Hybrid attention network with TabNet benchmarking for breast cancer and heart disease classification

CancerHybridNet is an enhanced Tabular Network (TabNet) incorporating global self-attention for improved clinical prediction. The model was evaluated on three medical datasets: Wisconsin Diagnostic Breast Cancer (WDBC), Breast Cancer Coimbra (BCCD), and Cleveland Heart Disease, against Logistic Regression, Support Vector Machine, Random Forest, Extreme Gradient Boosting, and TabNet. Using stratified splits and standardized preprocessing, CancerHybridNet achieved predictive accuracies of 97.8%, 85.0%, and 90.2% on WDBC, BCCD, and Cleveland, respectively. While gains over baselines were small on WDBC, larger improvements were observed on BCCD and Cleveland, with paired Wilcoxon signed-rank tests confirming statistical significance (p < 0.05). Computational analysis showed minimal overhead compared with classical classifiers and complexity comparable to TabNet. Results highlight that hybrid attention mechanisms improve feature masking and deliver competitive performance on small clinical tabular datasets.