EEG Criticality + Naive Bayes beats deep learning for deep sleep staging

A new study published at the Graz BCI Conference 2026 demonstrates that physics-inspired EEG features can outperform conventional deep learning for automated sleep staging. Researchers Narębski, Komendziński, and Rutkowski used Detrended Fluctuation Analysis (DFA) to extract criticality features from EEG signals—a measure of neural system dynamics near phase transitions. They analyzed 347,232 30-second epochs from overnight recordings of 290 older women, visualizing state transitions with UMAP manifold learning. Six classifiers were benchmarked via 10-fold cross-validation using balanced accuracy to identify the optimal "state-sensing" engine for passive BCIs.

Surprisingly, a simple Naive Bayes classifier achieved the highest mean balanced accuracy (87.17% ± 0.24%), significantly outperforming a fully connected deep neural network (FNN: 81.58%) and Random Forest (80.97%). Linear models (LDA: 57.21%, SVM: 51.01%) performed poorly, confirming that DFA-derived criticality features reside on a distinct, non-linear manifold. This probabilistic decoding of EEG criticality provides a high-accuracy sensing mechanism for closed-loop neurofeedback. The authors suggest their pipeline supports development of state-dependent interventions—like precisely timed auditory stimulation—to enhance cognitive recovery and sleep quality in clinical populations.