Closed-loop modulation of sleep in children undergoing intracranial recordings

Summary

Children with epilepsy suffer a vicious cycle in which disturbed sleep heightens seizure susceptibility, while seizures further disrupt sleep quality, particularly impairing the slow-wave sleep (SWS) critical for cognitive, immune, and metabolic function. We present a phase-targeted auditory stimulation (PTAS) system that delivers stimuli timed to endogenous slow oscillations. In 27 children undergoing epilepsy monitoring with simultaneous scalp, intracranial, and thalamic recordings, PTAS significantly enhances SWS power, with maximal effects on thalamic, frontal, and auditory regions in a randomized cross-over protocol. Stimulation also suppresses interictal epileptiform discharges (from 0.4 to <0.1 spikes/min) and improves cognitive performance on a response inhibition task (from 76% to 95% accuracy). These results provide direct intracranial evidence that closed-loop auditory stimulation modulates sleep architecture, suppresses pathological activity, and enhances cognition. PTAS represents a physiologically informed, non-invasive approach for addressing both neurophysiological and cognitive comorbidities in pediatric epilepsy.

Graphical abstract

Highlights

• •Phase-targeted auditory stimulation enhances slow-wave sleep in the epileptic brain
• •Thalamocortical networks involved in sleep modulation were probed with intracranial electrodes
• •Nighttime stimulation reduced interictal discharges during sleep by 75%
• •Next-day response inhibition improves from 76% to 95% accuracy after overnight stimulation

Phase-targeted auditory stimulation during sleep enhances slow-wave activity in children with epilepsy, with direct intracranial evidence of large-scale thalamocortical network engagement. This non-invasive approach significantly reduces epileptiform discharges during sleep and improves next-day cognitive performance on response inhibition tasks.

Introduction

Sleep plays a fundamental role in numerous aspects of human health and well-being. It is essential for physical, metabolic, and cardiovascular health; cognitive function; emotional well-being; growth; development; safety; and longevity.^{1,2,3,4,5,6,7} Chronic sleep disruption and disorders significantly increase the risk of hypertension, obesity, diabetes, and numerous other adverse health outcomes.^2,3,4

Efforts to optimize sleep have historically focused on addressing specific sleep disorders or prioritizing sleep hygiene.⁸ Recent advancements have attempted to directly enhance the restorative properties of sleep. Closed-loop phase-targeted auditory stimulation (PTAS) presents a promising avenue for directly modulating slow-wave sleep,^9,10 a phase of sleep that is critical for memory consolidation,¹¹ metabolic regulation,¹² and immune function.¹³ PTAS works by detecting slow-wave oscillations in real-time and then delivering auditory stimuli at specific phases of the oscillation.¹⁴ This targeted stimulation can amplify slow-wave activity, leading to potential improvements in cognition,⁹ metabolic health,¹⁵ and other sleep-dependent functions.¹⁶ Notably, PTAS offers a non-invasive and potentially customizable intervention that could target a core restorative component of sleep.¹⁴ This technology holds promise for improving both sleep quality and waking function in a variety of populations, including those with sleep disorders and those with sleep-related impairments to cognitive performance.

Children with epilepsy represent a population in need of sleep-targeted interventions. Adequate sleep is essential for seizure control,¹⁷ with some types of epilepsy responding favorably to improved sleep alone.¹⁸ Furthermore, sleep quality significantly impacts cognitive function^19,20,21,22—a critical aspect of quality of life for individuals with epilepsy. Moreover, intracranial recordings in this population present an opportunity to elucidate the source-based neural mechanisms underlying PTAS and sleep processes.²³

In this current work, we leveraged simultaneous scalp and intracranial electroencephalogram (EEG) recordings from cortical and subcortical regions, including the thalamus, in children with epilepsy to characterize the scalp and intracranial response to auditory stimulation during sleep, which we term the nocturnal auditory evoked potential (nAEP). We then systematically characterized the differential effects of evoking the nAEP at varying phases of the sleep slow oscillation (SO) and identified a phase target that robustly enhances sleep SOs. Using a closed-loop device that we developed specifically to address difficulties in delivering PTAS to persons with epilepsy, we explored the effects of PTAS on interictal epileptiform discharges (IEDs) and cognition by delivering up-phase targeted stimulation in a multi-day randomized, counterbalanced experiment in the epilepsy monitoring unit. This work addresses critical gaps in our understanding of the physiological mechanisms and pathways for auditory stimulation during sleep and the effects of PTAS in children with epilepsy.

Results

We developed a system (Figure 1A) to deliver PTAS in a pediatric epilepsy monitoring unit, an inherently dynamic and complex environment. Our system utilized a wavelet convolution approach (Figure 1B) to phase tracking to minimize the confounding effects of epileptiform events on slow-wave sleep detection, while optimizing closed-loop feedback time for improving phase targeting. Our approach outperformed commonly utilized phase-targeting methods, with pronounced benefits in the epileptic EEG.²⁴

Figure 1.

A system for PTAS in children with epilepsy using the TWave algorithm, and the experimental design overview

(A and B) Schematic overview of our closed-loop phase-targeted auditory stimulation system, consisting of a wireless EEG amplifier, bedside computer and processing software, bedside speaker to deliver auditory stimulation, and a trigger interface to the clinical EEG recording system for time-locking. ADC, analog/digital converter; μIC, microcontroller; LiPoly CHG, lithium polymer battery charger; DC-DC, DC-to-DC voltage regulator; SPI, Serial Peripheral Interface bus; BLE, Bluetooth Low Energy; UART, universal asynchronous receiver/transmitter; GPIO, general-purpose input/output; Spk, speaker. (B) Schematic overview of the TWave algorithm: incoming live EEG is convolved with a family of wavelets to estimate the frequency, amplitude, and phase and used to extrapolate the timing of the target phase in the upcoming oscillatory cycle.

(C) Overview of experimental blocks in the Discovery and Experimental cohorts. The Discovery cohort received random stimulation for one or more nights. The Experimental cohort received randomly counterbalanced blocks of PTAS (stimulation) and muted (withheld stimulation), followed by daily computer-based neurocognitive assessments with the PVT and Go/NoGo tasks (GnG).

