Aperiodic Activity Reflects Pathologic Waveform Shapes in Focal Epilepsy

Abstract

Epilepsy constitutes a clinically manifest excitability disorder that is characterized by aberrant electrophysiological activity in the electroencephalogram (EEG). The correct identification of the seizure onset zone relies on the visual detection of pathological waveforms and the assessment of their morphology, rhythmicity, and density. Recent advances in quantitative EEG analyses indicated that aperiodic EEG background activity might provide complementary information to traditional qualitative methods. Importantly, aperiodic activity, and specifically the slope of the 1/ƒ^χ decay function of the power spectrum, might constitute a biomarker of the underlying population excitability dynamics. Hence, in the context of epileptic activity, an altered spectral slope is often considered as a signature of pathological excitability. To date, it remained unclear if this straightforward interpretation also applies to states of manifest seizure activity. To address this question, we recorded intracranial electroencephalography (iEEG) during focal seizures from patients diagnosed with pharmacoresistant epilepsy (18 patients, 11 females). The results demonstrate that the spectral slope successfully delineates seizure activity. However, the spectral slope was sensitive to the presence and waveform shape of distinct epileptic components. By combining iEEG recordings with simulations, we demonstrate that epileptic spiking activity and associated slow-wave components differentially impact spectral slope estimates. These results offer a more parsimonious explanation for the biophysical origins of aperiodic activity as compared with the concept of an underlying balance between excitation and inhibition.

Significance Statement

It is under debate whether the nonoscillatory electroencephalogram background (aperiodic) component of electrophysiological brain activity provides diagnostic insights into clinical disorders such as epilepsy. Seizure-related changes in aperiodic activity have mostly been ascribed to pathological excitability. Yet, epileptic waveform shapes could potentially modulate spectral estimates of aperiodic activity. Heidiri et al. demonstrate that aperiodic activity tracks ictal activity, wherein the morphology and density of epileptic activity systematically distort spectral estimates and associated aperiodic estimates. These findings show that aperiodic activity is directly modulated by the presence of waveform shapes and may not necessarily reflect pathologic aberrations in the underlying population excitability.

Introduction

Electroencephalography (EEG) is an important diagnostic tool to detect abnormal brain activity in neurological conditions. Clinical EEG interpretation is descriptive and assesses the presence of pathological waveform shapes as well as the occurrence and distribution of prominent EEG rhythms. Visual assessment often neglects nonoscillatory EEG background activity as it is difficult to evaluate; wherein spectral power follows an inverse power scaling law as a function of 1/frequency^χ (1/ƒ^χ). Recent advances in quantitative EEG analysis by means of spectral parametrization, which disentangle oscillatory from nonoscillatory (aperiodic) EEG components, provided new means to evaluate aperiodic EEG background activity (Wen and Liu, 2016; Donoghue et al., 2020).

Computational modeling proposed an elegant biophysical interpretation of aperiodic activity, wherein the spectral slope (negative exponent χ) might index the excitation-to-inhibition ratio (E–I ratio) of underlying neural circuits (Gao et al., 2017). Specifically, a flattening of the spectral slope may reflect more excitability or a net decrease in inhibition, while a steepening signals less excitability or a net increase in inhibition. Related modeling and experimental work (Miller et al., 2009; Chini et al., 2022) provided evidence that the spectral slope indexes population dynamics. A steepening of the power spectrum was observed during general anesthesia with propofol (a GABA_A agonist) (Gao et al., 2017; Colombo et al., 2019; Lendner et al., 2020; Waschke et al., 2021) and physiologic sleep (Lendner et al., 2020, 2023; Kozhemiako et al., 2022; Ameen et al., 2024), whereas a flattening of the spectral slope was observed during cognitive engagement (Waschke et al., 2021) and physiologic aging (Voytek et al., 2015). Complementary studies in animal models confirmed a link between the spectral slope and excitability dynamics using simultaneous single-unit and field potential recordings (Diehl and Redish, 2024), two-photon imaging of large populations (Lendner et al., 2023), and optogenetic interventions (Chini et al., 2022).

