Real-time classification of cortical slow-wave states by a machine learning model.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Sleep Science

Background:

• Cortical slow oscillations, comprising Up/Down states (UDS), are vital for memory consolidation during slow-wave sleep.
• Accurate classification of UDS is essential for understanding sleep-dependent memory mechanisms.

Purpose of the Study:

• To develop and compare machine learning models for classifying UDS from local field potentials.
• To evaluate the performance of hybrid CNN+RNN and Transformer architectures against conventional methods.
• To assess the potential of real-time UDS classification for closed-loop experiments.

Main Methods:

• Local field potentials were analyzed using machine learning models, including a hybrid CNN+RNN and a Transformer architecture.
• Models were trained using labels derived from simultaneous membrane potential recordings.
• Performance was evaluated based on classification accuracy and error reduction compared to traditional techniques.

Main Results:

• Both the hybrid CNN+RNN and Transformer models significantly outperformed conventional methods in UDS classification.
• The Transformer model demonstrated robust performance and enabled real-time UDS inference.
• Classification accuracy was superior, and errors were markedly reduced by the proposed machine learning approaches.

Conclusions:

• Machine learning, particularly Transformer models, offers a powerful and accurate method for classifying UDS during sleep.
• Real-time UDS classification facilitates novel closed-loop experiments to investigate causal links in memory consolidation.
• These advancements provide a valuable tool for advancing research into the neurobiological basis of sleep and memory.