Study design and cohorts

To demonstrate the feasibility of auditory stimulation and to optimize stimulation parameters, we first characterized the effect of auditory stimulation delivered at varying phases of sleep slow waves on sleep electrophysiology in 15 children undergoing randomly timed stimulation (the Discovery cohort; 53% male, ages 8–16 years). Subsequently, we recruited a separate cohort of 12 children who received PTAS delivered in a randomized counter-balanced block-design experiment (the Experimental cohort; 42% male, ages 9–18 years). Full demographics of both cohorts are provided in Tables S1 and S2.

In all children, we recorded continuous EEG with simultaneous stereoelectroencephalography (sEEG) using implanted electrodes (Figure 2D; mean of 140 electrode contacts per participant, range 70–196) that sample brain regions with high anatomical precision and millisecond timescales. All children exhibited normative gross sleep architecture²⁵ (Figure 2A), with an average of 53%, 24.5%, and 18.5% of the night spent in N2, N3, and rapid eye movement (REM) sleep, respectively, and average time asleep of 8 h (Figure 2C). Stimulation did not significantly alter time within each sleep stage in the Discovery and Experimental cohorts. Mean spike-wave index,²⁶ the proportion of time during sleep where IEDs were observed, was 9.5% and did not differ significantly between cohorts (Figure 1B). The overall topology of IEDs across both groups is shown in Figure S3.

Figure 2.

Aggregate sleep and spike measures and electrode coverage map

(A) Sleep architecture across all recruited participants (percentage of time spent is represented as mean ± SEM).

(B) Spike index of the Discovery and Experimental cohorts (spike index boxplot represents the median, interquartile range [IQR], and ±1.5 IQR).

(C) Aggregate sleep likelihood at time of day demonstrates normative mean sleep onset and wake times (mean percentage ±SEM).

(D) Aggregate implanted electrode locations from all subjects (N = 27) registered in Montreal Neurological Institute standard space.

The nocturnal auditory evoked potential

We first aimed to characterize the intracranial sources and evolution of the nAEP, leveraging our tandem scalp and intracranial recordings. From the Discovery cohort, who received random stimulation, we identified the nAEP by computing the averaged stimulus-locked timeseries with auditory stimulation in 9 of 15 subjects in whom the nAEP was successfully evoked. In the remainder of the children in whom we could not measure the nAEP, we discovered that the audio headband shifted or loosened overnight. Children in the Discovery cohort received a mean of 284 ± 17 stimulations per night. The scalp topology of the nAEP in children with epilepsy showed a predominant central concentration and sharp waveform (Figure 3, top left), comparable to previous reports in healthy children.²⁷ The intracranial evolution of the nAEP is shown in Video S1. Regional time-domain waveforms summarizing the intracranial signal are also shown in Figure 3 (bottom left), averaged over lobar anatomy.

Figure 3.

The nocturnal auditory evoked potential

Grand average scalp EEG waveforms and time-resolved scalp topology of the nAEP (left). Topology and timeseries of four principal components of the time-locked nAEP from intracranial recordings, averaged within Automated Anatomical Labeling, v.3 (AALv3) regions.

To characterize group-level spatial patterns in the intracranial nAEP, we applied principal-component analysis directly to the time-domain data matrix (samples × regions), identifying four components based upon the elbow criterion (see Figure S1) that together account for 71% of the observed variance. This decomposition yielded time-invariant spatial loadings (rendered as brain maps) and corresponding time courses for each component (Figure 3, top right), capturing co-varying patterns of activity across regions, while reducing dimensionality.

The first principal component (C1) showed a large negative-going wave peaking at 450 ms in its time course, with a diffuse fronto-central-parietal medial topology in its spatial loadings with matching but opposing positive potentials observed in the inferior frontal and temporal regions. The second principal component (C2) consisted of a positive-going peak at 300 ms, followed by a negative-going peak at 660 ms in the parietal, inferior frontal, and medial frontal regions. Principal component 3 was largely similar to C2, with frontal involvement and an earlier and sharper peak at 297 ms followed by a negative peak at 450 ms. Principal component 4 exhibited the earliest peak at 200 ms, with a lower amplitude and seemingly oscillatory morphology in the inferior frontal, ventral anterior cingulate, inferior parietal, and posterior cingulate regions (Figure 3, right).

Up-phase stimulation selectively potentiates the sleep slow oscillation

PTAS relies on the interaction between auditory stimulation and specific phases of an endogenous oscillation, typically the sleep SO. We, therefore, sought to characterize the effect of sleep SO phase at the time of auditory stimulation on the resultant SO power. We observed that auditory stimulation at the rising and peak phases resulted in higher post-stimulation slow oscillatory power. Following delivery of PTAS, post-stimulation SO power in both early (500–2,000 ms) and late time windows (2,000–5,000 ms) was significantly associated with the phase of the SO at the time of stimulation (Figure 4A), as characterized by Rayleigh’s test^28,29 of circular uniformity (Rayleigh’s test Z score; Rz = 4.4 and Rz = 35.7, respectively), with higher Z score values indicating deviation from a uniform distribution. Post-stimulation amplitude was maximized at mean stimulation phases of 6.15 radians for the early window and 5.28 radians for the late window.

Figure 4.

Up-phase PTAS preferentially potentiates sleep SOs

(A) Post-stimulation SW power (0.5–2.0 Hz; radial axis) as a function of the phase (angular axis) of the SW at the time of stimulation shows power is higher during the rising and peak phases. The Rayleigh Z (Rz) value for both distributions indicates a significant non-uniformity of the amplitude-phase distribution.

(B) Mean stimulus-locked timeseries during the up- and down-phases of stimulation shows significantly increased post-stimulation activity after up-phase stimulation relative to down-phase stimulation (shading represents 95% bootstrap confidence interval [CI]).

(C) Up-phase targeted auditory stimulation increases SO power in the mid and inferior frontal, central, cingulate, and temporal regions.