This converging evidence enables testing if aperiodic activity constitutes a biomarker that provides diagnostic insights into clinical disorders. Pathologic alterations of the spectral slope have been observed in several neurological diseases, including epilepsy (Kundu et al., 2023; Kopf et al., 2024; Liao et al., 2024), stroke (Johnston et al., 2023; Maschke et al., 2023), tumors (Numan et al., 2021), traumatic brain injury (Maschke et al., 2023), dementia (Martínez-Cañada et al., 2023; Rosenblum et al., 2023), and Parkinson's disease (Darmani et al., 2023; Rosenblum et al., 2023; Bush et al., 2024). Epilepsy is considered a state of aberrant cortical excitability, wherein altered aperiodic activity has been observed before interepileptic discharges (Charlebois et al., 2024; Kopf et al., 2024), especially near the seizure onset zone (SOZ; van Heumen et al., 2021; A. I. Yang et al., 2023; Liao et al., 2024), and might also track successful treatment outcome (Coa et al., 2022; Kundu et al., 2023; Y. Yang et al., 2023).

Most studies interpreted changes in aperiodic activity as a modulation of the E–I ratio (Donoghue, 2024). However, it is critical to consider that epilepsy is hallmarked by pathological, high-amplitude waveform shapes that contain multiple spike and slow-wave components. As aperiodic activity is typically derived from spectrally decomposed (Fourier-transformed) time series data, epileptic waveforms may systematically bias spectral estimates, thereby influencing aperiodic parameters such as the spectral slope.

Since the majority of studies focused on interictal epileptic activity, the goal of this study was twofold. First, we tested if aperiodic activity delineates ictal activity during seizures, which constitutes an extreme case of aberrant cortical excitability and features a wide range of pathological waveform shapes. We assessed intracranial EEG activity before, during, and after manifest seizure activity to test if aperiodic activity tracks seizure progression within the SOZ. Subsequently, we tested to which extent distinct waveform features influence spectral estimates of aperiodic activity. In sum, we queried the clinical utility of spectral parametrization methods to assess EEG background activity and tested a more parsimonious interpretation of altered aperiodic activity in epilepsy: altered aperiodic activity reflects pathological waveform shapes and not necessarily altered E–I dynamics.

Materials and Methods

Participants

Intracranial EEG recordings (N = 31) were obtained from patients diagnosed with pharmacoresistant focal epilepsy during presurgical monitoring (N = 18, 11 female; M_age = 32.78 ± 3.71 years, mean ± SEM). Depth electrodes (DIXI medical) consisted of 5–15 contacts per electrode with variable interelectrode spacing [sampling rate of 512 Hz (N = 30) or 2,048 Hz (N = 1)]. All electrodes were placed stereotactically via robot-assisted surgery (ROSA One, Zimmer Biomet). Electrode localization was dictated by clinical needs to localize the SOZ. All recordings were conducted at the University Medical Center of Tübingen. All patients provided written informed consent prior to recordings. Procedures were approved by the Ethical Committee at the Medical Faculty of the University of Tübingen.

Procedure

A total of 31 spontaneously occurring (sub)clinical seizures were included in the analyses, and multiple seizures (up to 3) of the same patient were included if available. Seizures were selected to sample a diverse set of waveform shape patterns, hence avoiding to include multiple stereotypical seizures of the same patient. In addition, seizures following electric stimulation or those requiring termination by administration of benzodiazepines were excluded. Additional 5 min resting-state (RS) recordings were obtained from every patient and required maintaining fixation on a central cross presented on a laptop during bedside testing.

Data processing

Intracranial EEG data were band-stop filtered at 50 Hz and its harmonics, demeaned, detrended, rereferenced to a bipolar montage (next lateral neighboring electrode), and downsampled to 500 Hz. Artifacts were manually flagged and discarded. No electrodes were excluded from analysis as this study investigated pathologic activity. (Peri-)Ictal data were z-normalized relative to the RS recording. Data processing and analyses were conducted using MATLAB (The MathWorks Inc., 2023). Preprocessing and analyses utilized custom code in addition to functions from the FieldTrip toolbox (Oostenveld et al., 2011).

Seizure classification (conditions)

All seizure recordings were visually classified under supervision of a trained neurologist (Fig. 1A). The SOZ channel, as well as timepoints for seizure onset and seizure termination, was determined according to established clinical criteria (Frauscher et al., 2024). The interval preceding the seizure was classified as preictal, the seizure itself as ictal, and the interval following the seizure as postictal. Only postictal intervals that were not directly followed by another seizure were included in the analyses.

Figure 1.

