PREDICTING BREAST CANCER MALIGNANCY FROM DIAGNOSTIC IMAGING FEATURES: AN EMPIRICAL COMPARISON OF MACHINE LEARNING MODELS FOR HEALTHCARE ANALYTICS

Abstract

Early and accurate diagnosis of breast cancer has a direct bearing on patient survival, yet interpretation of fine needle aspirate (FNA) imaging features still depends heavily on clinician experience and remains subject to inter-observer disagreement. This paper presents an empirical case study evaluating two supervised machine learning models — logistic regression and random forest — for classifying breast masses as malignant or benign using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (n = 569, 30 quantitative morphological features). Both models were trained on a stratified 75/25 train-test split and validated with 5-fold stratified cross-validation. Logistic regression produced the stronger held-out result (accuracy = 0.986, ROC-AUC = 0.998), edging out random forest (accuracy = 0.958, ROC-AUC = 0.995); cross-validation confirmed that both models generalized consistently across folds (mean ROC-AUC of 0.995 and 0.989, respectively). A feature importance analysis identified worst-case perimeter, worst-case area, and concave-point counts as the strongest predictors of malignancy, which lines up with the clinical intuition that irregular, larger nuclei tend to signal malignant disease. Beyond the modeling exercise itself, this paper situates the results against a decade of published work on the same benchmark dataset, and expands the discussion to cover explainability, regulatory expectations for AI-enabled diagnostic software, and the equity concerns that arise whenever a model trained on a specific, historical population is proposed for broader clinical use. The intent throughout is not to advance a novel algorithm but to walk through, with full methodological transparency, how a healthcare analytics model should be evaluated and interrogated before anyone seriously considers putting it in front of a clinician.