To investigate the electrophysiology underlying differences in oscillatory power, we then analyzed the effect of phase on the differential temporal and topological evolution of the sleep SO. Based on the observed phase to SO power relationship (Figure 4A), the randomly delivered stimuli were dichotomized into up-phase (7π/6 to π/6) or the down-phase bins. First, we observed that the phase-aligned evoked potentials (Figure 4B. bottom) of the Fpz-Cz montage were significantly higher (q < 0.05) with up-phase stimulation. Second, up-phase stimulation was associated with significantly higher intracranial SO power across a widespread network spanning the mid and inferior frontal, cingulate, pre- and post-central, parietal, and temporal gyri (Figure 4C, q < 0.05). Third, we observed that up-phase stimulation promoted potentiation of the stereotypical SO frontocentral topology (Figure 4B, see 0.5–1.0 s). Down-phase stimulation was associated with the fractioning of this topology and diminished SO facilitation. After down-phase stimulation, the frontocentral topology previously observed in the baseline period was no longer present (Figure 4B, top at +0.5 s), the power in the SO frequency band was lower (Figure 4A), and there was no post-stimulation increase in scalp potentials (Figure 4B, bottom).

Up-PTAS increases slow-wave amplitude in children with epilepsy

Given that auditory stimulation targeted to the up-phase (“up-PTAS”) potentiated the sleep SO, we then sought to characterize the effect of up-PTAS overnight on sleep electrophysiology and epilepsy. To this end, we designed a randomized cross-over study design (Figure 1C). Children in the Experimental cohort were randomly allocated and received counterbalanced 4-h blocks of up-PTAS or withheld stimulation (the muted condition) over at least 2 nights’ duration with a target phase of Φ = 0 in the Fpz-Cz bipolar channel. In the Experimental cohort, children received on average 319 ± 42 stimulations and 330 ± 36 withheld stimulation triggers.

Compared to the muted condition, up-PTAS resulted in significant increase in slow-wave amplitude in the Fpz-Cz bipolar derivation up to 1.8 s post-stimulation (Figure 5A, q < 0.05). In a time-frequency-resolved analysis (Figure 5B), we found up-PTAS significantly increased SO power up to 3 s after delivery of the auditory stimulus.

Figure 5.

Up-PTAS potentiates the sleep SO

Mean stimulation-locked timeseries from the Experimental cohort compared between up-PTAS and muted conditions shows significant increases in SO activity in the Fpz-Cz montage (A, evoked; B, time-and-frequency-resolved decomposition), where shading represents 95% bootstrap CI. Scalp timeseries is shown for the muted (C) and up-PTAS (D) conditions, alongside the difference (F). At key time points selected by MNE’s peak finder, up-PTAS results in significantly increased SO activity across a wide set of scalp EEG sensors (E, q < 0.01, paired t test). Similarly in our intracranial recordings, the frontocentral and temporal regions and thalamus show increased slow oscillatory power with PTAS (G). In the thalamus, PTAS results in a strong evoked auditory response, despite the lack of pre-stimulation baseline SOs in the thalamus (H), where shading represents 95% bootstrap CI. Sleep architecture did not differ significantly between PTAS and muted conditions (I); boxplots representing the median, IQR, and ±1.5 IQR.

To identify the cortical regions affected by up-PTAS, we computed the whole scalp evoked response in both PTAS and muted conditions (Figures 5C–5F). In the up-PTAS condition, we observed a frontocentral-dominant evoked response that persisted synchronously for several additional cycles. The sleep SO in the baseline interval of both conditions did not differ significantly between groups. At time points identified by MNE-Python’s peak finder function,³⁰ up-PTAS resulted in significant increases in frontocentral EEG activity (q < 0.01), predominantly in frontal and central midline regions.

Using a linear mixed effects model accounting for the random effect of subject and night, we identified the intracranial correlates of up-PTAS and observed increased SO power localized to the anterior cingulate, medial superior frontal, Rolandic, and medial orbitofrontal cortices, and uniquely, the thalamus (Figure 5G). Among the 8 children with thalamic electrodes, a significant evoked response with up-PTAS was also observed in the thalamus (Figure 5H), but in contrast to our scalp recordings, we did not observe a corresponding baseline (i.e. detection) SO in the thalamus.

Up-phase PTAS reduces epileptiform discharges during sleep

To explore the effects of up-PTAS on epileptiform activity, we marked IEDs using a previously established automated spike detector.³¹ We then compared the rate of IEDs between up-PTAS and muted conditions. From the randomly counterbalanced Experimental cohort, we found that the rate of IEDs was significantly lower during blocks of up-PTAS, relative to blocks of withheld stimulation (Figure 6C) during both N2 and N3 sleep (p < 0.001). A similar pattern was observed in subgroup analyses considering only IEDs recorded over frontal, temporal, parietal, and occipital scalp EEG electrodes (Figure 6E). To assess whether reduction in IEDs occurs in the immediate temporal window of up-PTAS, we performed a time-resolved analysis of the likelihood of IEDs relative to stimulation and found that up-PTAS resulted in a significant decrease in spike likelihood during the 5-s period post-stimulation (Figure 6D).

Figure 6.

Up-PTAS reduces spikes and improves response inhibition

Children underwent daily neurocognitive testing with the Go/NoGo and PVT computer-based tasks (A). After PTAS, we observed significant increases in neurocognitive performance in response inhibition indexed by accuracy on the Go/NoGo task (∗p = 0.002 linear mixed effects [LME] in B). Alertness as indexed by the vigilance condition and PVT was non-significant. We observed significant decreases to spike rates overall with PTAS (N2: p < 0.001 LME, N3: p = 0.007 LME; C) and significant decreases in spike rate immediate post-stimulation (D; ∗p < 0.05 LME). The decline in spike rates was largely consistent across regions (E; ∗q < 0.05 LME). Boxplots represent the median, IQR, and ±1.5 IQR. Bar plots show the mean ± SEM. The dot plot (D) shows the estimate and 95% CI estimated by the LME.

Up-phase PTAS improves inhibitory control

Given that slow-wave sleep and PTAS have been shown to improve neurocognitive outcomes in adult populations,¹⁰ we sought to evaluate the neurocognitive sequelae of up-PTAS on children with epilepsy. We used the well-established Go/NoGo task³² (Figure 5J) to index response inhibition, a sleep-sensitive cognitive domain that is commonly impaired in children with epilepsy.^33,34 After nights of up-PTAS, we found significantly improved accuracy (p = 0.002) in response inhibition (mean of 95%) compared to after nights with no stimulation delivered (mean of 76%; Figure 6F).