Aperiodic activity delineates seizure activity in patients with focal epilepsy. A, Single-subject iEEG coverage (black circles). Time-domain signals in the SOZ (red) and a representative channel outside the SOZ (non-SOZ; black) channels are displayed during different conditions (RS, preictal, ictal, and postictal). B, Single-subject example of an ictal segment in the time domain (left) and the frequency domain (right). The FOOOF algorithm models the spectrum (dashed red line) by decomposing it into an aperiodic component (solid red line) and a periodic component. C, Group coverage (N_patients = 18) of iEEG channels using the same conventions as in panel A. D, Power spectra in log-log space per condition (group level, N_seizures = 31; mean ± SEM). E, Spectral slope per condition (group level, N_seizures = 31; p < 0.0001 LRT). P values reported from LME models. Data are visualized as probability density functions; boxplots indicate median, first/third quartiles, and extreme values. Note the flattening of the spectral slope during the ictal condition.

Data analyses

Spectral analyses

Condition-specific power spectra (Fig. 1D) and spectral slope estimates (Fig. 1E) were obtained following segmentation of seizure recordings (3 s segments, 90% overlap), wherein each segment was assigned to the respective condition (RS, preictal, ictal, or postictal). Time–frequency decomposition was implemented per segment by means of a fast Fourier transform (FFT) from 1 to 120 Hz in log₂(1/8) Hz steps with 2 Hz frequency smoothing. FFT analyses implemented a multitaper approach based on discrete prolate Slepian sequences (dpss; 11 tapers) to obtain smooth spectral estimates. Aperiodic activity was separated from periodic peaks within 1–120 Hz utilizing the FOOOF algorithm (Donoghue et al., 2020). This wide frequency range was utilized to ensure reliable spectral parametrization (Text S1; Fig. S1A). The FOOOF algorithm approximates periodic activity by iteratively fitting Gaussian functions to the spectral estimate that are then subtracted from the power spectrum. Aperiodic activity (i.e., spectral slope and offset) was subsequently estimated by fitting a first-degree polynomial to the peak-removed spectrum in log-log space (fixed aperiodic mode; see Text S3 and Fig. S2 for supporting analyses using a knee aperiodic mode). Spectral parametrization utilized the following parameters: max number of peaks, 3; peak width limits, 0.5–12 Hz; min peak height, 3 dB; peak threshold, 2 standard deviations; and aperiodic mode, fixed. Model fits were assessed through average R² and model error per seizure. Note that R² values depend on the number of samples (here, data points in the power spectrum), as well as the number of parameters (here, spectral slope, offset, and knee). To facilitate comparisons between model fits in different frequency ranges, as well as between fits with and without spectral knees, we adjusted the R² values (Hittner, 2020; Preston et al., 2025). Note that the maximum number of peaks for spectral parameterization was fixed across analyses (max number of peaks, 3) and therefore does not bias the R² statistic.

Analyses only included conditions with a minimum length of 10 s, corresponding to a minimum of 24 segments per condition given the 90% overlap. This minimum condition length reflects a trade-off between including short seizures and ensuring reliable spectral parameterization, which is known to be more stable in longer segments (Wiesman et al., 2022; for control analyses on increased condition lengths, see Text S4).

To control for the influence of seizure-specific condition length on spectral estimates, we matched the number of segments through a permutation procedure. For each iteration (N_iterations = 1,000), we randomly selected 24 segments (without replacement) and computed the spectral estimates (slope, offset, and power spectrum) through averaging across the 24 segments. These iteration averages were subsequently used to calculate mean spectral estimates per condition and seizure (Fig. 1D,E). To estimate the effect of condition on spectral slope, linear mixed effect models (LME) were performed on iteration means (Fig. 1E). Patient and seizure ID were included as independent variables to control for patient- and seizure-specific effects. The likelihood ratio test (LRT) was performed to compare the full model [slope ∼ 1 + condition + (1 | patient_ID) + (1 | seizure_ID)] against the baseline model [slope ∼ 1 + (1 | patient_ID) + (1 | seizure_ID)].

Continuous spectral slope and band-limited spectral power (Fig. 2B) were obtained using the same conventions (3 s segments with 90% overlap, FFT, dpss, spectral slope determined within 1–120 Hz). Band-limited spectral power was calculated per frequency for the following frequency ranges: 0.5–4 Hz (delta), 4–8 Hz (theta), 8–13 Hz (alpha), 13–30 Hz (beta), 30–80 Hz (gamma), and 80–200 Hz (high-frequency activity; HFA) in 1 Hz steps. Note that for all spectral analyses, frequencies at 50 ± 3 and 100 ± 3 Hz were omitted to avoid line noising leaking into adjacent frequencies through spectral smoothing.

Figure 2.

