Focal Changes to Human Electrocorticography With Drowsiness: A Novel Measure of Local Sleep

This study conformed to the Declaration of Helsinki. The protocol was approved by the institutional research board of the University of Texas Health Science Center at Houston (Institutional Review Board number, HSC-MS-15–0432). Recordings were obtained from adult (age >17 years old) patients suffering from medically refractory epilepsy undergoing intracranial monitoring prior to resective epilepsy surgery. All listed testing was completed within 6 months prior to the surgery, with the exception of the diseased hemisphere identity confirmed at the time of surgery. Subject characteristics are presented in Table 1.

We searched for evidence of local sleep using human ECoG recordings, allowing the capture of neural activity from multiple human cortical areas simultaneously. We measured the activity as changes in local field potentials (LFPs) in six human patients. The study was performed as a retrospective analysis of ECoG data from epilepsy surgery patients at the Texas Comprehensive Epilepsy Program. Cortical coverage included lateral, frontal, orbito-frontal, lateral temporal, basal and mesial temporal regions (Figure 1A). The patients underwent semichronic implantation with arrays of subdural platinum-iridium electrodes (PMT Corp., Chanhassen, Minn.) with a top-hat design (4.5-mm diameter, 3-mm diameter contact with cortex) embedded in silastic sheets (10-mm center-to-center spacing) using standard neurosurgical techniques. Sets of electrodes forming rectilinear grids or strips were placed subdurally during craniotomy. Electrodes were placed as dictated by the clinical needs of the patient after an in-depth review of each patient’s case at a multidisciplinary epilepsy surgery patient management conference. Data were recorded using a reference electrode placed in an area remote to that of the suspected seizure onset zone. Patients subsequently underwent several days of continuous video-ECoG monitoring (Nihon Kohden, Inc., Foothills Ranch, Calif.; raw signal was prefiltered in the passband 0.3–300 Hz and sampled at 1,000 Hz for ECoG recording) with their antiepileptic medications reduced to enable recording of habitual seizures. After recordings were completed, the analyses were performed using an average reference. To assess the impact of the choice of reference, the analyses were also performed using a single reference electrode selected to be remote from the seizure onset zone; this resulted in only minimal changes to the classifications. The raw ECoG data (Figure 1B) were analyzed at the single electrode level. We used 12 short recordings (1 hour–1.5 hours) from six subjects, with six recordings having only the wake state and six recordings having both the wake and sleep states. Additionally, we used six long recordings of 24 hours from four subjects. Electrodes with active interictal spiking were excluded from analysis (ranging from 6 to 10 electrodes per patient). Recording days during which seizures occurred were excluded from analysis. A human observer reviewed the video portion of each sample to identify a 1-hour section during the day when the patient was behaviorally clearly awake (morning circadian peak) and 1 hour at night after sleep onset when the patient clearly was and remained asleep. The sleep samples were chosen at time points prior to 80 minutes after sleep onset to attempt to limit the recordings to periods of non-REM sleep. When available, concurrent scalp EEG recordings were also reviewed for consistency with non-REM sleep. These 1-hour samples served as the baseline exemplar recordings (individual to each patient), from which the salient features associated with sleep and wake were derived. Except where otherwise specified, the data from each subject were analyzed independently of the other subjects. As this was a retrospective analysis, scalp recorded EEG and EMG data were not available so that formal sleep staging could not be performed. Due to fall precautions, patients undergoing invasive monitoring prior to surgical resection spent the majority of time sitting or lying down, and the videos with these recordings were reviewed to document significant changes in the behavioral state over time.