To ensure that the findings of greater response inhibition were not attributable to states of vigilance during the task, we also measured performance on the vigilance condition of the Go/NoGo task and an independent psychomotor vigilance task (PVT).³⁵ We found no difference in accuracy in the Go/NoGo vigilance condition (p = 0.50) and no difference in the PVT reaction time (p = 0.74) or accuracy (p = 0.45) between nights with and without PTAS.

Discussion

Sleep is fundamental to health, and children with epilepsy are particularly vulnerable to the effects of sleep fragmentation on seizure control and cognition. Modulation of sleep, namely, sleep SOs has recently emerged as a non-invasive strategy to enhance the restorative properties of sleep. Although closed-loop auditory stimulation during sleep is an area of active research, the intracranial effects of neuromodulation and the effects of PTAS in children with epilepsy are not well described.^36,37 In the current report, we introduce a system for tracking and estimating oscillatory phase in the EEGs of children with epilepsy. In the Discovery cohort with simultaneous scalp and intracranial EEG, we show that up-phase PTAS results in greater potentiation of sleep SOs compared to down-phase PTAS both in scalp recordings and within a thalamocortical network engaging auditory, frontal, central, and cingulate circuitry. The findings of thalamic involvement were uniquely confirmed with thalamic-targeted electrodes. We then designed a randomized, counterbalanced experiment administering blocks of up-PTAS and muted (withheld) stimulation to our Experimental cohort. We report enhancement of SOs, coincident with reductions in IEDs and improvement in the sleep-related cognitive domain of inhibitory control. The current findings present detailed intracranial mechanistic insights into PTAS and avenues for sleep modulation in epilepsy.

Intracranial networks of PTAS during sleep

Leveraging our unique dataset of tandem scalp, intracranial, and thalamic electrodes, we characterized the topography of the electroencephalographic response to phase-locked auditory stimulation during sleep, which we term the nAEP.

The nAEP consists of early peaks in temporal and parietal association cortices and auditory cortices (Figure 3, 100–200 ms, Components 2 & 3) and exhibited latency and topography consistent with auditory processing.^38,39 The sleeping brain is known to continuously receive and process sensory information, particularly auditory input, which increases alpha power in monitoring areas such as the prefrontal cortex (PFC) and anterior and mid cingulate cortices.⁴⁰ Similarly, we observed that the early auditory components were followed by activation in the prefrontal cortex and cingulate (Figure 3, Component 4). Similar transient sensory associative processing during sleep has been previously demonstrated. Error-related negativity responses⁴¹ have been observed during sleep with auditory oddball paradigms^42,43,44 and in response to unfamiliar voices.⁴⁵ The observed early intracranial components of the nAEP may, therefore, represent auditory and association “sentinel” activity⁴⁵ localized to the inferior and mid frontal, parietal, and posterior cingulate cortices.

Following these initial components, the nAEP includes a high-amplitude and widespread component resembling a k-complex (KC) in both waveform and topology^46,47 (Figure 3, Component 1). The KC has been proposed to serve as an instantaneous homeostatic response to maintain sleep in the face of non-threatening external stimuli.^48,49,50 KCs are thought to be variants of the sleep SO^51,52 as the KC is mechanistically, topologically, and functionally related to sleep slow waves^{27,49,51,27,53,54,55,56} and shares similar thalamocortical mechanisms.⁵⁷ KCs occur during and contribute to deepening of sleep,⁵⁸ especially preceding transition to N3.^59,60 In our data, the highest amplitude component of the nAEP is a widespread low-frequency wave propagating across the cortex with a peak latency of 450 ms. Therefore, the nAEP appears to reflect endogenous homeostatic sleep mechanisms, potentially deepening sleep in response to auditory stimulation.

The interaction between the nAEP and endogenous SO during N2 and N3 sleep leads to either potentiation or disruption of SO activity, dependent on the phase of the endogenous extant SO at the time of stimulation. Auditory stimulation during the positive-going phase of the Fpz-Cz waveform selectively potentiates SOs on scalp EEG, as well as within a widespread network involving frontal, central, parietal, cingulate, and temporal regions (Figure 4C). Given the shared mechanisms and shared roles of the KC and the SO, the interactions between the nAEP and SOs within intracranial brain topographies that we report provide evidence that PTAS potentiates sleep SOs through evoked KC-like activity.

We have long known that the thalamus and thalamocortical pathways are crucially involved in slow-wave sleep.^52,57,61,62 The role of the thalamus in mediating PTAS effects has been suspected indirectly through scalp EEG,⁶³ computational modeling,⁶⁴ and simultaneous EEG-fMRI.⁶⁵ By leveraging unique intracranial recordings within the central thalamus, we provide direct evidence of thalamic involvement in PTAS. We show that the increased SO power observed with up-PTAS is paired with a simultaneous increase in SO power in the thalamus (Figures 5B–5G and 5H) demonstrating that PTAS modulates thalamic SOs. Interestingly, up-PTAS evokes an SO in the thalamus despite the lack of a consistent antecedent SO in thalamic recordings and suggests that scalp SOs may not be universally expressed in the central thalamus. Taken together, our analyses of the intracranial nAEP provides evidence for the engagement of the brain’s endogenous homeostatic sleep protective response via the KC, through the auditory and association cortices, a phenomenon closely linked to the thalamocortical generators of sleep SOs.

Evoked response vs. entrainment of oscillations

Sleep SOs can originate from cortical (predominantly frontal) sources,^{54,66,67,68,69}, thalamic sources,^{70,71,72,73,74,75} or both,⁷⁶ and these interacting oscillatory mechanisms may respond to and modulate the observed effects of auditory stimulation during sleep.⁷⁷ It is not known whether the potentiation of the SO by PTAS represents merely constructive, additive interference between the nAEP and endogeneous SO or whether auditory stimulation may potentiate the oscillatory corticothalamocortical circuitry⁷⁸ underlying SOs.

