Comparison of Prediction Models for Lung Cancer Data: Accuracy Analysis of Lasso Regression, SVM and XGBoost

Abstract

This study aimed to evaluate the performance of three predictive models on a lung cancer prediction dataset including Lasso Regression, Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost). To ensure data quality, preprocessing included converting binomial variables to numeric values and adjusting model weights to improve the detection of negative cases and reduce missed diagnoses. Model training was performed using a robust 5-fold cross-validation repeated three times to ensure accuracy and generalizability. For Lasso Regression, feature selection was optimized using both Minimum Squared Error (MSE) and 1-Standard Error (1SE) criteria. The SVM model’s cost parameter was fine-tuned through cross-validation. For XGBoost, key hyperparameters include the number of boosting rounds (nrounds), maximum depth (max_depth), learning rate, column sampling by tree (colsample_bytree), and subsampling rate (subsample) were selected to maximize accuracy. Model performance was assessed primarily through accuracy. Among the three, XGBoost achieved the highest accuracy, indicating its superior capability in handling binary classification tasks such as lung cancer prediction. These findings suggest that XGBoost may be particularly effective for clinical decision-support systems. Future research could expand to larger datasets or explore threshold adjustments to reduce false positives and further enhance model performance. This study provides useful guidance for selecting effective predictive models in lung cancer diagnosis.