The instantaneous amplitude (IA) of a signal x(n) is the modulus of the corresponding analytical signal z(n)=x(n)+jHT[x(n)], where HT is the Hilbert transform.²⁸ IA is the positive envelope of the signal. The instantaneous frequency (IF) is the time derivative of the phase of analytical signal z(n), and it approximates the change with time of the signal x(n) frequency (for a short introduction of the IA and IF computation based on Hilbert transform, see the data supplement accompanying the online version of this article). We examined the change in IA and IF during wake and sleep states across EEG frequency bands defined as: δ: 0–4; δθ: 0–8; θ: 4–8; α: 8–14; θα: 4–14; β: 14–30; θβ:8–30; ϒ: 30–80, and βϒ:14–80 Hz. We used short records from six subjects containing both wake and non-REM sleep intervals. For each record, we selected 10 pairs of wake and sleep intervals of 500 seconds each. For each pair of wake and sleep intervals, we filtered the recorded ECoG signal into EEG bands with digital recursive elliptical filters²⁸ (see the online data supplement for the filters’ design). Next, for all EEG bands and channels, we computed IA and IF of the filtered signals (Figure 2A–2B). Finally, for each pair of wake and sleep intervals, and for each band, we made comparisons for all channels with rank-sum tests between the IA in wake and the IA in sleep, as well as between the IF in wake and the IF in sleep.²⁹ Then, for each band, we used a Holm-Bonferroni correction³⁰ to determine the threshold of significance for the results of the multiple rank-sum tests and computed the percentage of electrodes showing a significant increase/decrease in IA and IF in sleep versus wake state in that frequency band. The mean of the percentage changes in IA and IF, respectively, for 60 wake and sleep intervals selected from six subjects is shown in Figure 2. The IA in δ band (δIA) showed the highest increase from wake to sleep, while the IF in θα band (θαIF) showed the highest decrease from wake to sleep. These two features (δIA and θαIF) were subsequently used for determining the active/inactive state at the channel level.

We determined the active (wake) and inactive (sleep) state of channels (electrodes) by using a clustering procedure based on the k-means algorithm (Lloyd’s algorithm). The k-means clustering aims to partition N observations into K clusters in which each observation belongs to the cluster with the nearest mean. In our case, the observations represented all the pairs of smoothed, normalized IA in δ band (0–4Hz), denoted with δA, and smoothed IF in θα band (4–14 Hz), denoted by θαF, computed from the whole channel record (see the online data supplement for details about the computation of δA and θαF). Thus, for each channel we clustered, with k-means, N observations (δA(n), θαF(n)), N=1,2,.,N, computed from the channel record, into two clusters (K=2) for both the active and inactive states (Figure 3A). Lower amplitudes and higher frequencies characterized the active cluster. The cluster containing higher amplitudes and lower frequencies corresponded to inactive state. After clustering, we determined the state of the channel for every time moment as being inactive or active depending on the belonging of corresponding (amplitude, frequency) pair to the inactive or active cluster (binary wake sequence in Figure 3B, 1 for active, –1 for inactive). We also discriminated the active/inactive channel state by using one-dimensional k-means clustering for amplitude (Figure 4A) and for frequency (Figure 4B). We denoted by ΣIC(n) the sum (count) of inactive channels at the moment n (ΣIC(n) ≤M, M number of channels), obtained after the k-means clustering of channels based on δA and θαF (see Figure 3, Figure 5). The ΣIC was computed as often as δA and θαF: every 2 seconds for short records (around 1–2 hours long) and every 10 seconds for 24-hour long records.

We used the average across channels of normalized, smoothed IA in δ band, δA, as well as the average across channels of smoothed IF in θα band, θαF, to predict the number of inactive channels with a regression model. We denoted by mδA the average of δA across channels and with mθαF as the average of θαF across channels (Figure 5 [also see equations 9 and 10 in the online data supplement]). For each subject, we predicted the sum of inactive channels ΣIC as a function of averages across channels mδA and mθαF with a linear regression function:where SRC(n) was the approximation of ΣIC at the moment n and b_i (i=0,1,2,3) represented the linear regression coefficients, computed such that the mean square error between the SRC and ΣIC was minimized (Figure 3, Figure 5 [also see Figure S4C–S4D in the online data supplement]. We also predicted ΣIC by using only δA or only θαF. For these cases, ΣIC was the sum of inactive channels obtained after the k-means clustering of δA or θαF, respectively (Figure 4C).

