Abstract
INTRODUCTION: Gastrointestinal (GI) cancers present significant clinical challenges characterized by dismal survival outcomes and suboptimal prognoses. Currently, only partial indicators are available to predict the response of immunotherapy. A critical gap remains in the development of models capable of accurately predicting response rates to immunotherapy regimens. In this study, we developed a machine-learning (ML) model based on factorial, molecular, demographic, and clinical data to predict the response rate. METHODS: This multicentre retrospective study analyzed the clinical data of 506 patients, comprising 352 cases collected from Zhongnan Hospital of Wuhan University and Hubei Cancer Hospital, along with 154 cases obtained from the publicly available dataset of Memorial Sloan-Kettering Hospital. We used 14 features as input features, such as the patient's basic status, biochemical test results, and genetic test results. Eight ML methods were employed to build predictive models. Through rigorous validation using seven discriminative performance metrics (accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, and Brier score), the eXtreme Gradient Boosting (XGBoost) algorithm demonstrated superior predictive capability. Model interpretability was subsequently enhanced through Shapley Additive explanations (SHAP) analysis to elucidate feature contributions. RESULTS: We selected XGBoost with the best predictive performance to predict response (AUC: 0.829 [95% CI: 0.72-0.91], accuracy: 78.43%, sensitivity: 86.67%, specificity: 72.31%). The Delong test and calibration curve indicated that XGBoost significantly outperformed the other models in prediction. The SHAP values indicate that chemotherapy contributes the most to the model's predictive accuracy (contribution score = 0.28), while Ki-67 exhibits the lowest contribution rate (0.01). In addition, the study showed that chemotherapy, higher hemoglobin (HGB), body mass index (BMI), age, lower neutrophil-to-lymphocyte ratio (NLR), and tumor stage positively influenced the output of the model. CONCLUSION: Interpretable XGBoost models have shown accuracy, efficiency, and robustness in determining the association between input features and response rates. Among the input features, chemotherapy and tumor stage played the most important role in the prediction model. Due to the varying efficacy of ICIs in gastrointestinal cancers, personalized predictive models can greatly assist clinical decision-making. This model fills this gap in clinical practice and can provide more precise support for personalized treatment and risk avoidance.