Abstract:

Objective To construct a machine learning（ML）-based prediction model for esophageal cancer （EC） using routine clinical laboratory data， and to clarify the relative contribution of key laboratory indicators to model prediction. Methods A retrospective case-control study was conducted. The clinical laboratory data were collected from 2 763 patients with EC and 3 297 age-and sex-matched healthy controls（HC） treated at the First Affiliated Hospital of Air Force Medical University between January 2019 and December 2024. After missing value imputation， outlier removal， and data normalization， recursive feature elimination （RFE） combined with least absolute shrinkage and selection operator （LASSO） regression was used to screen predictive features. Based on the selected features， extreme gradient boosting （XGBoost）， logistic regression （LR）， light gradient boosting machine （LightGBM）， random forest （RF）， adaptive boosting （AdaBoost）， and decision tree （DT） models were constructed. Model performance was comprehensively evaluated using the area under the receiver operating characteristic（ROC） curve （AUC）， decision curve analysis （DCA）， calibration curves， and precision-recall （PR） curves. SHapley Additive exPlanations （SHAP） were applied to interpret the optimal model. Results From 67 routine laboratory indicators， 10 EC-related predictive features were identified， including carcinoembryonic antigen （CEA）， urinary cast quantity （UCQ）， red cell distribution width-standard deviation （RDW-SD）， platelet （PLT） count， platelet distribution width （PDW）， alkaline phosphatase （ALP）， uric acid （UA）， creatinine （Cr）， alanine aminotransferase （ALT）， and total protein （TP）. Among the compared models， the LightGBM model achieved the best performance in the test cohort， with an AUC of 0.953 ［95% confidence interval（95%CI）： 0.942-0.964］， and demonstrated good calibration and higher clinical net benefit in calibration and DCA. Conclusion The ML-based prediction models for EC was developed using routine clinical laboratory data and systematically validated in an internal dataset. Among the compared models， the LightGBM model showed the best performance， with high discriminative ability and good validation performance.

Key words: Esophageal neoplasm, Machine learning, Risk prediction model, Shapley additive explanation, Tumor screening