EEG Data Acquisition and Analysis

Preoperatively, a wireless, whole-scalp 16-channel silver/silver-chloride EEG system (Mobile-72 system, Cognionics, San Diego, CA, USA) was placed after measuring head circumference, and signal quality was established via laptop computer. The corresponding 10–20 montage channel locations are available in the Supplementary Figure S1. Data were recorded at 500 samples per second, and impedances were maintained below 100 kΩ per manufacturer recommendations. The EEG recording software was synchronized with the electronic medical record in order to timestamp critical events (e.g., induction, intubation, skin incision), and a research assistant remained in the room for the entire case to accurately document the timing of such events. After study completion, raw EEG signals were exported to MATLAB (version 2017a; MathWorks Inc., Natick, MA, USA) and down-sampled to 250 Hz. Full acquisition details are available as previously reported (Vlisides et al., 2019).

For this substudy analysis, EEG data were only analyzed from the maintenance anesthesia phase (i.e., from 30 s after skin incision to the last minimum alveolar concentration value of 0.7 towards the end of the procedure). Specifically, EEG data were analyzed from prefrontal (Fp1, Fp2), frontal (F5, F6, Fz), and parietal (P5, P6, Pz) regions given their postulated role in consciousness and anesthetic-induced unconsciousness (Ku et al., 2011; Koch et al., 2016; Flores et al., 2017). Functional connectivity among brain regions was estimated using a weighted phase lag index (wPLI; Vinck et al., 2011). This is a measure of phase synchronization that accounts only for non-zero phase lag/lead relationships. In this context, wPLI is relatively robust to volume conduction and reference montage. Between two neurophysiologic signals, if one signal consistently leads (or lags) the other, the phases are considered locked, and wPLI approaches 1 depending on the consistency of phase relationships. Alternatively, if the relationship between two signals is random, without any consistent phase relationship, then the wPLI value will be 0. To ascertain wPLI data, EEG signals were divided into 30-s windows at 10-s step sizes, which were then divided into 2-s sub-windows with 50% overlapping. The multitaper method (Mitra and Bokil, 2007) was then used to estimate the cross-spectral density with time-bandwidth product = 2 and the number of tapers = 3. WPLI values were then estimated, as a function of frequency, using a custom-written function adapted from the Fieldtrip Toolbox (Oostenveld et al., 2011). For this analysis, frontal-parietal, and prefrontal–frontal wPLI were calculated in the bandwidth between 0.5–35 Hz at 0.5 Hz step. Surrogate data were generated via the trial-shuffling method to mitigate potential bias of wPLI; subsequently, wPLI was calculated and subtracted from the original value as the final estimation of functional connectivity. Full methodological details related to wPLI are available as previously reported (Vlisides et al., 2019).

Lastly, temporal variations of connectivity were assessed over the entire anesthetic maintenance period as previously described (Vlisides et al., 2019). In brief, connectivity patterns were obtained using principal component analysis and k-means clustering. First, using principal component analysis, the 140-dimensional vector was reduced to 5-dimensional feature, and these patterns were classified into five clusters using the k-means algorithm with squared Euclidean distance and 100 replications of the initial centroids. The number of clusters and number of retained components were determined using the stability index, which quantifies the reproducibility of clustering solutions for the studied dataset, the amount of variance explained by principal components, and the interpretability of the clustering results (Lange et al., 2004). The results produced five distinct connectivity states, in addition to burst suppression, with distinct spatial and spectral properties over the anesthetic maintenance period (Vlisides et al., 2019). Given the focus on prefrontal-frontal and frontal-parietal oscillatory dynamics, these are the connectivity states further analyzed in this manuscript.

The cluster analysis also allowed for characterization of cortical connectivity data over time. For each subject, we quantified the occurrence rate that is defined as the fraction of time spent in a given connectivity state, compared to all states during anesthetic maintenance, for a given participant. We then assumed the connectivity state time sequence to be a Markov chain (i.e., the state transition depends only on the current state) and computed the state transition probability for each pair of states. We also computed the global state transition probability, which is defined as the number of state transitions between distinct states divided by the total time a subject spent in the anesthetic maintenance period.