A Comparative Study of Machine Learning Models for Cancer Survival Prediction: Feasibility and Performance Evaluation

Accurate survival prediction for cancer patients is critical for guiding treatment decisions, but traditional statistical models have limitations in capturing complex, nonlinear relationships. Machine Learning (ML) models offer a promising alternative, yet their feasibility for predicting continuous survival times remains to be determined. This study conducts a comparative evaluation of five regression-based ML algorithms: k-Nearest Neighbors (KNN), linear regression (LR), decision tree (DT), random forest (RF), and eXtreme Gradient Boosting (XGBoost), applied to three distinct cancer datasets (breast cancer, a multi-cancer cohort, and lung cancer) totaling approximately 45,000 patient records. Model performance was assessed via cross-validation using the coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) and error metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The results demonstrate that all models struggled to predict exact survival durations with high accuracy. The KNN model performed worst, exhibiting the largest prediction errors and negative R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values, indicating essentially no explained variance. The other models achieved only slightly better results; in most cases, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values were near zero, with the highest being approximately 0.17 in the breast cancer dataset. These findings highlight that, unlike some classification-based approaches achieving high accuracy, current ML regression methods face challenges capturing the multifactorial nature of survival outcomes. This study underscores the limited practical feasibility of these models for precise survival time prediction and suggests incorporating additional prognostic factors or adopting hybrid modeling strategies to enhance performance.