Privacy-Preserving Federated Learning Models for Accurate Diagnosis of Neurodegenerative Diseases in Distributed Healthcare Systems

Abstract

This paper presents a privacy-preserving federated learning framework aimed at the accurate diagnosis of neurodegenerative diseases, including Alzheimer’s and Parkinson’s, across distributed healthcare systems. Leveraging deep convolutional neural networks (CNNs) for image classification, we design a hybrid model capable of learning complex patterns from brain scan images. Our dataset includes 5,311 training images and 1,139 validation images classified into three categories: Normal, Alzheimer’s, and Parkinson’s. Extensive data augmentation techniques were applied to the training set to enhance generalization and mitigate overfitting. The hybrid CNN model achieved robust results after 30 epochs, with an overall test accuracy of 77.26% and a validation accuracy of 81.21% at its peak. The model performance correctly evaluated through all metrics including some test cases, achieving 83% accuracy for Alzheimer’s and 91% for Parkinson’s cases. The classification report and confusion matrix indicate that the model performs strongly in identifying neurodegenerative diseases, though some misclassifications remain in distinguishing normal cases. We also provide insights into model trade-offs by examining ROC-AUC curves, learning rates, and the effects of prediction confidence on diagnostic errors. Our results highlight the potential of federated learning in privacy-sensitive healthcare settings, particularly in providing accurate diagnoses while ensuring data privacy and resource efficiency.