Ictal patterns differentially impact the spectral slope. A, Exemplary 5 s segments within the ictal condition, illustrating the diversity of ictal patterns including LVFA, LFPS, and polyspike-wave discharges (PSWD). Variations in the time domain (left column) translate to the frequency domain (right column) and can be quantified by the slope (sl) of the aperiodic activity (red). B, Single-subject example of seizure progression. From top to bottom: z-normalized broadband time series, spectral slope, epileptic spike density, and band-limited spectral power (mean ± SEM across frequencies). Note that both the highest (arrows) and lowest (arrowhead) spectral slope estimates occur during the ictal condition. During the LVFA ictal pattern, a flattening of the spectral slope coincides with increased spike density and power increases in higher-frequency bands. During polyspike-wave discharges, a steepening of the spectral slope coincides with increased delta power. C, The spectral slope as a function of epileptic spike density (group level). Each point represents a 3 s segment during the ictal condition (N_segments = 11,042). P value reported from LRT. Data visualized using probability density functions; box plots represent median, first and third quartiles, and extreme values. D, Spectral slope as a function of delta power (group level; N_segments = 11,042). P value reported from LRT. Data visualized using the same conventions as in panel C. E, Spectral slope for two prevalent seizure patterns, group level; N_LVFA = 2,750; N_LFPS = 1,728. P value reported from LRT. Data visualized using the same conventions as in panels C and D.

Epileptic spike density

Spike density was determined by detecting peaks in the HFA band (80–200 Hz). Data were high-pass filtered (80 Hz) and z-scored. Peak detection was performed on the continuous filtered signal, implementing a minimum peak height of five standard deviations (SD) and a minimum peak distance of two samples, the latter corresponding to $\frac{\mathrm{sample}\mathrm{frequency}}{200\mathrm{Hz}}$. The detected peaks were subsequently assigned to 3 s segments to facilitate comparisons with spectral slope estimates. Spike density (the number of spikes per second; Hertz) was determined for each 3 s segment.

Relationship between spectral slope and waveform features

We assessed the relationship between the spectral slope and waveform features specifically during seizures, thus including all 3 s segments from the ictal condition (N_segments = 11,042). In particular, we examined whether the spectral slope changes as a function of epileptic spike density (Fig. 2C) and delta power (Fig. 2D). Delta power was obtained in the 0.5–4 Hz range using the same conventions as before (3 s segments with 90% overlap, FFT, dpss). LME models were fitted to all segments of the ictal condition (Fig. 2C,D). For each feature, the full model including spike density or delta power as a fixed effect (slope ∼ 1 + spike_density/delta_power + (1 | patient_ID) + (1 | seizure_ID)] was compared against the baseline model [slope ∼ 1 + (1 | patient_ID) + (1 | seizure_ID)] using the LRT.

Spectral slope per seizure pattern

Ictal segments were classified into the two most prevalent seizure patterns, namely, low-voltage fast activity (LVFA) and low-frequency periodic spiking (LFPS; Abdallah et al., 2024). The spectral slope difference between LVFA and LFPS was assessed by fitting LME models. Specifically, the model including the seizure pattern [full model; slope ∼ 1 + seizure pattern + (1 | patient_ID) + (1 | seizure_ID)] was compared against the baseline model [slope ∼ 1 + (1 | patient_ID) + (1 | seizure_ID)] using the LRT.

Simulation of pathologic waveform shapes

Simulation of spike–SW complexes

Epileptic spikes were approximated by combining two Gaussians into spike–slow-wave complexes (spike–SW; Fig. 3A). The first Gaussian simulated the spike component (10 ± 2.5 ms, mean ± SD) with an amplitude of 5 SD relative to the broadband signal to which the spikes were added. The second Gaussian simulated the SW following the spike component (255 ± 150 ms; mean ± SD) with a relative amplitude one-third of the spike component. Spike–SW complexes were added with increasing density (0–0.25 Hz in 0.025 Hz steps; 100 iterations per step) to simulated 1/f^χ noise with a spectral slope of −2. Power spectra and aperiodic activity were determined as outlined above (fixed aperiodic mode, Fig. 3B; for models fitting a knee, see Text S5, Fig. S3). The effect of spike–SW complex density on the spectral slope was analyzed utilizing a linear regression (slope ∼ 1 + spike–SW density).

Figure 3.

