Breast cancer prediction modeling based on SHAP interpretability analysis and XGBoost algorithm

This study aimed to predict breast cancer and identify associated risk factors by comparing the performance of the Extreme Gradient Boosting (XGBoost) model with four conventional machine learning models, followed by SHapley Additive exPlanations (SHAP) analysis to explain model predictions. Breast tumor datasets were retrieved from the University of California, Irvine (UCI) Machine Learning Repository. Feature selection was performed using heatmaps of correlation coefficient matrix to eliminate redundant and highly correlated variables. Five machine learning models, namely Extreme Gradient Boosting (XGBoost), Random Forest, K-Nearest Neighbors (KNN), Decision Tree, and Support Vector Machine (SVM), were comprehensively compared in terms of precision, recall, F1-score, training time, and classification accuracy. ROC curves were generated for all evaluated models, and the DeLong test was used to statistically compare the corresponding AUC values. Five-fold cross-validation was performed to assess the robustness of classification accuracy across models. Meanwhile, McNemar’s test was adopted for pairwise accuracy comparisons to further evaluate the predictive performance of models. Furthermore, SHAP analysis was employed to interpret model outputs, which enabled the identification of key predictive features associated with breast cancer pathogenesis and their interactive effects. As indicated by the confusion matrix results, XGBoost model achieved the highest accuracy at 97.3%, whereas the RF, KNN, SVM, and DT models achieved accuracies of 95.6%, 94.7%, 92.1%, and 89.5%, respectively. ROC curve analysis demonstrated that the XGBoost model yielded the highest AUC value of 0.97, whereas the AUC values for the RF, KNN, DT, and SVM models were 0.96, 0.94, 0.93, and 0.90, respectively. SHAP analysis further revealed that radius-worst, concave points-worst, and concavity-worst served as the key predictive factors associated with high risk of breast cancer. Notably, a significant interactive correlation was observed between radius-worst and concave points-worst features. Both XGBoost and RF demonstrated favorable predictive performance in the diagnosis of breast cancer. Specifically, the XGBoost model achieved superior numerical results in terms of precision, recall, F1-score and training time. Nevertheless, no statistically significant difference was observed in the overall performance between the two models. Radius-worst was identified as a pivotal feature associated with breast cancer occurrence, and a notable interactive effect was detected between radius-worst and concave points-worst.