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

During slow-wave sleep, cortical slow oscillations-composed of alternating Up/Down states (UDS)-are crucial for memory consolidation. We classified UDS from local field potentials using machine learning models, specifically a hybrid CNN + RNN and a Transformer architecture, trained on labels derived from membrane potential recordings. Both models outperformed conventional methods, achieving superior classification accuracy and markedly reducing classification errors. Notably, the Transformer model also enabled real-time inference while maintaining robust performance. This real-time classification capability facilitates closed-loop feedback experiments to manipulate neuronal activity contingent on UDS phase, providing a powerful tool for investigating the causal mechanisms underlying sleep-dependent memory consolidation.