Extended Data Fig. 4: Spectral density does not capture the full extent of high frequency state embedding.

To evaluate whether the embedding of brain state in > 750 hz activity is explained by spectral bandpower, a logistic regression (LR) was trained on either low frequency bandpower (1-16 Hz) or high frequency bandpower (750-3,000 Hz). The resolvability of state in the LR was compared to the performance of a CNN exposed to the same frequencies. Analyses consider the 71 single channel models that contribute to data in Fig. 3. A, Proportion of the CNN’s balanced accuracy that is achieved by LR. Data are presented as mean values +/- SEM. At both low frequencies (1-16 Hz blue), and high frequencies (750 - 3,000 Hz, red) the CNN substantially out performs a LR model using bandpower/FFT features. The difference between LR vs. CNN models is more pronounced in the 750 - 3,000 Hz range with the LR performance slightly better than ½ the balanced accuracy of CNN models. B, Unity line plot comparing the performance of 750 - 3,000 Hz LR versus 40 ms CNN models on a channel-by-channel basis. Note that points generally fall below the unity line, indicating higher balanced accuracies with the CNN than the LR. C, Confusion matrices for a single channel’s models - top is the CNN, bottom is the LR. Note that the CNN learns information about all three states - the diagonal is characterized by above chance accuracy (white lettering, red boxes). In contrast, the LR collapses to a two-state solution (sleep v wake). For A-C, n = 69 individual channels, where each implant is represented at least once. D,E To understand the frequency-based patterns that contain reliable state information, we examined the feature weights for the low and high frequency LR models. D, 14 high-accuracy 1 - 16 Hz LR model weights were averaged together. The signatures of state learned in a data-driven fashion are, unsurprisingly, highly consistent with human heuristics. For example, the 1-4 Hz (delta) band is highly weighted during NREM, and the 6-8 Hz (theta) band is highly weighted during REM. E, High frequency LR models did not exhibit consistent patterns across channels. Here we show two exemplary 750 - 3,000 Hz LR models. While each model learned to score state at > 50% balanced accuracy, the two weight distributions show distinct learning patterns, characterized by various irregularly interspersed frequency bands. This supplemental analysis employing logistic regression models trained on bandpower features provides valuable insight into the spectral components of dynamical signatures of sleep and wake states. While a CNN-based approach confirms the critical role of high frequency patterns in state classification, logistic regression analyses offer a complementary perspective, suggesting that state information persists in the high frequency domain, albeit less comprehensively than captured by CNNs. This dual approach underscores the complexity of neural embedding of states and the limitations of relying solely on traditional frequency domain analyses to understand such dynamics.