We studied the relationship between δA and θαF, as well as the relationship between the sum of the inactive channels ΣIC, mδA, and mθαF by using the Pearson correlation (also see the online data supplement). First, we computed the histograms of the Pearson correlations between the channels’ δA and θαF for records having only the wake state (Figure 6A), as well as for records having both wake and sleep states (Figure 6B). The wake and sleep intervals were selected from the short records of all the subjects by inspecting the video records. Then, for each subject we computed the Pearson correlation between the averages across channels mδA and mθαF, R(mδA,mθαF), between mδA and ΣIC, R(mδA,ΣIC), and between mθαF and ΣIC, R(mθαF,ΣIC) (Figure 6C).

We used short records from six subjects containing wake and non-REM sleep intervals. Only one of the four patients had an additional limited montage of scalp electrodes placed. Review of the scalp electrode recordings confirmed sleep and wakefulness during the sections of the recording used as baseline wake and sleep. The presence or absence of REM sleep could not be assessed, but all the baseline sleep recordings were obtained within the first hour after clinical (video) sleep onset so that entry into REM sleep was unlikely. For each record, we selected 10 wake and 10 sleep intervals of 500 seconds each (Figure 2A–2B) and compared the medians of corresponding IA and IF in all channels and EEG bands. Figure 2C–2D illustrate the comparison in δ band for IA and θα band for IF, respectively. For each EEG band, and for each pair of wake and sleep intervals, we compared the IA in wake and IA in sleep, as well as the IF in wake and IF in sleep, in all channels by using rank-sum tests. We corrected the results of the multiple rank-sum tests with the Holm-Bonferroni method,³⁰ and then we computed for each subject the percentage of electrodes (channels) showing a significant increase/decrease in IA and IF in sleep versus wake state, in each frequency band. In Figure 2E–2F, we show the mean percentages for 60 pairs of wake and sleep intervals, selected from six subjects. We found the highest percentage of channels showing a significant increase of IA, sleep versus wake, in δ band (0–4 Hz) (Figure 2E). At the same time, we found the maximum percentage of channels showing a significant decrease of IF, sleep versus wake, in θα band (4–14 Hz) (Figure 2F). We concluded that the most salient features for inactive state determination were IA in δ band and IF in θα band.

It is widely accepted that the signal power in δ band can be related to the sleep state, and thus it could be used in sleep-state detection.^17,18 Our analysis adds a new salient feature for the sleep detection: the IF in θα band. The normalized, smoothed IA in δ band, denoted with δA, and the smoothed IF in θα band, denoted with θαF, were negatively correlated during the sleep state (see Figure 6B [also see the Methods section]). Thus, during sleep, the increase in δA is accompanied by a decrease in θαF. Based on this observation, we developed a detection scheme for the sleep state (termed “inactive” when applied to an individual channel) that simultaneously uses δA and θαF.

We estimated the active or inactive state of each channel (ECoG electrode) by clustering the δA and the θαF in two clusters corresponding to active and inactive states, respectively (see the Methods section). We used a k-means clustering algorithm having as inputs (δA(n), θαF(n)), computed for every 2 seconds, and in each subject separately identified two centroids corresponding to the active state (low amplitude, high frequency) and to the inactive state (high amplitude, low frequency), respectively. We assigned the state of the channel at the moment n as active or inactive depending on the shortest distance of the point (δA(n), θαF(n)) to the active or inactive centroids in the feature space (Figure 3A). Thus, for every 2 seconds, the classification determined the active or inactive state of each of the channels throughout the record (Figure 3B).

The sum of the inactive channels, denoted by ΣIC in the following, was interpreted as a measure of percentage of cortex in local sleep: during wake, ΣIC was significantly smaller than the number of recorded channels, while during sleep, ΣIC was close to the total number of channels. It is also possible to use only one feature like δA (Figure 4A), or θαF (Figure 4B), followed by one-dimensional k-means clustering to determine the state of the channel. However, we found that using only δA or only θαF led to a large underestimation or overestimation of ΣIC in comparison to using both δA and θαF (Figure 4D).

