Aperiodic component of the electroencephalogram power spectrum reflects the hypnotic level of anaesthesia

Abstract

Background

Aperiodic (nonoscillatory) electroencephalogram (EEG) activity can be characterised by its power spectral density, which decays according to an inverse power law. Previous studies reported a shift in the spectral exponent α from consciousness to unconsciousness. We investigated the impact of aperiodic EEG activity on parameters used for anaesthesia monitoring to test the hypothesis that aperiodic EEG activity carries information about the hypnotic component of general anaesthesia.

Methods

We used simulated noise with varying inverse power law exponents α and the aperiodic component of EEGs recorded during wakefulness (n=62) and maintenance of general anaesthesia (n=125) in a diverse sample of surgical patients receiving sevoflurane, desflurane, or propofol, extracted using the Fitting Oscillations and One-Over-F algorithm. Four spectral EEG parameters (beta ratio, spectral edge frequency 95, spectral entropy, and alpha-to-delta ratio) and two time-series parameters (approximate [ApEn] and permutation entropy [PeEn]) were calculated from the simulated signals and human EEG data. Performance in distinguishing between consciousness and unconsciousness was evaluated with AUC values.

Results

We observed an increase in the spectral exponent from consciousness to unconsciousness (AUC=0.98 (0.94–1)). The spectral parameters exhibited linear or nonlinear responses to changes in α. Using aperiodic EEG activity instead of the entire spectrum for spectral parameter calculation improved the separation between consciousness and unconsciousness for all parameters (AUC_aperiodic=0.98 (0.94–1.00) vs AUC_original=0.71 (0.62–0.79) to AUC_original=0.95 (0.92–0.98)) up to the level of ApEn (AUC=0.96 (0.93–0.98)) and PeEn (AUC=0.94 (0.90–0.97)).

Conclusions

Aperiodic EEG activity could improve discrimination between consciousness and unconsciousness using spectral analyses.

Editor's key points.

• •Monitoring the electroencephalogram (EEG) provides valuable information about the hypnotic state during general anaesthesia, but it can be inaccurate.
• •This study analysed both simulated and patient data to evaluate the impact of the aperiodic components of EEG records in distinguishing between wakefulness and general anaesthesia.
• •There was an increase in the spectral exponent in transitions from consciousness to unconsciousness; use of aperiodic EEG activity instead of the entire spectrum for spectral parameter calculation improved the separation between consciousness and unconsciousness.
• •Monitoring aperiodic spectral EEG activity could help optimise EEG-based patient monitoring by improving discrimination between consciousness and unconsciousness.

Patient monitoring systems assessing the electroencephalogram (EEG) provide valuable information about a patient's brain during general anaesthesia.¹ However, using index-focused systems can lead to inaccuracies in the estimation of arousal levels.^2,3 In order to learn more about how signal properties track transitions in consciousness, we investigated the impact of aperiodic EEG activity on quantitative EEG parameters used for anaesthesia monitoring. For GABAergic anaesthetics, the shift from a high-frequency and low-amplitude signal during wake states to a slow signal with high amplitudes as arousal diminishes is readily appreciated in frontal cortex EEG leads.⁴

Aperiodic EEG activity is characterised by a decay of spectral power with frequency. In the log-log presentation of the power spectral density (PSD), the decrease follows an inverse power law: $PSD\left(f\right)=\frac{1}{f^{\alpha }}$.^5,6 Slowing of EEG frequencies during general anaesthesia increases the spectral exponent α.⁷ Here we used both simulated and patient data to evaluate the impact of the exponent change on EEG-based monitoring, extracting α with the aperiodic components from EEG records of wakefulness and general anaesthesia using the Fitting Oscillations and One-Over-F (FOOOF) algorithm.⁸ The behaviour of spectral parameters was investigated with simulated signals and human EEG data. Using simulated signals allowed us to systematically analyse changes in spectral parameters caused by changes in α with a defined, homogeneous dataset. Parameters that measure the complexity of time domain signals were also assessed as they perform equally or better⁹ than spectral approaches in assessing arousal levels. We also calculated permutation entropy (PeEn)¹⁰ and approximate entropy (ApEn),¹¹ and evaluated each parameter's performance in discriminating between anaesthesia-induced unconsciousness and the awake state.