Epileptic spike and SW components shape the spectral slope. A, Left, Comparison of the observed epileptiform spiking activity (gray) and a simulated approximation comprising spike and SW components (green). (Right) 1/f^χ noise (gray; slope, −2) with simulated spike–SW complex (green). B, Power spectrum (left) and spectral slope (right) as a function of spike–SW density. Power spectra depict mean ± SEM per density level. Slope data visualized using probability density functions; box plots represent median, first and third quartiles, and extreme values. Each point represents an iteration of the density level (N_iterations = 100). P value reported from linear regression. Note how increased spike–SW density affects the power spectrum in both low (2–3.7 Hz) and high (32–83 Hz) frequency ranges, hence flattening the spectral slope. C, Effects of spike density (left) on power spectra and the spectral slope as compared with 1/f^χ noise with a spectral slope of −2 (gray). Middle, Spike density impacts power estimates across all frequencies, but effects were most pronounced for higher frequencies (∼10–80 Hz). Right, Increasing spike density flattens the spectral slope. P value reported from linear regression. Data visualized using the same conventions as in panel B. D, SW density (left) solely affects power in the lower-frequency range (<4 Hz; middle), resulting in a steeper spectral slope (right) as compared with 1/f^χ noise with a spectral slope of −2 (gray). P value reported from linear regression. Data visualized using the same conventions as in panels B and C. E, Spectral slope as a function of spike and SW densities. P values are reported from multiple linear regression. Each cell represents the mean spectral slope across 100 iterations.

Simulation of spike and SW components

To disentangle the effects of epileptiform waveform shapes on spectral estimates, we separately repeated the analyses for simulated epileptic spiking (density 0–10 Hz in 1 Hz steps; Fig. 3C) and SW components (density 0–0.25 Hz in 0.025 Hz steps; Fig. 3D; fixed aperiodic mode, for models fitting a knee, see Text S5, Fig. S3). The effects of spike and slow-wave density on the spectral slope were examined utilizing linear regressions (slope ∼ 1 + spike density and slope ∼ 1 + SW density, respectively). The spectral slope was determined as a function of both spike and SW component density (Fig. 3E). The joint effect of spike and SW component density on the spectral slope was examined by a multiple linear regression (slope ∼ 1 + spike density + SW density).

Results

We investigated whether aperiodic activity delineates seizure activity, irrespective of the precise waveform shape, by recording intracranial electroencephalography (iEEG) during ictal and nonictal (RS, preictal, and postictal) conditions from patients diagnosed with pharmacoresistant epilepsy (N = 18, 11 female; M_age = 32.78 ± 3.71 years, mean ± SEM). A total of 31 seizure recordings were included in the analyses, which were selected for their waveform shape diversity to maximize the seizure pattern variability.

Aperiodic activity marks seizure activity

First, we compared the spectral decompositions of ictal and nonictal conditions to test if aperiodic activity delineates seizure activity. The model provided accurate fits [R² = 0.9743 ± 0.0019; mean ± standard error of the mean (SEM); range, 0.9435–0.9863; MSE, 0.0384 ± 0.0025; mean per seizure computed across all segments and conditions] which included reliable estimates for the ictal condition (R² = 0.9207 ± 0.0116; model error, 0.0907 ± 0.0094; mean per seizure computed across all ictal segments). Given the high goodness-of-fit, no fits were excluded from analyses. Notably, spectral parametrization was significantly more reliable when considering the full spectrum (1–120 Hz) as compared with narrower frequency ranges (Text S1; Fig. S1A). During seizures, the power spectrum was modulated across all frequency bands (Fig. 1D). Spectral slope estimates, i.e., the steepness of the power spectrum, differed significantly between conditions (Fig. 1E; p < 0.0001, LRT). The spectral slope was flatter during the seizure as compared with the RS condition [p = 0.0036, estimate of regression coefficient β = −0.23; 95% confidence interval (CI₉₅) = [−0.38, −0.08]; t₍₁₁₉₎ = −2.97, LME], the preictal condition (p = 0.0005; β = −0.28; CI₉₅ = [−0.43, −0.12]; t₍₁₁₉₎ = −3.56, LME), and the postictal condition (p < 0.0001; β = −0.48; CI₉₅ = [−0.63, −0.32]; t₍₁₁₉₎ = −6.08, LME). Similarly, the spectral offset marked seizure activity (Text S2; Fig. S1C). These results were not sensitive to segmentation parameters (i.e., the minimum number of segments and percentage overlap between segments; Text S4). Note that aperiodic estimates delineated the ictal condition regardless of the selected model (fixed or knee aperiodic model), although the direction of the effects changed when including a spectral knee parameter (Text S3, Fig. S2). In sum, these results show that the broadband spectral slope and offset differentiate the ictal from nonictal conditions, expressed as a flattening of the spectral slope.

