Secure federated learning for Alzheimer's disease detection

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Machine Learning (ML) shows promise for neuroimaging analysis but is hindered by data privacy concerns and the lack of large public datasets.
• Federated Learning (FL) offers a solution by enabling model training on decentralized data without compromising patient privacy.

Purpose of the Study:

• To train a Machine Learning (ML) model for Alzheimer's Disease (AD) detection using structural MRI (sMRI) data within a Federated Learning (FL) framework.
• To compare the performance of FL algorithms (Federated Averaging and Secure Aggregation) against centralized ML training.
• To investigate the impact of data heterogeneity (demographics, diagnosis, imbalance) on FL model performance for AD detection.

Main Methods:

• Implementation of two FL aggregation algorithms: Federated Averaging (FedAvg) and Secure Aggregation (SecAgg).
• Simulation of heterogeneous environments to assess the influence of demographic and diagnostic data distributions.
• Comparison of FL models with a centralized ML model trained on the same data.
• Evaluation of privacy guarantees using simulated membership inference attacks.

Main Results:

• FL models demonstrated comparable performance to centralized training for Alzheimer's Disease detection.
• Simulated heterogeneous environments revealed the impact of data distribution differences on FL model accuracy.
• Secure Aggregation (SecAgg) provided enhanced privacy protection against membership inference attacks compared to FedAvg and centralized methods.

Conclusions:

• Federated Learning (FL) is a viable and privacy-preserving approach for developing ML models for Alzheimer's Disease detection from neuroimaging data.
• Addressing data heterogeneity is crucial for robust FL model development in clinical applications.
• FL, particularly with Secure Aggregation (SecAgg), offers significant advantages in protecting patient privacy in medical AI research.