Methods

Simulated EEG signals

We used simulated noise signals representative of aperiodic EEG activity with different spectral exponents α that were generated with the MATLAB dsp.ColoredNoise function (DSP Systems Toolbox, version 9.15; MathWorks, Natick, MA, USA). We created 1-min signals with spectral exponents ranging from α=0.7 to α=2.0 in steps of 0.1 (sampling rate 250 Hz). For each spectral exponent, we generated 20 independent samples. For preprocessing, we applied a bandpass filter (1–47 Hz) to mimic the frequency range typically used for anaesthesia monitoring.^12,13 Supplementary Figure S1 shows exemplary traces for different spectral exponents.

Patient EEG recordings

We used patient EEG signals recorded during wakefulness and general anaesthesia to transfer our findings to clinical EEG data sets. The analysed data were taken from two studies where all patients gave their informed written consent. The ethics committee at the Klinikum rechts der Isar, Technical University of Munich, Munich, Germany, approved the study protocols. The approval date was May 24, 2017,¹⁴ for the general anaesthesia data study and August 13, 2018, for the baseline data study.¹⁵ All data were obtained as 10-channel recordings with the electrodes placed according to the international 10–20 system. Our analysis used recordings from the (pre)frontal electrodes Fp1, Fp2, F3, and F4 since the frontal EEG information is used for anaesthesia monitoring. All recordings were bandpass filtered (1–47 Hz) before further processing. Data contaminated by artifacts were excluded. A random sampling of cases with various analgesic strategies, opioid dosing, and ages were chosen from our sample.

For the awake state, we used 30-s baseline EEG recordings with closed eyes (n=62, sampling rate 250 Hz).¹⁵ EEG recordings during maintenance of general anaesthesia were obtained from a different study.¹⁴ We extracted 1-min segments of non-burst suppression EEG mid-surgery to ensure stable surgical anaesthesia (n=125, sampling rate 250 Hz). General anaesthesia was maintained with propofol (n=58), sevoflurane (n=40), or desflurane (n=27).

Calculation of absolute and relative power spectral density

The absolute PSD was computed using the MATLAB pwelch function with a frequency resolution of 0.488 Hz. For the simulated data, the time series was used for PSD calculation. We calculated PSD for the four frontal EEG channels and used the median PSD of these four channels per patient for evaluation. This (median) absolute PSD was also normalised by dividing each power value of the absolute PSD by the cumulative power in the 1–47 Hz range. The relative PSD indicates the contribution of each frequency to the total power.

Fitting Oscillations and One-Over-F

To extract the aperiodic component from PSD, we used the FOOOF (version 1.1.0) algorithm⁸ executed in Python (Spyder, Python 3.9; Python Software Foundation, Beaverton, OR, USA) with further processing steps in MATLAB (version R2023b; MathWorks, Natick, MA, USA). We used the FOOOF algorithm with following parameters: parameterisation range (1.4–45) Hz; maximum width limits (2–12); maximum number of peaks=3; minimum peak height=0.3; maximum threshold (in standard deviations)=2.0. The aperiodic mode was set to fixed, that is, the aperiodic model equalled a linear fit in the log-log space. The FOOOF frequency was narrower than the input frequency range to avoid overfitting noise as small bandwidth peaks. We used two output variables of the model for subsequent analyses: the spectral exponent α and the aperiodic model, which is a PSD following a $\frac{1}{f^{\alpha }}$ function with the same frequency resolution as the input frequency vector, but with only valid values within the specified FOOOF frequency range. A visual representation of the aperiodic component is presented in Supplementary Figure S2.

