A research team from India has introduced a new AI system designed to detect Alzheimer's disease early – while protecting patient privacy and providing explainable decisions. The model, named FuzzyFed-CNN, combines image analysis of MRI scans with clinical data using a federated learning approach. The study was published in the journal Frontiers in Artificial Intelligence (DOI: 10.3389/frai.2026.1852196).
Background
Early detection of Alzheimer's is crucial for slowing disease progression. Previous AI models mostly analyze only MRI images or clinical data centrally – which raises data protection concerns and complicates the explainability of decisions. Furthermore, many systems lack the ability to adequately account for uncertainties in clinical data (e.g., in MMSE scores or hippocampal volume).
The Method: FuzzyFed-CNN
The proposed system processes two data streams in parallel:
- CNN Branch: Analyzes T1-weighted MRI images and extracts structural changes (e.g., hippocampal atrophy, cortical thinning).
- Fuzzy Inference Branch: Processes clinical and demographic data (age, MMSE score, hippocampal volume) using a rule-based fuzzy system capable of modeling uncertainties and linguistic variables (e.g., "high age" or "severe atrophy").
The two feature vectors are fused. Training is performed federated (Federated Learning) using the FedAvg algorithm: participating clinics train locally on their own data and only exchange model parameters – no raw data. This preserves privacy (HIPAA/GDPR compliant).
For explainability, Grad-CAM is used, which visually shows which brain regions the AI uses for its decision.
Results
The model was tested on a combination of ADNI and OASIS-3 datasets (403 subjects, divided into CN, MCI, and AD). It achieved the following values:
- Accuracy: 97.7%
- Sensitivity: 98.0%
- Specificity: 99.0%
- F1 Score: 98.0%
FuzzyFed-CNN significantly outperformed the comparison models MobileNet, ResNet50, DenseNet121, and EfficientNet-B0. Ablation studies showed that both the fuzzy component and federated fusion significantly contribute to the performance improvement.
Grad-CAM heatmaps confirmed that the model pays particular attention to clinically relevant regions such as the hippocampus and cortex.
Explainability and Data Privacy
A key advantage of the approach is the combination of high performance and interpretability. The fuzzy rules provide human-readable justifications (e.g., "high age + low MMSE + severe atrophy ? high risk"), while Grad-CAM visualizes the image-based decisions. At the same time, federated training ensures that sensitive patient data does not leave the clinic.
Evaluation and Outlook
The authors see FuzzyFed-CNN as a promising step towards privacy-preserving and explainable AI systems for neurodegenerative diagnostics. However, the results are based on public datasets. For real-world clinical use, further prospective studies in heterogeneous hospital environments and the integration of additional modalities (e.g., PET, blood biomarkers) are necessary.
FAQ
What is FuzzyFed-CNN?
An AI model that analyzes MRI images with clinical data using Convolutional Neural Networks and a fuzzy inference system – trained in a privacy-preserving federated learning approach.
Why is the method privacy-preserving?
The clinics do not share patient data, only model parameters. Training takes place locally on the respective servers.
How explainable is the model?
Through fuzzy rules (human-understandable decision rules) and Grad-CAM visualizations, which show which brain regions the AI looked at.
How good is the performance?
On the tested datasets, the model achieved an accuracy of 97.7%, significantly outperforming common CNN architectures.
When could such a system be used in the clinic?
Not immediately. Further validation studies in real clinical environments and with additional data modalities are required.