In both our findings (Figure 4) and prior work in PTAS,^9,79 auditory stimulation targeted to the opposite phase reduces SO amplitude. We propose four related and interconnected mechanistic explanations. First, the simultaneous, overlapping, and anti-phase oscillations measured on scalp EEG additively interferes to cause a lower measured amplitude. Depending on the timing of auditory stimulation, the nAEP “up”-state may temporally coincide with the “down”-state of the ongoing endogenous SO. Conversely, when stimulation aligns with the endogenous rhythm, these oscillations can constructively interfere. Second, simultaneous and conflicting excitatory and inhibitory inputs at the thalamus and brainstem from the nAEP and ongoing SOs to the cortex could result in destructive inference and the observed disruption of the sleep SO topology (Figure 4). Third, auditory stimuli arriving at sensory nuclei in the brainstem and thalamus alter firing patterns in the thalamic reticular nuclei, important in the control and propagation of sleep SOs.^80,81 Fourth, the down-phase of a sleep SO may affect the sensory processing of the auditory stimulus^82,83 and subsequently impact the tendency of that auditory stimulus to evoke the nAEP. SOs consist of alternating up-states facilitating brief windows of coordinated neural processing followed by down-states broadly suppressing neuronal activation across the cortex coordinated by thalamocortical circuits.^52,84

Entrainment describes the phenomenon whereby cortical oscillatory activity synchronizes with periodic external stimuli^85,86,87 and may potentially be driven by auditory stimulation during non-REM (NREM) sleep.⁷⁷ Neural oscillators, including corticothalamocortical networks^74,78 involved in sleep SOs can be driven to oscillate at specific frequencies by external oscillatory input.^88,89 Although prior work has generally applied phase-targeting only for the first auditory pulse, while second and subsequent pulses in the train are delivered at fixed intervals,¹⁴ we sought to deliver stimulation entirely in-phase with the endogenous oscillator for both the initial and subsequent stimulations in a train. Rather than entraining the brain’s oscillatory network to an external periodic stimulus, we entrained our stimuli to the brain’s endogenous rhythms.

PTAS in children reduces IEDs

The bi-directional relationship between sleep and epileptic events including IEDs is well established. Poor sleep exacerbates the risk of seizures^90,91,92 and increases the propensity for IEDs.^92,93,94,95 Furthermore, IEDs are more frequently observed with increasing sleep depth^96,97,98,99 and these IEDs can fragment sleep,^97,100,101 leading to a pattern of high IED burden and lower sleep quality in many children with epilepsy.^102,103,104 As a result, affected children often suffer from impaired daytime cognitive function.^105,106 IEDs also disrupt the normative network maturation and cortical development that occurs during sleep,^31,107 with sleep being a particularly vulnerable interval when many processes critical to brain maturation occur.^108,109 There is a need for novel therapies that reduce IEDs and improve sleep in this population.

The precise mechanisms underlying the interaction between IEDs and sleep architecture are incompletely understood. Physiological transient events, including hippocampal sharp-wave ripples are known to couple with specific phases of the cortical SO.^110,111 These interactions subserve sleep-related cognition, including memory consolidation.¹¹² Interestingly, IEDs are also preferentially expressed during specific phases of the SO,^113,114 particularly the transition from up- to down-states, which provide brief temporal windows characterized by high inter-regional brain synchronization.¹¹³

Analogous to the windows of excitation and inhibition provided by SOs, there exists a quiescent window following IEDs whereby brain-wide networks (including within healthy tissue) and neural activity are inhibited.¹¹⁵ This has been attributed to a transition to a global cortical down-state¹¹⁶ and the resultant transient suppression of all synaptic and spiking activity in thalamocortical networks.¹¹⁵ This refractory period^{117,118,119,120} momentarily inhibits the occurrence of a subsequent IED. Given that the nAEP induces a similar thalamocortically mediated down-state, it may, therefore, initiate a similar refractory mechanism for spike suppression following PTAS.³⁶ In support of this view, thalamocortical SOs and associated KC responses are known to exhibit a decline with repeated stimulations, indicative of a refractory response.^89,121 The KC is also elicited more often when the interstimulus interval (ISI) is longer than 30 s and less frequently when the ISI is between 5 and 10 s.⁵⁰ Given that these two mechanisms utilize the same thalamocortical circuits, there may be competition between physiological (PTAS) and pathologic (IEDs) transients within these neural structures. The refractory period of the nAEP may prevent expression of IEDs during future SO cycles, which is supported by the reduction in IEDs we observed during the 5-s window following an auditory stimulation in NREM sleep (Figure 6).

We also observed an overall reduction in spike likelihood in both N2 and N3 sleep (Figure 6) during intervals wherein up-PTAS was administered, suggesting that thalamocortical refractoriness does not entirely account for the impact of up-PTAS on IEDs. In addition to thalamocortical refractoriness, we propose two further mechanisms through which up-PTAS may reduce IEDs. First, disrupted sleep architecture^{97,122,123,124} and microarousals lead to increased epileptic activity, whereas slow-wave and delta oscillations homeostatically regulate cortical excitation.^125,126 Up-PTAS may, therefore, potentiate these regulatory SOs. Second, up-PTAS may stabilize physiological thalamocortical synchrony, potentially through an entrainment effect or by supporting the expression of normative architectures, such as spindles. As IEDs arise from altered and dysregulated thalamocortical rhythms in anatomical,¹²⁷ computational,^128,129 and cellular models,¹³⁰ it is thought that the same circuitry that produces physiologic oscillatory activity such as sleep spindles is also the likely source of IEDs^102,131,132 in the presence of dysregulated synchrony within the thalamic network.⁸¹ Indeed, spindle rate has been shown to be reduced in the epileptogenic cortex proportional to IEDs^131,133 and previous studies employing auditory stimulation found that the frequency of sleep spindles was negatively correlated with IED rate.³⁶

Taken together, up-PTAS reduces IEDs by (1) inducing a refractory and broadly inhibitory down-state in thalamocortical circuits, (2) stabilizing NREM sleep and potentiating sleep SOs, and (3) predisposing thalamocortical mechanisms toward producing spindles instead of IEDs by stabilizing physiological thalamic regulatory mechanisms.