Calculation of processed EEG parameters

Frequency domain measures

We evaluated the performance of four spectral parameters used in anaesthesia monitoring when exposed to the simulated signals and the aperiodic component of the EEG only. These parameters were: beta ratio,¹² spectral edge frequency 95 (SEF),¹⁶ spectral entropy (SpEn),¹³ and alpha-to-delta ratio.¹⁷ Beta ratio:

$$\text{beta}\text{ratio}={\log }_{10}\left(\frac{\text{spectral}\text{power}\left(30-47\text{Hz}\right)}{\text{spectral}\text{power}\left(11-20\text{Hz}\right)}\right)$$

is reported to be one of the processed EEG parameters used in calculating the bispectral (BIS) index.¹² However, some controversy exists about whether it is used specifically for level of anaesthesia estimation.¹⁸ SEF is the cut-off frequency below which 95% of the total EEG power is located within a defined frequency range.¹⁶ It is incorporated in the BIS monitor¹² and the patient state index.¹⁹ The SpEn evaluates the shape of the power spectrum by applying the Shannon entropy formula to the relative PSD¹³:

$$\text{SpEn}=-\sum \text{relative}\text{PSD}\times {\log }_{10}\left(\text{relative}\text{PSD}\right)$$

SpEn is part of the state and response entropy algorithm used in the Entropy Module.¹³ Alpha-to-delta ratio is a spectral parameter used in EEG research¹⁷ calculated by dividing the absolute power in the alpha frequency band (8–12 Hz) by the total power in the delta band (1–4 Hz).

Complexity measures for time series

We calculated ApEn and PeEn for the simulated signals and patient EEG recordings. In contrast to spectral methods, ApEn and PeEn also capture nonlinear or chaotic EEG behaviour. PeEn and ApEn have been successfully applied to EEG signals recorded during anaesthesia.^20,21 We calculated ApEn with an embedding dimension m=2, a tolerance r=0.2∗standard deviation, and a time lag of τ=1,²² and PeEn with embedding dimension m=3 and a lag of τ=1.²³ We evaluated the normalised PeEn with values ranging from 0 to 1. ApEn and PeEn were calculated independently from the recordings of each frontal electrode. Subsequently, the median of the four electrodes was calculated for each patient.

Visualisation of data and statistics

We used MATLAB for curve fitting (MATLAB Curve Fitter; Curve Fitting Toolbox version 23.2; MathWorks, Natick, MA, USA) to obtain the fitting parameter and the corresponding R² values and create the plots. The boxplots use the default settings with outliers being defined as points that are by more than the 1.5-fold interquartile range away from the bottom (25th percentile [Q1]) or top (75th percentile [Q3]) of the box. We calculated the effect size area under the receiver operating characteristic (AUC) with 10k-fold bootstrapped 95% confidence intervals (CIs) for statistical comparisons of the two groups using the MATLAB-based MES toolbox.²⁴ This is related to the Mann–Whitney U statistic²⁵ and a difference can be considered significant for P<0.05 if the 95% CI excludes AUC=0.5.²⁴ A rough classification was used to interpret AUC values: 0.7–0.8 was considered acceptable, 0.8–0.9 excellent, and >0.9 outstanding.²⁶ For AUC<0.5, the effect size is equivalent to 1-AUC.

