Development and Validation of An Explainable Machine Learning Model for Predicting Pathological Complete Response to Neoadjuvant Chemoradiotherapy in Esophageal Squamous Cell Carcinoma: A Machine Learning Study
Domaine:
healthcareType de record:
modelpaper
Créateur:
LouMenYukXin
Éditeur:
Hôte:
Background: Esophageal squamous cell carcinoma (ESCC) is a common gastrointestinal malignancy with a poor prognosis; pathological complete response (pCR) following neoadjuvant chemoradiotherapy (nCRT) is a key determinant of long-term survival. This study established and validated a machine learning model that integrates baseline and dynamic inflammatory markers to predict pCR in ESCC following nCRT.
Methods: A multicenter retrospective cohort study was conducted, including 231 patients with ESCC treated at Shandong Cancer Hospital (divided into a training set and an internal validation set in an 8:2 ratio) and 57 independent patients from Anyang Cancer Hospital for external validation. We evaluated over 20 candidate predictive factors, including clinical characteristics, inflammation, nutrition, and metabolism, with a focus on indicators of dynamic changes before and after treatment. Feature selection employed a three-step process combining univariate logistic regression, the Least Absolute Shrinkage and Selection Operator (LASSO) regression, and multivariate logistic regression, and variance inflation factor (VIF) analysis was used to confirm the absence of multicollinearity. Five supervised algorithms were trained using the training-validation split: logistic regression (LR), random forest (RF), XGBoost, support vector machine (SVM), and k-nearest neighbors (KNN). Model performance was evaluated using discriminatory power, calibration, and receiver operating characteristic (ROC) curve analysis, and model interpretability was assessed using Shapley additive explanations (SHAP).
Results: The optimal model was the RF, with an internal validation AUC of 0.894 (95% CI: 0.850–0.939) and an external validation AUC of 0.808 (95% CI: 0.691–0.925). Four core predictors were ultimately retained: pre-treatment neutrophil-to-lymphocyte ratio (PreNLR), pre-treatment lymphocyte-to-monocyte ratio (PreLMR), change in neutrophil-to-lymphocyte ratio (△NLR), and change in platelet-to-lymphocyte ratio (△PLR). This study successfully developed and deployed a web-based online visualization calculator for individualized pCR prediction.
Conclusion: The machine learning model developed in this study integrates static baseline indicators with dynamic rates of inflammatory change and can predict pCR status following nCRT for ESCC. The model is stable and highly interpretable, and the accompanying online tool provides non-invasive, efficient support for clinical decisions regarding individualized treatment.
Methods: A multicenter retrospective cohort study was conducted, including 231 patients with ESCC treated at Shandong Cancer Hospital (divided into a training set and an internal validation set in an 8:2 ratio) and 57 independent patients from Anyang Cancer Hospital for external validation. We evaluated over 20 candidate predictive factors, including clinical characteristics, inflammation, nutrition, and metabolism, with a focus on indicators of dynamic changes before and after treatment. Feature selection employed a three-step process combining univariate logistic regression, the Least Absolute Shrinkage and Selection Operator (LASSO) regression, and multivariate logistic regression, and variance inflation factor (VIF) analysis was used to confirm the absence of multicollinearity. Five supervised algorithms were trained using the training-validation split: logistic regression (LR), random forest (RF), XGBoost, support vector machine (SVM), and k-nearest neighbors (KNN). Model performance was evaluated using discriminatory power, calibration, and receiver operating characteristic (ROC) curve analysis, and model interpretability was assessed using Shapley additive explanations (SHAP).
Results: The optimal model was the RF, with an internal validation AUC of 0.894 (95% CI: 0.850–0.939) and an external validation AUC of 0.808 (95% CI: 0.691–0.925). Four core predictors were ultimately retained: pre-treatment neutrophil-to-lymphocyte ratio (PreNLR), pre-treatment lymphocyte-to-monocyte ratio (PreLMR), change in neutrophil-to-lymphocyte ratio (△NLR), and change in platelet-to-lymphocyte ratio (△PLR). This study successfully developed and deployed a web-based online visualization calculator for individualized pCR prediction.
Conclusion: The machine learning model developed in this study integrates static baseline indicators with dynamic rates of inflammatory change and can predict pCR status following nCRT for ESCC. The model is stable and highly interpretable, and the accompanying online tool provides non-invasive, efficient support for clinical decisions regarding individualized treatment.