PTAS in children improves cognition

Sleep plays a crucial role in cortical maturation and neural network plasticity, especially during early childhood, and the relationship between slow-wave sleep and cognition is well-established.^11,134,135 Sleep-related cognitive-restorative processes^136,137 are disrupted in children with epilepsy.^{31,107,138,139}

Previous studies with healthy adults^9,15 report that PTAS has focused exclusively on sleep-related memory consolidation and identified improvements following PTAS even when it was administered during a nap. In the current work, we sought to expand these findings by studying sleep-related inhibition. Response inhibition, operationalized through the Go/NoGo task, is sensitive to the sleep quality of the previous night.^140,141,142 Deficits in response inhibition, which are prevalent in children with epilepsy,^33,143 are at least partially driven by fragmented sleep.^140,144 We report that up-PTAS significantly improves response inhibition during the subsequent day (Figure 6). Importantly, the effects we observed were not driven by changes in the level of arousal or vigilance, evidenced by comparable performance in both PVT and during the vigilance condition of Go/NoGo. Although we had also planned to characterize memory consolidation, we found that most children we were able to recruit were either unable or unwilling to complete daily paired-word association verbal memory tasks commonly used in prior PTAS research.¹⁰

These findings of improved next-day neurocognition may arise due to, first, the increase in SO power in cortical regions important to response inhibition and, second, the decrease in IEDs during up-PTAS. Children with epilepsy often suffer from impaired slow-wave sleep.^145,146 Sleep deprivation slows neural processing¹⁴² and reduces performance in response inhibition, indexed by the Go/NoGo and the stop signal task.^142,147 The improvement in Go/NoGo performance we observed in children with epilepsy after up-PTAS may be mediated in part by improved restoration during slow-wave sleep in these regions critical for response inhibition. Specifically, the inhibitory control is positively associated with activity in the PFC and anterior cingulate.^148,149,150 The amplitude of sleep SOs predict next-day task-related activity in the PFC¹⁵¹ and resulting performance in executive function assessments,^152,153,154 which may be mediated by the role of slow-wave sleep in cortical reorganization.¹⁵¹

Second, IEDs during sleep directly disrupt cognitive networks and sleep restorative processes critical to cognition. In children with self-limited epilepsy with centrotemporal spikes (SeLECTS or previously known as BECTS), the activation of IEDs and more generalized discharges during sleep was associated with increased daytime cognitive deficits,^155,156 and children with high spike rates during sleep, such as those with continuous spike-and-wave during sleep, experience significant cognitive impairment.^157,158 IEDs are thought to interfere with physiological sleep oscillations at the site of the IED and impair the local neuroplastic changes critical to learning and brain maturation,^132,159,160 and a reduction in IEDs correlates with subsequent recovery of cognitive function.¹³¹ The reduction in IEDs observed with up-PTAS may, therefore, improve cognition by reducing the deleterious impacts of IEDs on cognitive networks during sleep and sleep restorative processes.

In summary, we observed significantly higher SO power in the regions critical to mediating response inhibition (Figure 4) and a reduction in IEDs (Figures 6C and 6E) following up-PTAS. Improved SO-mediated restoration of the PFC and reduced IED-mediated disruption of sleep processes may underlie the observed significant increase in response inhibition the next day (Figure 6B). Understanding the mechanistic relationship between sleep-specific phenomena, epileptiform activity, and cognitive outcomes facilitates the development of more comprehensive treatment strategies in epilepsy.

Our work demonstrates the promising potential of PTAS as an innovative, non-invasive treatment modality for epilepsy. We present a method for phase estimation and PTAS delivery tolerant to the epileptic EEG. We then deploy our algorithm to demonstrate the intracranial correlates of auditory stimulation and an ability to reduce IEDs and improve next-day cognitive performance. PTAS offers a method for personalized interventions based on real-time EEG oscillation estimation, and here, we demonstrate its potential for tailored epilepsy management. This approach represents an advancement in our understanding of PTAS and its application in children with epilepsy, a vulnerable population in need of sleep-targeted intervention.

Limitations of the study

Due to the inclusion of pediatric populations in a complex and dynamic clinical environment, there are inherent limitations to our ability to interpret these findings. First, this study was conducted in the epilepsy monitoring unit, which differs from the typical controlled sleep laboratory environment. Our participant cohort did exhibit variability in their demographics, epilepsy conditions, and intracranial monitoring surgical plan. Second, our data are unable to characterize the impact of prolonged PTAS on seizure activity or neurocognition. Finally, while all stages of sleep are important, our experiment only administered PTAS during NREM sleep.

Our findings and these limitations, therefore, suggest several avenues for further study, notably, a longer-duration experiment testing the effects of PTAS over several weeks to months, which would not be practical with sEEG electrodes. Future experiments may also target SOs intracranially in cortical and subcortical regions implicated in the control of sleep, including the thalamus and the anterior insula, or in regions implicated in the patient’s individualized epileptogenic network.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, George Ibrahim (george.ibrahim@sickkids.ca).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Clinical data: The EEG recordings (scalp and intracranial), video data, neuroimaging (CT and MRI), neurocognitive assessment scores, and other clinical data generated in this study contain sensitive personal health information from pediatric participants and cannot be made publicly available to protect patient privacy and comply with the research ethics board requirements and privacy regulations. These data may be made available to qualified researchers, after deidentification, through a data access agreement that includes appropriate institutional review board approval, data use agreements, and data security protocols. Requests for data access should be directed to the lead contact, who will facilitate communication with the institutional data access committee.
• •Code: The TWave algorithm is publicly available at https://doi.org/10.5281/zenodo.17666777 or https://github.com/gmilab/ptas_benchmarks. The WebPVT assessment code is publicly available at https://doi.org/10.5281/zenodo.15009246 or https://github.com/gmilab/webpvt.
• •Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request.

Acknowledgments

We extend our deepest gratitude to all the children and their families who participated in this study. Your invaluable contributions have made this research possible.

We also acknowledge the generous support of funding agencies. This work was supported by Canadian Institutes of Health Research (CIHR) project grants, the Abe Bresver Chair in Functional Neurosurgery, Garry Hurvitz Centre for Brain and Mental Health and SickKids Precision Child Health grants held by G.M.I., and a CIHR Banting & Best Doctoral Canada Graduate Scholarship (CGS-D) and a SickKids Restracomp PhD Scholarship held by S.M.W.

Finally, we thank the clinical and research staff at The Hospital for Sick Children for their dedication and assistance throughout the study.