Results

Simulated EEG data

Figure 1 shows exemplary raw traces of simulated data (Fig. 1a) and the derived median PSD plots (Fig. 1b: absolute; Fig. 1c: relative) for each trace with a different spectral exponent α. The four spectral parameters monotonically decreased as α increased (Fig. 2). For beta ratio and SpEn, the best fit was a linear model: (1) $\text{beta}\text{ratio}=p_1-p_2·\alpha$, p₁=0.31 (0.27–0.34) (p₁ (95% CI)), p₂=0.41 (0.38–0.43), R²=0.99; (2) $\text{SpEn}=p_1-p_2·\alpha$, p₁=1.14 (1.12–1.16), p₂=0.27 (0.26–0.29), R²=0.99. SEF change was best described with a quadratic fit ($\text{SEF}=p_1-p_2·\alpha -p_3·{\alpha }^2$, p₁=42.26 (38.72–45.81), p₂=7.65 (2.08–13.21), p₃=10.38 (8.34–12.42), R²=1.00). The alpha-to-delta ratio decreased exponentially (alpha-to-delta ratio = $p_1·e^{{-p}_2·\alpha }$, p₁=1.43 (1.30–1.55), p₂=1.50 (1.41–1.59), R²=0.99). The entropic parameters decreased monotonically following a linear (ApEn) or quadratic fit (PeEn): (1) ${\text{ApEn}=p}_1-p_2·\alpha$, p₁=1.54 (1.52–1.55), p₂=0.47 (0.46–0.48), R²=1.00; (2) $\text{PeEn}=p_1+p_2·\alpha -p_3·{\alpha }^2$, p₁=0.82 (0.81–0.83), p₂=1.11e-2 (0.49e-2–2.70e-2), p₃=0.02 (0.01–0.03), R²=1.00.

Fig 1.

Raw traces and power spectral density (PSD) of simulated EEG signals. (a) Example raw EEG traces of 5 s of simulated signals (sampling rate 250 Hz) for four different spectral exponents α. An increase in the exponent α was associated with increasingly slower dynamics in the raw signal. (b) Absolute PSD. Each line represents the group median for each spectral exponent ranging from α=0.7 (dark blue) to α=2.0 (light blue). (c) Relative PSD. Each line represents the group median for each spectral exponent ranging from α=0.7 (dark purple) to α=2.0 (light purple).

Fig 2.

Spectral and entropic parameters derived from simulated EEG signals follow the change in the spectral exponent α, either in a linear, quadratic, or exponential manner. (a) Beta ratio linearly decreases with α. (b) Spectral edge frequency (SEF) quadratically decreases with α. (c) Spectral entropy (SpEn) linearly decreases with α. (d) Alpha-to-delta ratio exponentially decreases with α. (e) Approximate entropy (ApEn) linearly decreases with α. (f) Permutation entropy (PeEn) quadratically decreases with α.

Patient EEG data during general anaesthesia

The spectral exponent shifts from wakefulness to general anaesthesia

Figure 3a shows exemplary EEG traces recorded during wakefulness or anaesthetic-induced unconsciousness. Figure 3b presents the absolute PSDs for both stages. As expected, the aperiodic model of the PSDs was characterised by straight lines in log-log plots, with a steeper slope during anaesthesia and a more uniform distribution of frequencies during wakefulness (Fig. 3c). This was also demonstrated by quantifying α under both conditions (Fig. 3d). The exponent α was significantly lower during wakefulness (α_awake=1.61 [1.40–1.80]; median [Q1–Q3]) compared with anaesthesia (α_{anaesthetised}=2.60 [2.36–2.79], AUC=0.98 (0.94–1)).

Fig 3.

Raw traces, power spectral density (PSD), and spectral exponents for EEG signals during wakefulness and anaesthesia. (a) Example raw EEG traces for EEGs recorded during wakefulness and anaesthesia. The awake EEG is characterised by higher frequencies and lower amplitudes, whereas the anaesthetised EEG shows lower frequencies with higher amplitudes. (b) The PSD group medians of the EEG recordings during wakefulness and anaesthesia. (c) The aperiodic model of the PSD. The shaded areas indicate the median absolute deviation. (d) Spectral exponent for EEG recordings during wakefulness and anaesthesia.

