Decentralized Data-Privacy Preserving Deep-Learning Approaches for Enhancing Inter-Database Generalization in...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Sleep Medicine

Background:

• Automatic sleep staging is crucial but faces challenges in clinical adoption due to poor generalization across diverse datasets.
• Data privacy restrictions further complicate the development and deployment of reliable automatic sleep scoring systems.
• Existing methods struggle to perform consistently across data from different clinical settings.

Purpose of the Study:

• To propose and evaluate decentralized deep learning approaches for robust, privacy-preserving automatic sleep staging.
• To assess the generalization capabilities of ensemble models and federated learning across heterogeneous sleep databases.
• To compare decentralized methods against traditional single-database and centralized multi-database models.

Main Methods:

• Exploration of four ensemble strategies: max-voting, output averaging, size-proportional weighting, and Nelder-Mead.
• Introduction of a novel federated learning algorithm: sub-sampled federated stochastic gradient descent (ssFedSGD).
• Evaluation using a leaving-one-database-out cross-validation on six independent sleep staging databases.

Main Results:

• Decentralized learning methods demonstrated superior performance compared to baseline local models.
• The proposed approaches achieved generalization results comparable to centralized multi-database models.
• Ensemble and federated learning effectively addressed inter-database performance variability.

Conclusions:

• Decentralized deep learning offers a promising solution for improving automatic sleep staging generalization and data privacy.
• These methods provide advantages in scalability, design flexibility, and privacy preservation over traditional approaches.
• The findings support the adoption of decentralized techniques for more robust and widely applicable automatic sleep scoring systems.