Development and Validation of an Interpretable Machine Learning Model Based on Peripheral Blood Biomarkers for Esophageal Cancer Risk Prediction.

BACKGROUND

Noninvasive, low-cost prescreening tools are needed to improve risk stratification for esophageal cancer (EC) before endoscopic confirmation. We developed and validated explainable machine-learning (ML) models using routine peripheral blood biomarkers.

METHODS

This dual-center retrospective case-control study enrolled 454 participants (198 EC cases, 256 non-EC controls) from two hospitals between March 2021 and June 2025. Data were randomly split 7:3 into training (n=319) and validation (n=135) sets. LASSO regression selected nine features (SIRI, MLR, AST, ADA, CREA, UA, K, PT, and TT). Seven algorithms-logistic regression, decision tree, random forest (RF), XGBoost, LightGBM, support vector machine, and artificial neural network-were trained with 10-fold cross-validation and grid-search hyperparameter tuning. Performance was assessed by discrimination, calibration, clinical utility, and confusion matrices, with Shapley additive explanations (SHAP) for interpretation.

RESULTS

Baseline demographics and comorbidities were comparable between groups. In the validation set, RF performed best (AUC=0.973; accuracy=0.926; sensitivity=0.881; specificity=0.961; F1-score=0.912), achieved the lowest Brier score (0.059), and showed favorable net benefit. SHAP analysis identified creatinine and SIRI as the most influential features, where lower creatinine and higher SIRI increased predicted EC risk.

CONCLUSION

This explainable RF model showed excellent discrimination and good calibration. As a retrospective case-control study using healthy controls, it is intended as a prescreening tool to guide endoscopic referral rather than a diagnostic test, and requires prospective external validation before clinical use.