A Comparative Study of Machine Learning Models for Breast Cancer Diagnosis Using the Wisconsin Diagnostic Dataset

Accurate and interpretable diagnostic models are vital for early detection of breast cancer. This paper presents a comparative analysis of five classifiers Logistic Regression (LogReg), Random Forest (RF), Gradient Boosting (GBRT), a compact Multilayer Perceptron (MLP), and an attention-based TabNet model for the Breast Cancer Wisconsin (Diagnostic) dataset. Classical models were optimised via grid search and stratified 5-fold crossvalidation (CV), while the MLP used validation-based early stopping. On the held-out test set, LogReg achieved the strongest overall balance of discrimination and calibration (AUC $\approx 0.996, \mathbf{A P} \approx 0.998$), with competitive $\mathbf{F 1}$ accuracy. LogReg trained approximately ten times faster than ensemble and neural network models while preserving near real-time inference; GBRT was slightly faster per sample but required significantly longer training. These results reinforce that transparent, well-calibrated linear models can match ensemble and neural network baselines on small structured clinical data, supporting their suitability for time-sensitive decision support. These findings highlight the importance of model calibration and efficiency for reliable AI-assisted screening in clinical settings.