Machine learning to predict distant metastasis risk and prognostic analysis in T1-2 esophageal cancer: a study based on the SEER database.

Background: The risk of distant metastasis (DM) in T1-2 esophageal cancer (EC) is frequently underestimated, and the prognosis for patients who develop metastasis remains poor. Currently, there is a lack of effective tools based on routine clinical data to accurately identify high-risk patients preoperatively or to perform subsequent prognostic stratification for those with metastatic disease. This study aimed to develop an integrated tool to preoperatively identify T1-2 EC patients at high risk of DM and stratify prognosis for metastatic cases, providing evidence-based clinical support. Methods: This retrospective study utilized data from the Surveillance, Epidemiology, and End Results (SEER) database, including 12,928 patients. The cohort was split into training and testing sets (7:3). In the training set, univariate and multivariate logistic regression (LR) were first applied to initially identify factors associated with DM. Feature selection was subsequently conducted using the least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. Variables consistently selected by both methods and confirmed as independent predictors were retained for model construction. Next, 12 machine learning algorithms were applied to build prediction models. Model interpretability was achieved using SHapley Additive exPlanations (SHAP). Performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), calibration curves, precision-recall (PR) curves, and decision curve analysis (DCA) curves. For patients who developed DM, a prognostic nomogram was constructed based on factors identified through Cox regression and evaluated using the concordance index (C-index), AUC, calibration curves, and DCA. Results: Among the DM prediction models, the Naive Bayes classifier (NBC) model achieved AUC values of 0.740 and 0.750 in the training and testing sets, respectively, indicating favorable discriminative ability. Additionally, it exhibited the highest sensitivity among all models and was therefore selected as the final prediction model. SHAP analysis identified N stage as the most important predictive factor, with N stage, primary site, and tumor size being confirmed as risk factors for DM, whereas T stage and older age were associated with reduced risk. In the prognostic analysis, race, T stage, chemotherapy, marital status, and median household income were independent prognostic factors for metastatic patients. The resulting overall survival (OS) and cancer-specific survival (CSS) nomograms demonstrated good discrimination, with C-indexes of 0.653 and 0.654 in the testing set, respectively, along with satisfactory calibration and clinical net benefit. Conclusions: We developed a dual-strategy risk assessment tool that predicts DM in T1-2 EC and provides individualized survival estimates for metastatic patients, offering a practical framework for stratified clinical decision making and precision oncology.