Extended Data Fig. 3: CNN accuracy by region as a function of high-pass and sample size reduction.

A, Balanced accuracy of CNNs trained and tested on progressively high-passed raw data from each recorded brain region (n = 45 implants, 9 animals, 10 regions). Dot color represents the region from which this model was trained. The region-colored line traces the average balanced accuracy of models trained on data from that region across various levels of high-pass. High-pass filtering significantly decreased brain state information above 1,000 Hz (y = -9.892e-05*x + 0.84, p < 0.0001, r2 = 0.43). B, To directly test the minimum time interval in which sleep and wake states reliably structure neural dynamics, we trained and tested a series of CNNs on single channel data, each model operating on a progressively smaller interval of data (from 2.6 s to 0 s). Accuracy declined as a function of number of input sample points (pearson correlation: r = 0.650464, p < 0.0001). Example data at various input sizes is shown below the x-axis. Model accuracy is shown as a function of region (marker color). Regional means are shown in colored lines (+/- SEM, shaded area).