Threshold-Optimized and Calibrated Logistic Regression for Breast Cancer Classification

Breast cancer affects over 2.3 million individuals annually worldwide. Traditional diagnostic methods face limitations in consistency and objectivity, particularly in resource-constrained settings. This study developed a logistic regression-based clinical decision support system for breast cancer classification. We analyzed the Wisconsin Diagnostic Breast Cancer dataset containing 569 samples with 30 quantitative morphological features from fine needle aspirate cytology. The dataset comprised 357 benign and 212 malignant cases. Data underwent standardization via StandardScaler, followed by 75-25 train-test partitioning (426 training, 143 testing samples). We evaluated the logistic regression model through confusion matrix analysis, ROC curve assessment, threshold optimization via Youden's Index, and probability calibration using Expected Calibration Error (ECE). The model achieved 95.8% accuracy, 96.2% sensitivity, and 95.6% specificity on independent testing data, with AUC-ROC of 0.993. Threshold optimization identified 0.560 as the optimal decision boundary, yielding 3.77% false negative rate and 4.44% false positive rate. Probability calibration demonstrated reliable predictions with ECE of 0.0390, improved to 0.0328 through isotonic regression. The model correctly classified 137 of 143 test samples (86 true negatives, 51 true positives, 4 false positives, 2 false negatives). The logistic regression model demonstrated strong discriminative performance for breast cancer classification. However, single train-test validation and dataset-specific characteristics require cautious interpretation. Cross-validation and external validation remain necessary for clinical translation.