Federated AI Framework Pairs Medical Imaging Accuracy with Auditable Explanations

Researchers have introduced FedXAI, a federated learning framework designed to train medical imaging models across multiple hospitals without moving raw patient data, while simultaneously addressing three goals that existing systems usually treat separately: privacy, interpretability, and consistency of explanations across sites. The work, published in the journal Machine Learning, reports results on a single simulated benchmark rather than a live clinical deployment.

FedXAI combines three components in one architecture. The first, a Self-Explainable Federated Architecture, pairs an EfficientNet-B0 backbone with a 32-dimensional hybrid concept-bottleneck layer whose concepts are anchored to terms from the RadLex radiological ontology, so the model’s reasoning is tied to defined radiological vocabulary rather than being reconstructed after the fact. The second, a Federated Explanation Aggregation Protocol, extends standard server-side aggregation with a cosine-consistency penalty that pushes the explanation spaces of participating clients into alignment. The third, a Privacy-Preserving Explanation Verification module, audits the quality of explanations across clients using the subsampled Gaussian mechanism with Rényi differential privacy accounting.

In the reported evaluation, the framework was tested on the NIH ChestX-ray14 dataset, partitioned across five simulated hospital clients using a non-IID Dirichlet split with alpha equal to 0.5, a setup that mimics the uneven data distributions typical of real hospitals. FedXAI achieved a mean AUC of 0.7349, compared with 0.7193 for FedAvg, 0.7241 for FedProx, 0.7268 for SCAFFOLD, 0.7285 for FedBN, and 0.7301 for MOON. Cross-client explanation consistency rose from 83.6 percent to 99.9 percent under the new protocol.

The privacy guarantee applies specifically to the explanation-audit channel. The verification module provides formal (epsilon, delta)-differential privacy, with a certified epsilon of 0.1394 at delta of 10 to the power of minus 5. The authors note that model parameter transmission follows the standard federated learning architectural separation, while fully end-to-end differentially private SGD training remains future work, meaning the reported privacy certification does not extend to the entire training pipeline.

The study is a simulation-based benchmark on one public chest X-ray dataset with five synthetic clients, so the results do not establish performance in real multi-hospital deployments, and the reported AUC values indicate moderate diagnostic discrimination rather than clinical readiness. The framework’s contribution lies in showing that, on the benchmark studied, privacy, diagnostic accuracy, and interpretability can be advanced together through architectural integration rather than traded off against one another. Whether the gains hold under different partitioning schemes, larger client networks, and prospective clinical data remains to be tested.

Subject of Research: A federated learning framework for medical imaging that integrates intrinsic explainability, cross-client explanation consistency, and differentially private explanation auditing.

Article Title: FedXAI: Federated Medical Imaging with Intrinsic Explainability and Differentially Private Explanation Auditing

Article References: Rahman, M. A. (2026). FedXAI: Federated Medical Imaging with Intrinsic Explainability and Differentially Private Explanation Auditing. Machine Learning, 115(11), Article 253. https://doi.org/10.1007/s10994-026-07191-y

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07191-y

Keywords: federated learning, medical imaging, explainable AI, differential privacy, chest X-ray, ChestX-ray14, concept bottleneck, RadLex, non-IID data, privacy-preserving machine learning, Rényi differential privacy, EfficientNet