Response of spectral EEG parameters

When calculating the spectral parameters using the aperiodic model of recordings during wakefulness and general anaesthesia, the beta ratio decreased linearly ($\text{beta}\text{ratio}=p_1-p_2·\alpha$, p₁=0.25 (0.25–0.25), p₂=0.41 (0.40–0.41), R²=1.00, Fig. 4a), SEF and SpEn in a sigmoidal way ($\text{SEF}=p_1+\frac{p_2-p_1}{1{+\left(\frac{x}{p_3}\right)}^{{-p}_4}}$, p₁=39.33 (39.17–39.50), p₂=2.96 (2.84–3.07), p₃=1.79 (1.79–1.80), p₄=5.04 (4.98–5.10), R²=1.00, Fig. 4c; $\text{SpEn}=p_1+\frac{p_2-p_1}{1{+\left(\frac{x}{p_3}\right)}^{{-p}_4}}$, p₁=1.00 (1.00–1.00), p₂=0.14 (0.14–0.14), p₃=2.11 (2.11–2.12), p₄=2.55 (2.54–2.56), R²=1.00, Fig. 4e) and the alpha-to-delta ratio exponentially (alpha-to-delta ratio = $p_1·e^{{-p}_2·\alpha }$, p₁=1.51 (1.50–1.52), p₂=1.53 (1.52–1.53), R²=1.00, Fig. 4g). Calculating the spectral parameters with the aperiodic model instead of the original spectrum revealed a shift of the parameter values, especially during general anaesthesia (Fig. 4). Moreover, using the aperiodic model led to homogenisation of the parameter values compared with the original spectrum. Figure 4b, d, f, and h present the boxplots for the spectral parameter values calculated with the original and the aperiodic spectrum and the corresponding AUC values. The aperiodic model outperformed the original PSD in distinguishing between wakefulness and anaesthesia for all four spectral parameters, demonstrating outstanding separation of the groups (AUC_aperiodic=0.98 (0.94–1.00) for all parameters vs AUC_original=0.71 (0.62–0.79) to AUC_original=0.95 (0.92–0.98)).

Fig 4.

Spectral parameters calculated with both the aperiodic model and the original power spectral density (PSD) of the EEG recorded during wakefulness and anaesthesia. Beta ratio (a), SEF (c), SpEn (e), and alpha-to-delta ratio (g) were calculated with the aperiodic model and the original PSD of awake and anesthetised EEG signals and plotted against α. Dots represent the parameter values calculated using the aperiodic model. Using the aperiodic model resulted in a shift and homogenisation of the parameter values. By calculating the parameter values with the aperiodic model instead of the original PSD, the ability to discriminate between wakefulness and unconsciousness increased (b, d, f, and h). SEF, spectral edge frequency 95; SpEn, spectral entropy.

Discriminating capacity of entropic time-series parameters

Consistent with previous observations, both ApEn and PeEn decreased with increasing α (Fig. 5a and c). Assessing the discriminative capacity of ApEn and PeEn between wakefulness and anaesthesia revealed outstanding performance for both parameters (AUC_ApEn=0.96 (0.93–0.98), AUC_PeEn=0.94 (0.90–0.97), Fig. 5b and d). The AUC values of the entropic parameters were higher than those for the spectral parameters calculated from the entire PSD, but using the aperiodic model for spectral parameter calculation led to equally good, if not better AUC values.

Fig 5.

Approximate entropy (a) and permutation entropy (c) calculated with EEG recordings during wakefulness and general anaesthesia plotted against the spectral exponent α. Both parameters showed outstanding performance for discrimination between wakefulness and unconsciousness (b and d). ApEn, approximate entropy; PeEn, permutation entropy.

Discussion

The PSD slopes of frontal EEG signals, that is, the aperiodic exponents α, were superior in distinguishing wakefulness from anaesthesia-induced unconsciousness. Our findings show that signal properties of aperiodic EEG activity reflect the hypnotic component of general anaesthesia. Conversely, periodic EEG activity seems to be a nuisance factor for the assessment of ‘pure’ hypnosis. It should be noted that periodic information, in particular alpha power, might represent a marker for good quality of anaesthesia care, that is, when hypnosis and analgesia are in balance with surgical stimulation.²⁷

