Classifying Sleep Slow Oscillations in Low Density EEG

Slow oscillations (SOs, 0.5–1.5 Hz) are electrographic signatures of highly synchronous cortical events contributing to slow wave activity (SWA, 0.5–4 Hz) during sleep. An objective, EEG-manifold-based characterization of SOs could enable clinically actionable investigations on their functional role in cognition and health. In an earlier study, our group presented a data-driven approach showing that SOs group into Global, Frontal, and Local patterns, based on near-simultaneous detection across multiple scalp electrodes. However, this method requires sleep acquisitions with at least 24 head electrodes, limiting its applicability in standard diagnostic polysomnography and at-home studies. Here, we introduce a deterministic classification method that leverages machine learning to identify space-time profiles of SO types in low-density EEG (2–8 electrodes). Using an xgboost architecture trained on data from eight or fewer scalp EEG channels, our gradient-boosted tree model drew from more than 300 features to predict SO class (i.e., space-time profile) with high fidelity. Leveraging the feature-based nature of our classifier, we also used the SHAP feature-importance package to isolate the biophysical properties most characteristic of each SO type. Notably, we found that large amplitudes at central electrodes distinguish Global from Frontal SOs, while large amplitudes at frontal electrodes distinguish both types from Local SOs. Finally, we obtained strong predictions using only a minimal setup – 20 features gathered from only F3-F4-C3-C4. Using as few as four leads, this interpretable, low-density EEG approach enables routine spatiotemporal classification of SOs in clinical and research settings and clarifies the biophysical signatures that distinguish them.