Spectral slope changes with seizure patterns

Ictal patterns are variable between (Fig. 2A) and within seizures (Fig. 2B; Perucca et al., 2014; Abdallah et al., 2024). Similarly, ictal slope estimates varied across seizures (Fig. 1E), as well as within seizures, the latter being indicated by broad confidence intervals (Fig. S1B). To determine the impact of different waveform shapes on the spectral slope, we epoched the ictal data into 3 s segments and examined segment-specific variations in the spectral slope, as well as their link to density and waveform shape of various ictal patterns. Specifically, we examined whether the slope changes with density of epileptic spikes and spectral power in the SW/delta band. We observed a flattening of the spectral slope as a function of epileptic spike density (Fig. 2C; p_LRT < 0.0001, LRT; p_LME < 0.0001; β = 0.02; CI₉₅ = [0.02, 0.02]; t_(11,040) = 55.03, LME). Moreover, the spectral slope steepened as a function of delta band power (Fig. 2D; p_LRT < 0.0001, LRT; p_LME < 0.0001; β = −0.09; CI₉₅ = [−0.10, −0.09]; t_(11,040) = −37.66, LME). Similarly, the spectral offset changed with underlying waveform features (Text S2; Fig. S1D). These results were not sensitive to segmentation parameters (i.e., the minimum number of segments and percentage overlap between segments; Text S4). Inclusion of a knee parameter in the aperiodic model supported the central finding that aperiodic activity reflects underlying waveform features. Notably, while delta power remained associated with the spectral slope, epileptic spiking activity was instead captured by the knee frequency (Text S3, Fig. S2). To test how these observations relate to clinically established seizure patterns, we examined the spectral slope difference between LVFA and LFPS patterns. We found a flatter slope for LVFA patterns (characterized by high spike density) as compared with LFPS patterns (hallmarked by slow-wave activity; Fig. 2E; p_LRT < 0.0001, LRT; p_LME < 0.0001; β = 0.13; [CI₉₅] = [0.10, 0.16]; t_(4,476) = 7.76, LME).

Together, these findings reveal that the spectral slope tracks the density of underlying waveform features. A flattening of the slope reflects increased spike density, while a steepening of the spectral slope is related to enhanced delta power and may thus index increased SW activity.

Pathologic waveform shapes impact the spectral slope

Having established that the spectral slope reflects the density of epileptic spikes and SW activity, we investigated whether this relationship may be explained by the Fourier transform of these pathologic waveform shapes. We simulated epileptic spiking activity by combining a short, high-amplitude Gaussian function (reflecting an epileptic spike) with a longer, low-amplitude Gaussian function (Fig. 3A, left; simulating low-frequency SW activity). We added the simulated spike–SW complex to 1/f^χ noise with a spectral slope of −2 (Fig. 3A, right) and examined the spectral slope as a function of spike–SW complex density (ranging from 0 to 0.25 Hz; Fig. 3B). Increasing spike–SW complex density resulted in increased power for both low (<4 Hz; Fig. 3B, left; inset in top right) and high frequencies (>30 Hz; Fig. 3B, left; inset in bottom left). Notably, the spectral slope flattened as a function of spike–SW complex density [Fig. 3B, right; p < 0.0001; slope 2.37 ± 0.04 (mean ± SEM); t_(1,098) = 59.11; R² = 0.76; linear regression].

To disentangle the specific effects of spike and SW components on the aperiodic activity, we independently manipulated spike and SW density. Increasing spike density (Fig. 3C, left; ranging from 0 to 10 Hz) resulted in a broadband power increase that was most prominent in high-frequency bands (Fig. 3C, middle), resulting in a flattening of the spectral slope [Fig. 3C, right; p < 0.0001; slope 0.09 ± 0.00 (mean ± SEM); t_(1,098) = 29.40; R² = 0.44; linear regression]. Interestingly, simulated spiking activity induced a spectral knee whose frequency was modulated by spike density (Text S5, Fig. S3). In turn, increasing SW density (ranging from 0 to 0.25 Hz; Fig. 3D, left) led to power increases in lower frequencies (Fig. 3D, middle) and steepening of the spectral slope [Fig. 3D, right; p < 0.0001; slope −0.39 ± 0.01 (mean ± SEM); t_(1,098) = −21.05; R² = 0.29; linear regression]. Independent comodulation of both spike and SW density revealed the flattest spectral slope for high spike density and low SW density and the steepest spectral slope for low spike density and high SW density [Fig. 3E; p_spike < 0.0001; slope_spike 0.09 ± 0.00 (mean ± SEM); t_{(12,097, spike)} = 100.66; p_slow-wave < 0.0001; slope_slow-wave −0.31 ± 0.04 (mean ± SEM), t_{(12,097, slow-wave)} = −8.42; R² = 0.46; multiple linear regression]. Collectively, this set of findings demonstrates that the density and waveform shapes of pathologic activity modulate the spectral slope. Epileptic spikes result in high-frequency power increases, thus flattening the spectral slope, whereas SW enhance delta power and thereby steepen the spectral slope.