Author contributions

Conceptualization, S.M.W., K.M., N.M.W., H.S., O.A., A.G.W., A.H., H.O., P.J., A.O., J.T.R., M.L.S., S.W., E.D., and G.M.I.; data curation, S.M.W., K.M., N.M.W., O.A., C.M., R.S., H.O., P.J., A.O., E.N.K., M.L.S., L.S., S.W., E.D., and G.M.I.; formal analysis, S.M.W., C.M., E.N.K., and G.M.I.; funding acquisition, S.M.W. and G.M.I.; investigation, S.M.W., V.L., K.M., N.M.W., S.C.C., O.A., P.J., A.O., E.N.K., M.L.S., and G.M.I.; methodology, S.M.W., O.A., R.S., A.G.W., A.H., H.O., P.J., A.O., E.N.K., M.L.S., S.W., E.D., and G.M.I.; project administration, S.M.W., E.D., and G.M.I.; resources, C.M., R.S., and G.M.I.; software, S.M.W., V.L., and G.M.I.; supervision, A.G.W., A.H., H.O., P.J., A.O., J.T.R., M.L.S., E.D., and G.M.I.; validation, S.M.W., V.L., S.C.C., H.O., P.J., A.O., and G.M.I.; visualization, S.M.W., V.L., H.S., S.C.C., O.A., and G.M.I.; writing – original draft, S.M.W.,. K.M., N.M.W., H.S., S.C.C., and G.M.I.; writing – review & editing, S.M.W., V.L., and G.M.I.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
TWave Algorithm	Li et al.24	https://github.com/gmilab/ptas_benchmarks
Python 3.11 and 3.12	Python Software Foundation	https://python.org
MNE-Python 1.7.1	Gramfort et al.30	https://github.com/mne-tools/mne-python
Autoreject	Jas et al. 2017161	https://autoreject.github.io/stable/index.html
YASA	Vallat et al.162	https://yasa-sleep.org/index.html
WebPVT	Mithani et al.35	https://github.com/gmilab/webpvt
PsychoPy	Peirce et al.163	https://www.psychopy.org/

Experimental model and study participant details

Participants were prospectively recruited through study protocols approved by the research ethics board of the Hospital for Sick Children, Toronto, Ontario, Canada. Informed consent and assent were obtained from all participants and their caregivers in compliance with the Code of Ethics of the World Medical Association (Declaration of Helsinki).

Twenty-seven children admitted for intracranial epilepsy monitoring at the Hospital for Sick Children participated in this study (Tables S1 and S2). All enrolled children were diagnosed with drug resistant epilepsy with a focal hypothesis and underwent surgical placement of SEEG depth electrodes (0.86 mm diameter, AdTech Medical, Oak Creek, WI, USA) for seizure mapping. The location, laterality, and number of electrodes implanted were dictated solely by clinical needs and varied between patients. Simultaneous scalp EEG were acquired at standard 10–20 locations with Ag-AgCl electrodes affixed with collodion¹⁶⁴ and further secured by gauze head wrapping. Primary seizure semiology was classified as focal aware in 52%, with 67% having motor involvement. Over half of the participants (52%) had sleep-related seizure activity. A known genetic cause of epilepsy was identified in 21% of participants. Full participant demographics are provided in Tables S1 and S2.

Discovery and experimental cohorts

The first fifteen children were recruited into the Discovery cohort and received auditory stimulation at random intervals (3–30 s) throughout the night after sleep onset. Children were recruited after their second night of invasive monitoring, after a sufficient number of seizures were recorded. Auditory stimuli consisted of 50 ms bursts of 1/f pink noise with a 5 ms sinusoid taper at the beginning and end. Stimuli were delivered through sleep-compatible headphones (SleepPhones, AcousticSheep LLC, Erie, PA, USA), with volume calibrated to 55 dB by holding the sound emitting segment of the headband against an R8050 Sound Level Meter (REED Instruments, Newmarket, Canada). The stimuli were delivered at random intervals (interstimulus delay of 3–30 s) during N2 and N3 sleep throughout the night, without regard for the phase of the slow oscillation.

The subsequent twelve children were recruited into the Experimental cohort and received at least two nights of counterbalanced up-PTAS (up-phase targeted auditory stimulation). Children were recruited after their second night of invasive monitoring, after approval by the neurology team and received up-PTAS during N2 and N3 sleep. Stimulation and no-stimulation blocks were randomized across nights and patients to control for medication changes, day-to-day variability, and other potential confounds.

In the latter cohort, auditory stimuli consisted of 50 ms bursts of 1/f pink noise with a 5 ms sinusoid taper at each end. Stimuli were delivered through a Dell AE515M Pro Sound Bar (Dell Technologies Inc., Round Rock, USA) mounted above the bed using a custom-built wall mount, with volume calibrated to 55 dB at the patient’s sleeping head position using the R8050 Sound Level Meter (REED Instruments, Newmarket, Canada).

Method details

Electrophysiological data acquisition and signals processing

Scalp and intracranial EEG data were acquired simultaneously using the 256-channel Natus Quantum amplifier and Natus Neuroworks 8 software at 2048 Hz, referenced to a quiet intracranial white matter electrode selected by a board-certified electrophysiologist.

EEG data were exported in EDF+ format and imported into MNE-Python.³⁰ Electrode localization followed previously established protocols.²³ Artifacts in scalp channels were removed using default parameters with the MNE autoreject toolbox.¹⁶¹ Artifactual intracranial channels were excluded if localization placed them outside of the brain (e.g., localized to the dura or fixation bolt), or if marked as faulty in visual review by a board-certified epileptologist, or in segments where signals exceeded 2500 μV peak-to-peak. The EEG data were re-referenced to the common average separately for intracranial and scalp electrodes. Sleep staging was performed with the YASA toolbox¹⁶² using default classifiers,¹⁶⁵ and six subjects were independently staged by a certified sleep technologist (C.M.) with good agreement (Figure S4). YASA has been previously used and validated in pediatric clinical populations, including children with epilepsy.^166,167,168 All analyses were conducted in Python 3.12 using MNE 1.7.1.³⁰ Electrode localization was performed in a semi-automated manner as detailed previously.²³ Briefly, post-insertion CT images were linearly co-registered with each participant’s pre-operative T1 structural MRI, then non-linearly co-registered to the MNI152 atlas. Electrodes were identified automatically within the CT image and clustered into trajectories, then manually matched to the surgical plan and EEG recording. Electrode contacts were labeled using the Automated Anatomical Labeling, version 3 (AALv3) atlas,¹⁶⁹ based on their coordinates in template space.

