<![CDATA[Breast cancer is one of the most significant global health issues. Machine learning approaches offer the potential to accurately analyze clinical data and aid in early diagnosis. However, conventional machine learning models are often limited in their ability to model complex nonlinear relationships in medical data, which can reduce predictive accuracy. This study employs a deep learning architecture because of its ability to model such relationships. Specifically, the TabNet model was chosen because it is designed for tabular data and offers better interpretability. The public Wisconsin Diagnostic Breast Cancer (WDBC) dataset, which has 30 features and an imbalanced class distribution, was used in this study. Feature selection was necessary to handle the high-dimensional data, and SMOTE-ENN was used for class balancing. Two feature selection methods, Chi-Square and Information Gain, were compared to determine the most effective approach. Hyperparameter optimization was performed using Optuna and validated with stratified k-fold cross-validation to ensure optimal performance. The results of the experiment demonstrate that feature selection and optimization significantly improve performance. The base model with Chi-Square feature selection achieved an accuracy rate of 64.91%. Meanwhile, the Chi-Square model with Optuna optimization increased accuracy to 98.25%. This is 3.51% higher than the accuracy of 94.74% achieved by the optimized model without feature selection. In the final comparison, both methods demonstrated distinct advantages: Chi-Square (75% features) excelled in achieving 100% precision and more efficient computation time. Information Gain (75% features), on the other hand, was the only method to achieve 100% recall, which is crucial for minimizing false negatives. These results demonstrate that the optimal method depends on the context. Information Gain is best for maximum diagnostic sensitivity, and Chi-Square is best for performance balance and efficiency.]]>
Copyrights © 2025
