Comparative Evaluation of Supervised Learning Algorithms for Cancer Treatment Response Prediction Using Clinical and Biomarker Features

Abstract

Accurate prediction of cancer treatment response remains essential for personalized oncology. This study conducts comprehensive evaluation of six supervised learning algorithms---Logistic Regression, Decision Tree, Support Vector Machine, Random Forest, XGBoost, and LightGBM---for predicting chemotherapy and targeted therapy responses. We developed a feature engineering pipeline integrating neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and treatment history from 847 cancer patients across five tumor types. XGBoost achieved superior performance with AUC-ROC of 0.873, accuracy of 81.4%, and F1-score of 0.796, significantly outperforming baseline Logistic Regression (AUC-ROC 0.742, p < 0.001). SHAP analysis identified NLR (mean |SHAP| = 0.187), tumor stage, and prior treatment lines as the most predictive features. Kaplan-Meier survival analysis stratified by predicted risk demonstrated strong clinical validity, with high-risk patients exhibiting median progression-free survival of 4.7 months versus 14.3 months for low-risk patients (log-rank p < 0.001, hazard ratio 3.24). Gradient boosting algorithms provide optimal balance between predictive accuracy, computational efficiency, and clinical interpretability for treatment response prediction.