Federated Learning for Predicting Mild Cognitive Impairment to Dementia Conversion

Area of Science:

• Artificial Intelligence in Medicine
• Neurodegenerative Disease Research
• Privacy-Enhancing Technologies

Background:

• Dementia is a progressive cognitive decline, with mild cognitive impairment (MCI) as a common precursor.
• Predicting MCI-to-dementia conversion is crucial for early intervention.
• Traditional machine learning (ML) methods for prediction require sharing sensitive clinical data, posing privacy risks.

Purpose of the Study:

• To propose and evaluate a privacy-enhancing Federated Learning (FL) framework for predicting MCI-to-dementia conversion.
• To enable collaborative model training without the need for sensitive data sharing among clinical sites.
• To compare the efficacy of FL against traditional centralized ML and site-specific models.

Main Methods:

• Implemented and compared two FL network architectures: Peer-To-Peer (P2P) and client-server.
• Trained predictive models using socio-demographic and cognitive measures within a federated environment.
• Assessed model performance against centralized ML models trained on pooled data and individual site-specific models.

Main Results:

• Federated learning achieved predictive performance comparable to centralized machine learning.
• Each participating clinical site demonstrated similar model performance without sharing local data.
• FL models outperformed site-specific models trained independently, highlighting the benefits of collaborative learning.

Conclusions:

• Federated learning offers a viable and effective solution for predicting MCI-to-dementia conversion while preserving data privacy.
• FL eliminates the necessity for sensitive data sharing, making collaborative research more feasible and secure.
• This approach maintains model efficacy and enhances predictive power through decentralized collaboration.