A High-Fidelity Predictive Analytics Framework for Chemotherapy-Induced Toxicity Using Explainable Gradient Boosting Models

The side effects of chemotherapy play a role in the treatment of cancer. Here we compare the prediction of XGBoost, Random Forest, and Decision Tree classifiers in predicting severe chemotherapy adverse events and their explainability by SHAP. The clinical data set was comprised of the variables of demographics, of biochemistry, and of treatment. The data in the dataset was processed and then modeled, validated, and tested on accuracy, precision, recall, F1score, and ROC-AUC. The accuracy of XGBoost is 93.9%, and the AUC is 0.96, which is better than the models using random forest with 89.6% and decision tree with 85.2% accuracy. As the ROC curves show, XG-Boost classification and false prediction were better than the other models. The SHAP summary plot, dependence plots showed that the best and most important predictors were drug dosage, liver enzyme, treatment cycles, and age of patients. XGBoost/SHAP suggests an interesting and interpretable predictive modeling system that provides a logical data-driven explanation of clinical decision-making and the development of precision oncology.