Abstract Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, where early detection is crucial to slowing progression and improving patient outcomes. This study investigates the integration of Split Federated Learning (SFL) with neural networks for multiclass classification of AD stages: cognitively normal (CN), mild cognitive impairment (MCI), and AD. We evaluated ResNet-50 and DenseNet-169 across federated settings with 3, 5, and 7 clients, using FedAvg, FedAvgM, and FedProx as aggregation algorithms. In Phase I (548 patients), DenseNet-169 with FedAvg achieved 86. 97% ± 1. 14% accuracy. In Phase II (1287 patients), ResNet-50 with FedProx (batch size 32, 3 clients) outperformed all configurations, reaching 99. 26% ± 0. 73% accuracy, 99. 10% ± 0. 96% F1-score, and 99. 96% ± 0. 06% AUC. Statistical validation confirmed significant improvements from Phase I to Phase II (p 1. 0 for all metrics). In contrast, FedAvgM showed instability in most experiments despite occasional gains. These results demonstrate the effectiveness of combining SFL with deep learning for AD classification. ResNet-50 proved more suitable for large-scale datasets, while FedProx improved robustness under data heterogeneity. Overall, the findings highlight the need to align model architecture and aggregation strategies with dataset characteristics and confirm the potential of FL as a secure, scalable, and clinically applicable approach in medical imaging.
Mantegazine et al. (Wed,) studied this question.