Development and Cross-Dataset Replication of a Reproducible Explainable Machine Learning Pipeline for Breast Cancer Classification Using Cytological Measurements

Abstract

Machine-learning models for breast-cancer cytology are rarely published in a form that others can rerun, and their explanations are rarely checked for stability. This study develops a reproducible, configuration-driven pipeline — automated preprocessing, SHAP-based feature selection, a comparative model panel following the explainable-model comparison of Siddiqui et al. [1], a soft-voting ensemble, probability calibration, a screening threshold and a prediction interface returning a probability, a decision and the three features that drove it — and replicates it, unchanged, on two independent Wisconsin cytology datasets with different measurement systems: the Wisconsin Diagnostic Breast Cancer dataset (WDBC; 569 samples, 30 computer-measured nuclear features) and the original Wisconsin Breast Cancer dataset (WBCD; 699 samples, 9 pathologist-scored features, 16 missing values). On held-out test sets the ensemble reached, at a screening threshold chosen on development data, a sensitivity of 0.943 and specificity of 0.944 on WDBC (AUC 0.971) and a sensitivity of 0.983 and specificity of 0.957 on WBCD (AUC 0.986), with Brier scores of 0.032 on both. The pipeline was exactly reproducible under a fixed seed, and changing the seed altered test AUC by at most 0.0007. Feature selection retained 6–10 of 30 features on WDBC and 4–6 of 9 on WBCD, with mean pairwise Jaccard similarity of 0.74 and 0.73 across folds, and SHAP rankings were consistent across folds (mean Spearman ρ 0.83 and 0.75). In both datasets the explanation was dominated by nuclear size and shape irregularity — worst area, perimeter and concave points in WDBC; uniformity of cell size and shape and bare nuclei in WBCD — consistent with cytopathological criteria. Prediction took about 0.8 ms per case and a full explained prediction about 0.26 s. Because the two datasets use different features, this is a replication of the pipeline rather than external validation of a single fitted model; validation on an independent measured-feature dataset remains necessary.