Discussion

The present results demonstrate that aperiodic activity tracks epileptic dynamics in the electrophysiological signal. The morphology and density of epileptic activity systematically distort spectral and associated aperiodic estimates: epileptic spikes flatten the spectral slope, while SW activity steepens it. These findings highlight the importance of scrutinizing waveform morphology when interpreting aperiodic activity in the context of pathological brain states.

Aperiodic activity delineates seizure activity

Aperiodic activity might be a promising biomarker of the population E–I ratio (Gao et al., 2017). As epileptic seizures constitute states of aberrant cortical excitability (McCormick and Contreras, 2001; Badawy et al., 2009a,b), the spectral slope may provide novel diagnostic insights into underlying E–I dynamics. The present study replicates prior findings that the spectral slope is modulated during seizures. However, while previous studies reported a steepening of the spectral slope during ictal activity (van Heumen et al., 2021; A. I. Yang et al., 2023; Kopf et al., 2024; Liao et al., 2024), we observed a flattening. This discrepancy may be attributed to differences in the selected frequency range and model parameters (cf. Ameen et al., 2024), as well as to the underlying electrophysiological waveforms, which can vary substantially within and across seizures (Perucca et al., 2014; Abdallah et al., 2024).

Epileptic activity typically comprises spiking and SW activity, and their densities define distinct seizure patterns. Our results show that epileptic spiking activity flattens the spectral slope, while SW activity steepens it. This clear dissociation was corroborated by simulations demonstrating how the Fourier transform of these waveforms systematically alters spectral and associated aperiodic estimates (Fig. 3B–D). In sum, these findings demonstrate that aperiodic modulations reflect the spectral decomposition of the underlying waveform shapes and may not directly index pathologic aberrations in the underlying E–I ratio.

Does the spectral slope track underlying E–I ratio in epileptic seizures?

The present study offers a more parsimonious explanation why the spectral slope tracks ictal activity, namely, its sensitivity to waveform features. This consideration raises the question whether the occurrence of distinct waveform shapes per se signal altered E–I dynamics. Optogenetic modulation and in vitro physiology in humans and animal models demonstrated that GABAergic (inhibitory) interneurons are involved in the generation of LVFA seizure patterns (de Curtis and Avoli, 2016; Shiri et al., 2016). In contrast, glutamatergic (excitatory) pyramidal cells are thought to be involved in the generation of LFPS seizure patterns (Shiri et al., 2016). This dissociation between the primarily involved cell type and resulting seizure patterns is reflected in the spectral decomposition, albeit in opposite direction of the model prediction: the spectral slope was flatter for LVFA (I↑) as compared with LFPS patterns (E↑). This set of findings challenges the assumption that a flattening of the spectral slope directly indexes increases in E–I ratio.

The underlying neural mechanisms giving rise to the 1/f decay function of electrophysiological power spectra are not fully understood (Buzsáki et al., 2012). Several studies have highlighted factors beyond the E–I balance interpretation that shape background 1/f aperiodic activity (Brake et al., 2024). Theoretical work has shown that the superimposition of damped oscillators may also underlie the characteristic 1/f shape of aperiodic activity (Milotti, 2002), and empirical evidence has demonstrated this association between aperiodic and periodic dynamics for the alpha frequency range (Muthukumaraswamy and Liley, 2018; Evertz et al., 2022). Our findings contribute to this body of work by demonstrating that waveform features are an additional, critical factor shaping aperiodic activity.

