jinling Hospital between June 2014 and November 2016, with follow-up until October 2021 (median of 36 follow-up months, range 0-93.5 months). After excluding cases with incomplete data (n=1), 310 eligible patients were randomly allocated into test (n=106), validation 1 (n=103), and validation 2 (n=101) cohorts. Using SVM algorithms, patients were stratified into high- or low-recurrence-risk groups. Model performance was assessed using sensitivity, specificity, the Youden index, positive predictive value, and negative predictive value. Calibration curves were generated to evaluate model accuracy and reliability. Statistical analyses were performed using SPSS (version 22.0; IBM Corp) and R (version 3.6.1; R Foundation for Statistical Computing). Results: In all cohorts, SVM7 (incorporating tumor node metastasis [TNM] stage, adjuvant therapy, differentiation, tumor size, and complications) demonstrated significantly higher sensitivity in predicting recurrence than SVM6 (based on the Eastern Cooperative Oncology Group performance status, neutrophil-to-lymphocyte ratio, and CY211) (P<.001). The composite model SVM6+8 (combining SVM6 and SVM8 [SVM7 excluding complications]) achieved recurrence prediction sensitivities of 94%, 79.59%, and 72.73% in the test, validation 1, and validation 2 groups, respectively; with specificities of 98.11%, 69.84%, and 78.43%. These results were comparable to SVM6+TNM (SVM6 combined with TNM staging) but outperformed SVM6 alone (P<.001). Survival analysis revealed significantly longer disease-free survival in the SVM6+TNM-predicted low-risk group compared to the high-risk group, with a marked difference in recurrence rates (P<.001). Conclusions: The proposed SVM-based model enables accurate prediction of postoperative recurrence in ESCC patients with high sensitivity, specificity, and discriminative power, offering a valuable tool for clinical risk stratification.