FedTrust-XAI: Algorithmic Trust in Federated Health Communication Systems through Explainability and Privacy-Preserving Breast Cancer Diagnosis

Abstract

When using Artificial Intelligence(AI) in CDSS, two important and related problems come up: making sure that federated learning over institutional data silos doesn't violate patient privacy and getting doctors to trust AI models' diagnostic suggestions. The literature primarily focused on addressing each issue in isolation, resulting in a limited comprehension of their interconnections. This study proposes an innovative approach to address the two issues through the utilization of the FedTrust-XAI architecture, which integrates federated learning and explainable artificial intelligence (XAI). This study also examines its impact on diagnostic accuracy and the level of trust clinicians place in it. The proposed system employs federated logistic regression utilizing the FedAvg method across three distinct clinical sites and multiple interaction rounds. This lets the model learn together without sharing any information, which keeps patient privacy safe. The experiment using the UCI Wisconsin Diagnostic Breast Cancer (WDBC) dataset shows that the federated model can give diagnostic results that are just as good as those from the central model, while still keeping strong privacy protections. A method for understanding decisions is presented to help clarify how choices are made. We use linear XAI attribution, which uses model coefficients and feature inputs to give feature-level attribution scores. The model chooses worst_concave_points, worst_perimeter, and worst_radius as the most important things to look at when trying to figure out if a tumor is cancerous. These results are consistent with existing medical knowledge and enhance clarity. From a sociotechnical perspective, a factorial experiment involving physicians examines the influence of system architecture and explainability on the credibility of AI-generated advice. The federated structure and the ability to explain the output work together to make clinicians trust the system the most. Also, studies show that clinicians' levels of trust in federated AI systems depend on their own privacy concerns. In particular, doctors who care more about privacy issues are more likely to trust AI-powered solutions. The proposed architecture has an accuracy of 99.12% and an AUROC score of 99.23%, which is close to the accuracy of the centralized baseline (99.60% AUROC). The results show that you can learn without giving up your privacy or the ability to make predictions. Overall, FedTrust-XAI adds to important conversations in many areas about learning that protects privacy, AI that can be explained, and how people and AI work together in healthcare apps. The study shows that for AI technology to be accepted in healthcare, the algorithms must be clear and take privacy into account.