Breast Cancer Prediction Using CatBoost Classifier: An Interpretable and Well-Calibrated Machine Learning Approach for Clinical Decision Support
Main Article Content
Abstract
This article is about the performance of CatBoost on breast cancer diagnosis based on WI, sconsin Diagnostic Breast Cancer (WDBC)Dataset using a CatBoost classifier. As can be seen from the learning curve, the model has excellent predictive power, obtaining nearly perfect training accuracy, along with a cross-validation accuracy of ~97. The best predictive features found using permutation importance were concave points_worst, concave points_mean, area_worst, and texture_worst. Analysis of partial dependence reinforces that greater magnitude of these features increase predicted probability of malignancy. The Precision– Recall curve depicts an AUC value of 0.9968 reflecting the sensitivity of the model to the prediction of malignant in class imbalance settings. We show that predicted probabilities are close to observed outcomes in our calibration analysis, which is reassuring for risk estimation. CatBoost offers great accuracy, interpretability, and well-calibrated prediction probabilities, making it a suitable choice for supporting clinical decision-making in breast cancer detection.