Physiological correlate of aperiodic EEG activity

Viewed in the frequency domain, aperiodic EEG activity, sometimes referred to as the PSD ‘background’,^5,7 decays according to the inverse power law: $PSD\left(f\right)=\frac{1}{f^{\alpha }}$.^5,6 The spectral exponent α can be estimated from the PSD slope presented in a log-log plot. Existing research indicate a shift in α from consciousness to unconsciousness during general anaesthesia.⁷ Simulation experiments performed by Gao and colleagues²⁸ demonstrated that the spectral exponent is indicative of the global balance between synaptic excitation and inhibition. Therefore, increased input of inhibitory neurones decreases the excitation-to-inhibition ratio and steepens the decay of the aperiodic component.²⁸ We used simulated signals and clinical data to investigate how spectral and entropic time-series parameters respond to changes in the exponent α, corresponding to a steepening of the PSD decay in the log-log representation. The simulated signals facilitated analysis of the dependence of EEG parameters on the exponent α in a homogeneous dataset with defined α. Although the range of exponents was restricted by the MATLAB function used for simulation, it appeared to be suitable for examining the transition from consciousness to unconsciousness.⁷ Nevertheless, the lower exponents can explain the relatively high SEF in the simulated data.

Our results demonstrate that all investigated spectral parameters (beta ratio, SEF, SpEn, and alpha-to-delta ratio) and the two entropies ApEn and PeEn react to changes in aperiodic EEG activity. The observed trend indicates a decrease in all parameters with the steepening of the PSD, occurring linearly (beta ratio, SpEn, ApEn) or nonlinearly (SEF, alpha-to-delta ratio, PeEn).

Aperiodic EEG activity during general anaesthesia

In this study, we compared aperiodic EEG activity during wakefulness and during maintenance of general anaesthesia with GABAergic anaesthetics. Our results showed an increased spectral exponent α from wakefulness to anaesthesia. This agrees with the existing literature reporting the steepening of the PSD under propofol anaesthesia,^7,29 which, as previously outlined, is a sign of inhibited signal propagation as observed during anaesthesia with GABAergic anaesthetics.²⁸ A recent study validated the spectral slope as a parameter for continuous evaluation of consciousness levels during propofol anaesthesia.²⁹ This aligns with our findings demonstrating that the spectral exponent α was an outstanding separator (AUC ∼0.98) of consciousness from unconsciousness. Given that the balance between synaptic excitation and inhibition appears to govern the aperiodic component, we tentatively gain mechanistic insights on transitions in consciousness at the systems neuroscience level through our study. GABAergic anaesthetics influence thalamic and cortical neuronal firing patterns by potentiating GABAergic inhibition, thus suppressing action potential generation and hyperpolarisation of the thalamocortical network.³⁰ This altered activity pattern manifests as slowing of the EEG, characterised by spindle-like alpha and background delta wave activity,^30,31 and, consequently, a steeper PSD decay.⁷

Aperiodic EEG activity tracks the hypnotic component of general anaesthesia

Reducing the EEG signal information to aperiodic EEG activity led to outstanding performance of the spectral EEG parameters compared with the original PSD for distinguishing hypnosis from wakefulness. When utilising the aperiodic model to calculate spectral parameters, performance increased to the level of the entropic parameters ApEn and PeEn, which have been suggested to perform equally, if not better, than spectral parameters derived from the original PSD.²¹ Interestingly, the spectral parameters calculated with the aperiodic model demonstrated the same performance in separating wakefulness and anaesthesia as the spectral exponent itself, suggesting that they adequately reflect the signal properties of the aperiodic spectrum in its entire frequency range. Therefore, our results support the idea that aperiodic EEG activity tracks the hypnotic component of general anaesthesia.

In contrast, periodic EEG activity, especially individual alpha oscillations, are processed by spectral EEG parameters in manifold ways, leading to inconsistent implications for EEG-based monitoring.³² This suggests that periodic EEG activity imparts greater complexity to EEG information, leading to inconsistencies when evaluating the hypnotic level of general anaesthesia. Contrary to aperiodic EEG activity, periodic EEG activity does not reflect the global excitation-to-inhibition ratio but is linked to neural population synchrony.³³ Alpha oscillations during general anaesthesia reflect thalamocortical oscillations³⁴ and seem to hold information about the quality of general anaesthesia rather than the hypnotic state; thus alpha oscillations could be a marker for adequate levels of general anaesthesia,³⁵ nociception,³⁶ and brain frailty.³⁷ However, low alpha power does not necessarily indicate inadequate quality of anaesthesia, for instance when additional hypnotics are administered.³⁸

