Federated, explainable AI framework classifies Alzheimer’s disease from MRI and clinical data

Researchers have built an artificial intelligence framework that classifies Alzheimer’s disease from a combination of brain scans and clinical data, while keeping patient information private and explaining its decisions. The study, published in Multimedia Tools and Applications, reports a deep residual convolutional neural network that sorts subjects into three diagnostic groups: normal control, mild cognitive impairment, and Alzheimer’s disease.

The model was trained on multimodal data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), a dataset that combines high-resolution longitudinal structural MRI scans with rich tabular clinical metrics. According to the abstract, the ADNI data provide clean diagnostic separation between the three classes, which supports both multimodal fusion and explainability analysis.

On its own, the ResNet-CNN achieved a reported classification accuracy of 92.4 percent, with a macro-averaged F1-score of 91.7 percent and an AUC-ROC of 0.963. The authors then constructed a hybrid ensemble model that combines machine learning and deep learning techniques, which they report improved accuracy further, to 94.1 percent.

Interpretability is a central design goal of the framework. For the tabular clinical features, the team used SHAP and LIME to identify the features that contribute most to each prediction. For the MRI scans, Grad-CAM and saliency maps visualize which image regions drive the classification. The authors report that these highlighted regions align with clinically recognized biomarkers, such as hippocampal atrophy, giving clinicians a way to check whether the model is looking at diagnostically meaningful structures.

Privacy is addressed through federated learning, a training strategy in which a model is learned across multiple institutions without the raw patient data leaving each site. The implementation is based on the Federated Averaging (FedAvg) algorithm. The federated model reportedly reached 91.1 percent accuracy, which the authors describe as a negligible loss in performance compared with centralized training.

On top of federated training, the framework integrates additional privacy-enhancing mechanisms. Differential privacy is applied with parameters set to epsilon equal to 1.0 and delta equal to 10 to the power of minus 5, and secure aggregation is used to further mitigate the exposure of individual patient data during model training.

The study also explores two auxiliary techniques. Generative AI methods, specifically variational autoencoders and generative adversarial networks, were used to synthesize minority-class data, which the authors report improved recall by 12.2 percent and helped mitigate class imbalance in the dataset. In addition, a reinforcement learning study was run in simulation mode to optimize treatment pathways, yielding a reported 18.5 percent gain in cumulative reward over static policies. Because this reinforcement learning component was explored only in simulation, it represents a planning-stage exploration rather than a demonstrated clinical benefit.

The overall result, as summarized by the authors, is a scalable, interpretable, and privacy-aware diagnostic framework for Alzheimer’s disease. The reported performance figures come from the study’s own evaluation on ADNI data, and the abstract does not describe external validation on independent cohorts or deployment in clinical settings.

The work sits within a broader effort to apply AI and machine learning to neurodegenerative disease classification. By combining multimodal inputs, explanation methods that map onto known biomarkers, and federated training with differential privacy, the framework addresses several of the practical barriers, including data privacy and trust, that the authors identify as relevant to the secure and ethical adoption of AI in clinical settings.

Subject of Research: Federated and explainable AI for multimodal classification of Alzheimer's disease using MRI and clinical data

Article Title: Federated and explainable AI for accurate medical diagnosis of Alzheimer’s disease multimodal classification

Article References: Rashmi, U., & B M, B. (2026). Federated and explainable AI for accurate medical diagnosis of Alzheimer’s disease multimodal classification. Multimedia Tools and Applications, 85(11), Article 826. https://doi.org/10.1007/s11042-026-21983-3

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21983-3

Keywords: Alzheimer's disease, federated learning, explainable AI, deep learning, MRI, ADNI, differential privacy, SHAP, Grad-CAM, machine learning, neurodegenerative disease, healthcare AI