Personalizing brain stimulation: continual learning for sleep spindle detection.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Sleep Science

Background:

• Personalized stimulation algorithms are key for consistent brain stimulation quality.
• Sleep spindles are crucial for memory consolidation but challenging to target due to variability and fleeting nature.
• Real-time detection of neural events is necessary for effective closed-loop stimulation.

Purpose of the Study:

• To develop a continual learning framework for personalized sleep spindle detection.
• To enable subject-specific, real-time targeting of sleep spindles for closed-loop stimulation.
• To improve the accuracy and consistency of brain stimulation for research and clinical applications.

Main Methods:

• A novel continual learning framework was implemented using a pre-trained model for sleep stage classification and spindle detection.
• The algorithm refines spindle detection, adapting it to individual users over multiple nights without manual intervention.
• Weight averaging was employed to mitigate catastrophic forgetting and enhance performance over extended periods.

Main Results:

• The methodology achieved accurate, subject-specific targeting of sleep spindles.
• Fine-tuning alone showed minimal benefits for single nights, but weight averaging significantly improved performance over multiple nights.
• The approach effectively mitigated catastrophic forgetting, demonstrating robust personalization.

Conclusions:

• This work advances signal-level personalization for brain stimulation, applicable to various paradigms and neural events.
• The findings enhance the potential of brain stimulation for investigating cognitive processes like memory consolidation during sleep.
• The developed framework may lead to novel therapeutic applications by improving understanding and manipulation of neural activity.