Implications for EEG-based monitoring of general anaesthesia

Current EEG-based monitoring systems are designed to estimate ‘anaesthetic depth’, or the hypnotic component of general anaesthesia,³⁹ with the initial goal of preventing intraoperative awareness and conserving anaesthetic drugs.⁴⁰ These indices are based on evaluating spectral EEG components, the oscillatory power of selected EEG frequencies.³⁹ This approach is feasible because the EEG shifts from a high-frequency, low-amplitude signal during wakefulness to low-frequency, high-amplitude activity during general anaesthesia.⁴ This shift is reflected in the change of the spectral exponent, likely owing to a change in the global excitation-to-inhibition ratio induced by GABAergic anaesthetic drugs.²⁸ In addition to the overall slowing of the EEG, certain oscillations, such as dominant alpha rhythms, can develop under general anaesthesia.³⁵ These rhythms can create peaks in the spectral representation of the EEG. By applying the FOOOF algorithm, we performed a mathematical separation of the steepness of the spectrum (the aperiodic component) from the oscillatory activity. Our results suggest that this separation is essential for improving the quality of depth of anaesthesia monitoring. The steepness of the spectrum appears to carry hypnotic information, whereas the role of the oscillatory component is less clear. Recent studies suggest that alpha oscillations are associated with analgesia,³⁶ the quality of anaesthesia,³⁵ or the brain's health.³⁷ Therefore, both aperiodic and periodic EEG components should be considered separately when assessing the EEG visually or automatically.

Limitations

Our investigation was of a technical and signal-analytical nature, and purposefully did not control for drug concentrations or demographic factors. Data were generated during general anaesthesia under propofol, sevoflurane, and desflurane. While sharing GABAergic mechanisms, the EEG signatures of these anaesthetics differ owing to additional molecular targets.³⁵ However, current monitoring systems follow a one-size-fits-all approach and also do not include this information. Further studies are needed to evaluate the detailed dynamics of aperiodic component parameters during surgeries with varying nociceptive stimuli, different anaesthetics, and different patient cohorts. Previous studies identified deviations from linearity in the PSDs of EEG recordings under anaesthesia, characterised by ‘knee’-frequencies around 20 Hz.^7,28 We fitted the FOOOF with a default setting without a ‘knee parameter’. This could influence the separation between states as the fit increases delta power. The impact of different FOOOF settings needs to be evaluated in the future. Finally, we chose the four frontal EEG electrodes Fp1, Fp2, F3, and F4 for analysis. Commercial EEG systems typically use proprietary electrode layouts. Although we can offer strong theoretical evidence supporting the superiority of aperiodic EEG, this must be validated through simulation experiments using actual monitoring systems.

Conclusions

Spectral EEG information, reduced to the aperiodic component, was superior in distinguishing the anaesthetised from the awake state. This finding could help to optimise EEG-based patient monitoring in the future.

Authors’ contributions

Analysed the data: SW, JO, MK

Discussed the results: SW, JO, SZ, SP, PSG, SK, GS, MK

Helped to write the manuscript: SW, JO, SZ, SP, PSG, SK, GS, MK

Declarations of interest

SK, MK, PSG, and GS are co-inventors on several patents related to intraoperative EEG analysis owned by Columbia University and TUM. PG is co-founder of Lantern Laboratories, Inc. that has a license to build software and hardware for intraoperative monitoring. MK received funding from Masimo Corporation, Narcotrend-Gruppe, Medtronic GmbH, and Fresenius Kabi Deutschland GmbH for conducting EEG-based training for anaesthesiologists.

Funding

Elite Network of Bavaria (Elite Graduate Program Biomedical Neuroscience, S-LW-2016-351/2/58 to SW).

Handling Editor: Hugh C Hemmings Jr

Appendix A. Supplementary data

The following is the Supplementary data to this article: