Explainable Federated Deep Learning for Low-Cost and Privacy-Preserving Early Breast Cancer Screening to Reduce U.S. Healthcare Burden

Abstract

Breast cancer, which overall has an established survival rate of 74% - compared to a progressively increasing 99% within the last 30 years, making early detection essential in improving prognoses - is both sensitive and difficult for clinicians. To overcome the above challenges, in this paper, we propose FedXAI-DP: a novel framework for privacy-preserving breast cancer classification that synergistically combines Federated Learning (FL), Explainable Artificial Intelligence (XAI), and Differential Privacy (DP). FedXAI-DP performs a SHAP-importance-weighted aggregation, where more important clients are given higher or lower weights proportionally in constructing the global model, as compared to existing FL strategies that aggregate model parameters uniformly for each client. This allows more discriminative feature information to rule the global update in the heterogeneous and non-IID client data. Surprisingly, we find under IID data conditions by experimenting on the Wisconsin Breast Cancer Dataset (WBCD, n=569) that FedXAI-DP achieves 98.25% accuracy and AUC-ROC =0.9950, exactly matches centralized training performance while assuring zero access to raw data. Even under realistic non-IID scenarios, the federated model only costs a 1.76% overhead compared to the centralized upper bound whilst achieving an accuracy of 96.49%. We perform a detailed privacy-utility tradeoff analysis across epsilon 2.0, 5.0, 10.0, and 20.0 that quantifies the cost in accuracy from formal privacy guarantees. SHAP analysis ranks radius_mean, area_worst and concave_points_worst as the clinical features most impactful, offering actionable explainability from a pathologist's perspective.