Abstract
This study evaluated machine learning (ML) models on the Wisconsin Breast Cancer Dataset (WBCD), refined to 554 unique instances after addressing 5% missing values via mean imputation, removing 15 duplicates, and normalizing features with Min-Max scaling. Data were split into 80% training and 20% testing, maintaining a 63% benign and 37% malignant distribution. Using 10-fold cross-validation, the random forest, XGBoost, and deep neural network (DNN) models achieved accuracies of 96.5% (95% CI: [93.1-98.6]), 97.4% 95% CI: [94.2-99.1], and 98.0% (95% CI [95.1-99.5]), respectively. The DNN demonstrated a benign precision of 0.97, malignant precision of 1.00, benign recall of 1.00, malignant recall of 0.95, and F1-scores of 0.99 and 0.98, with an ROC-AUC of 0.992 (p < 0.001); its accuracy further improved to 98.9% after Bayesian hyperparameter tuning. Additionally, a convolutional neural network (CNN) using transfer learning (VGG16) achieved 99.3% accuracy, with precision and recall of 99.4% and 99.2%, respectively, although potential domain mismatch issues warrant caution. Optimized DNN and CNN models achieved high accuracy, demonstrating highly reliable diagnostic performance with promising clinical applicability.