Implications for studying aperiodic activity as a clinically relevant biomarker

Aperiodic activity has been of great interest in recent years (Donoghue, 2024) and is commonly regarded as a noninvasive biomarker for various cognitive and pathologic neural states including epilepsy (van Heumen et al., 2021; Jiang et al., 2022; Kundu et al., 2023; Liao et al., 2024) but also stroke (Johnston et al., 2023; Maschke et al., 2023), malignant neoplasias (Numan et al., 2021), traumatic brain injury (Maschke et al., 2023), dementia (Rosenblum et al., 2023), or Parkinson's disease (Darmani et al., 2023; Rosenblum et al., 2023; Bush et al., 2024). Our results show that, in the case of epilepsy, utilization of aperiodic activity as a surrogate marker of the underlying E–I ratio is biased by the pathological waveforms. Similar considerations might apply to other pathological waveforms, such as interictal epileptic activity (Kundu et al., 2023; Kopf et al., 2024). Hence, it is best practice to remove epileptic transients to mitigate their impact on aperiodic estimates (Kopf et al., 2024). Similarly, it seems necessary to attenuate prominent waveform shape artifacts when considering the spectral slope as a biomarker in the context of other neurological diseases hallmarked by pathologic waveforms. For example, focal lesions in stroke (Jordan, 2004; Ramanathan et al., 2018; Galovic et al., 2021; Raposo et al., 2023) or brain tumors (Kennedy and Schuele, 2013; Selvam and Devi, 2015) are delineated by high-amplitude, low-frequency EEG activity. Other diseases, such as autoimmune-mediated encephalitis, are characterized by prominent delta brush patterns (Gillinder et al., 2019), while traumatic brain injuries might give rise to diffuse EEG slowing (Wallace et al., 2001; Ronne-Engstrom and Winkler, 2006). A prominent example of pathological sharp waveform shapes is the aberrant central beta rhythm (∼13–35 Hz) in patients suffering from Parkinson's disease, where increased waveform sharpness in the beta range (Cole et al., 2017; Geraedts et al., 2018) might bias aperiodic estimates (Wiest et al., 2023; Bush et al., 2024). Lastly, altered aperiodic activity has also been described in Alzheimer's and dementia (Al-Qazzaz et al., 2014; Kamondi et al., 2024), where occult epileptic activity might be present (Lam et al., 2017) that could bias aperiodic estimates. The present study emphasizes the necessity to evaluate waveform shapes in the time domain to correctly interpret changes in aperiodic activity, specifically in the context of pathological brain states.

Limitations

All recordings were conducted using intracranial electrodes in pharmacoresistant patients with epilepsy for diagnostic purposes, and coverage consequently favored (sub)cortical structures with a high likelihood of pathological activity. As we lack a true baseline condition without any pathologic activity for these structures, we compared the ictal condition to various nonictal conditions (i.e., RS, preictal, and postictal) to mitigate this limitation. Second, different seizure types and pathologies may be associated with distinct neural mechanisms that are reflected in the aperiodic activity. Including a broader range of pathologies and seizure types might benefit the identification of pathology-specific alterations and ensured generalizability of our findings, which was not feasible given the limited patient sample. Third, although manual classification of seizures and seizure patterns constitutes the clinical gold standard, it introduces subjectivity (Abdallah et al., 2024). In the present study, seizure patterns were defined visually as well as by quantifying spike density and SW activity. Fourth, comparing absolute spectral slope estimates may introduce biases, since the absolute value of the spectral slope is shaped by cellular microstructure (Gao et al., 2020), cortical depth (Halgren et al., 2021), and age (Voytek et al., 2015). We adopted a within-subject design to mitigate this limitation. Fifth, the assessment of aperiodic activity may be impeded during seizures, as neural activity is highly pathological. Power spectra may therefore exhibit abnormal oscillatory peaks and spectral knees (Gerster et al., 2022). In the context of epileptic seizures, the distinction between periodic and aperiodic activity is often unclear, and the identification of spectral knees remains ambiguous. However, it is critical to highlight that pathological rhythmicity and synchronization is a defining feature of epileptic EEG activity. Sixth, supplementary analyses revealed that the selection of aperiodic model parameters is crucial, as aperiodic estimates depend on the specific model (Text S3, Fig. S2). Future work should specifically address the presence of spectral knees during epileptic seizures. Critically, altered aperiodic activity that resulted from pathologic waveforms was observed irrespective of the selected model; hence, our results provide qualitatively similar results irrespective of the precise aperiodic parameters.

Conclusion

Collectively, these findings demonstrate that the spectral slope tracks ictal activity. In contrast to the notion that aperiodic activity reflects the E–I ratio, the present results provide a more parsimonious explanation and illustrate that waveform morphology systematically biases aperiodic estimates.