FuzzyFed-CNN: secure and explainable multimodal federated learning for early Alzheimer's diagnosis

Alzheimer's disease (AD) is a progressive neurodegenerative condition that has a great effect on cognitive impairment and quality of life. Timely intervention requires the early and reliable diagnosis of the patient, but current diagnostic systems are frequently troubled with the limitations of data privacy, their lack of interpretability, and the fusion of heterogeneous clinical and imaging data.

The proposed study suggests FuzzyFed-CNN, an explainable and privacy-oriented multimodal FL system that incorporates CNNs as well as fuzzy inference systems to enhance the early detection of AD and model interpretability and data privacy.

The suggested framework involves CNN-based extractions of features using the T1-weighted MRI structural scans, and the use of the fuzzy-rule-based reasoning with the demographic and neuropsychological features, such as age, MMSE scores, and hippocampal volume. Experiments were done using a subset of the ADNI and OASIS-3 datasets. The training was conducted in a FL setting using the FedAvg algorithm. The metrics of accuracy, sensitivity, specificity, F1-score, and AUC were used to evaluate model performance.

Experiments that FuzzyFed-CNN with its accuracy, sensitivity, and specificity measure 97.7, 98.0, 99.0, and F1-score of 98.0. The suggested framework was better at performing compared to baseline models such as MobileNet and ResNet., DenseNet, EfficientNet. Grad-CAM visualizations also supported that the model paid attention to clinically significant brain regions, including the hippocampus and the cortical areas.

The results demonstrate that combining multimodal learning, fuzzy reasoning, and federated training can be used to achieve considerable improvements in the diagnosis of AD without damaging patient privacy and improving the interpretability of the models.