We included all of the active/inactive states in each of the channels in a restgram, defined as the matrix containing the variation of the subject channel states over time. The restgrams revealed evidence of inactive channels during wakefulness, consistent with the presence of local sleep. As an example, the restgram of subject C553 (Figure 3C) reveals local sleep: during wake period the majority of channels are active, but some are inactive. During sleep, the majority of channels are inactive, but some are active. Local sleep is also evident in the sum of inactive channels (ΣIC, blue line in Figure 3D). During wake (0–1,500 seconds), ΣIC had small positive values, showing that some channels are inactive. During sleep (t >1,500), ΣIC had higher values, but in most cases ΣIC < number of recorded channels, meaning that even in sleep some channels remained active.

As ΣIC might be a potentially important indicator of the global state of the brain, we approximated ΣIC with a linear regression function that used, as predictors, the average overall recorded channels of the δA and θαF denoted by mδA and mθαF, respectively (see equations 9 and 10 in the online data supplement). First, we determined the equation of the model by checking the interaction of mδA and mθαF with ΣIC, as well as the interaction between mδA and mθαF, in wake and sleep state (see Figure 6C). We found, in both wake and sleep, a high Pearson correlation between mδA and ΣIC, and between mθαF and ΣIC, (R(mδA,ΣIC) and R(mδA,ΣIC), respectively [Figure 6C]). We also found high negative correlations between mδA and mθαF (mean R(mδA,mθαF)=−0.74 [Figure 6C]) for the records having both wake and sleep states. This showed a clear interaction, at the level of the whole brain, between the IA in δ band and IF in θαF band: during sleep mδA increases while mθαF decreases. Given these high interactions, the regression model used as predictors mδA, mθαF, and their product mδA*mθαF (see equation 1 in the Methods section). For short records containing both wake and sleep states, the model (Figure 3D) explained a high percentage of the variability in ΣIC (R²=0.89 in Figure 3D, mean R²=0.92 in Figure 6C).

We used 24-hour records from four subjects to study the variation of instantaneous amplitude and frequency for periods of prolonged wakefulness during the entire day. The δA and θαF were computed for every 10 seconds (see the Methods section) to detect the wake/sleep state of each channel through the same classification scheme, applied to the entire 24-hour record. The restgram from Figure 4A shows evidence of local sleep that coincided with the wake periods of the subject, extracted from video record examination. For all subjects, we found strong correlation between mδA, mθαF, and ΣIC (mean R(mδA,ΣIC)=0.7, mean R(mθαF,ΣIC)=–0.83 in Figure 6C), as well as between mδA and mθαF (R(mδA, mθαF)=−0.89 in Figure 5, mean R(mδA,mθαF)=−0.59 in Figure 7). Therefore, we approximated the sum of inactive channels with a regression model similar to the one used for short records (see equation 1 in the Methods section). The regression model for 24-hour records explained a high percentage of ΣIC variability (R²=0.89 in Figure 5, mean R²=0.86 in Figure 6C).

As we used only video inspection and not a polysomnographic system, for the majority of the patients, our sleep-state analysis served only for a crude determination of the wake/sleep state (e.g., we considered the subject at sleep when the video inspection showed the subject laid in bed with eyes closed) so that changes to the ECog metrics over the course of a night’s sleep could not be meaningfully interpreted. However, the video inspection allowed us to select, with acceptable approximation, long periods of time when the subjects were awake during the 24-hour record (e.g., mean duration 3.1 hours [Figure 5]). We observed a clear increase in ΣIC during these wake periods. Figure 5 shows the increase in ΣIC for subject 1 in day 3, between 9:00 a.m. and 1:00 p.m. By approximating the change in ΣIC with a regression model (ΣIC(t)≈a+b_*t, mean R²=0.7 [Figure 5]), we found a consistent increase with time (b>0). By averaging the slopes b on all the selected wake intervals, we found an increase of eleven inactive channels per hour for wake intervals of approximately 3 hours long (Figure 5).

We illustrate the ΣIC increase with time in the wake state in Figure 7, where the activity of the lateral temporal occipital (LTO) grid of subject C544 (Figure 7A; also see Table S1 in the online data supplement) is shown at different time moments. When the subject was at sleep at 5:00 a.m., almost all the electrodes (channels) were inactive (Figure 7B). In a wake state, at 9:00 a.m., almost all of the channels were active, with only a few inactive channels (Figure 7B), while at 11:00 a.m., the number of inactive channels increased considerably (from 7 to 12 channels).