A system for phase-targeted stimulation in children with epilepsy

The pediatric epilepsy monitoring unit (EMU) offers a unique environment to study sleep improvements in children with epilepsy. To perform phase-targeted auditory stimulation (PTAS), also referred to as closed-loop auditory stimulation (CLAS) or phase-locked auditory stimulation (PLAS), in the EMU, we developed a system specifically designed for overnight studies in this population. This system addresses critical considerations in the pediatric EMU. Our system is designed to be unobtrusive in the limited space available in the hospital room, allow for patient mobility within the room, operate overnight in a completely autonomous manner, and reliably deliver auditory stimulation despite patient movement and sleeping positions, while considering the substantial head dressing applied for clinical scalp and intracranial EEG.

Our PTAS system records data live from the Fpz and Cz scalp EEG electrodes and a ground EMG electrode using a battery-powered OpenBCI Cyton amplifier¹⁷⁰ within a custom-built enclosure. The data is sampled at 256 Hz and wirelessly transmitted live to a Microsoft Surface Go 2 tablet at the bedside. A custom application built on Python 3.11 and PyQt 5 receives the live EEG data, applies a one-pass 60 Hz notch filter (Q = 20), and stores the data into a 2 s FIFO buffer. The EEG analysis is run upon every buffer update. The amplitude, frequency, and phase within the SO band (0.5–2 Hz) is estimated with the TWave algorithm. The presence of an SO is evaluated based on the current sleep stage and the peak-to-peak amplitude of the EEG within the SO band. If an SO is detected, and if the target phase is anticipated to occur within the next 100 ms, the algorithm schedules a stimulation to occur at the anticipated time to occurrence of the target phase (e.g., SO peak).

The TWave real-time oscillation estimator algorithm

We designed the TWave algorithm to estimate and measure sleep slow wave activity in real-time, while tolerating movement artifact and IEDs in the epileptic EEG. Oscillation parameters are estimated through convolution with a family of truncated complex Morlet wavelets at varying scale factors (LF: 20 wavelets with center frequencies spanning 0.5 to 2 Hz, HF: center frequencies of 10, 20, and 30 Hz). The low-frequency wavelet of best fit when convolved with the 2 s FIFO buffer is considered the dominant frequency. An estimate of broadband amplitude is derived from an average of the high frequency wavelets. The vector norm and angle estimate the oscillatory amplitude and phase, respectively. Amplitude derived from LF wavelets must exceed 150 μV peak-to-peak, equivalent to the standard criteria of a 75 μV deflection from baseline,¹⁷¹ to be considered a sleep slow oscillation. The HF/LF amplitude ratio must also be below 0.1, to exclude IEDs, artifact, and as a gate for NREM sleep. Otherwise, the current signal is considered not to be an SO and no stimulation is cued. The TWave algorithm has been shown to offer higher phase precision and immunity toward epileptic activity²⁴ and does not deliver stimulation in response to high amplitude interictal epileptiform activity nor ictal activity. An overview of the phase targeting accuracy of the TWave algorithm, using data from the Discovery cohort is presented in Figure S2.

Neurocognitive testing

When feasible, considering compliance and clinical scheduling, these children also received daily neurocognitive testing with the Go/NoGo response inhibition test and the psychomotor vigilance task. Neurocognitive testing was conducted at least 2 h before or after any ictal events.

The Go/NoGo (GNG) response inhibition test (Figure 6A), adapted from the previous work of our colleagues,^172,173 is a well-established neurocognitive paradigm that primarily evaluates the ability to inhibit pre-potent responses. We employed a reciprocal-ratio GNG task consisting of serially presented geometric stimuli consisting of a single randomized shape with randomized color designed to maintain interest from younger children, delivered using PsychoPy¹⁶³ on Python 3.11. Children were asked to respond with a button press on a handheld gaming controller (Logitech F310, Logitech International S.A., Lausanne, Switzerland) as quickly as possible upon presentation of any stimuli regardless of shape or color (“Go” trials), except when shapes were superimposed with a white ‘X’ mark (“NoGo” trials). Stimuli were presented for a duration ranging from 300 to 700 ms, with an interstimulus interval of 650–1300 ms, with a random jitter of 200 ms. The stimulus duration and the interstimulus duration were adjusted algorithmically based on task performance as previously described.³² The task consists of an inhibition block (75% Go trials, 25% NoGo trials) and a vigilance block (25% Go trials, 75% NoGo trials), both consisting of 80 NoGo trials, presented in a randomized order. Children were asked to prioritize response speed, and therefore the higher proportion of Go trials in the inhibition block requiring a button press elicits a pre-primed motor response. Accuracy in withholding the button press is therefore an index of response inhibition.³² Performance (accuracy and reaction time) in the vigilance condition was assessed to confirm that changes in response inhibition accuracy were not driven by moment-to-moment fluctuations in vigilance or alertness.

The psychomotor vigilance task (PVT), previously described by our group,³⁵ is another commonly used neurocognitive paradigm indexing alertness. Children attend to a centrally presented red square on the screen (Figure 6A), and are asked to respond as quickly as possible with a button press when the square changes color to a luminosity and saturation-matched blue square. Blue squares are presented at a random interval between 2 and 5 s. Alertness is indexed by the mean reaction time or proportion of responses exceeding specific response time thresholds. The PVT paradigm was written in ES2016, presented in a Chromium browser, and timing was measured through the ES2016 Performance API.

Both tasks were presented to children from a Windows 11 laptop on a bedside cart connected to an external display affixed to a moveable overhead arm and positioned at a comfortable height and angle and at a viewing distance of approximately 60 cm. Task timing was recorded on the EEG using a trigger signal provided via an MMBT-S Trigger Interface Box (Neurospec AG, Switzerland).

The nAEP (nocturnal auditory evoked potential)

Scalp and intracranial EEG data acquired from the Discovery cohort were imported into MNE-Python³⁰ and preprocessed as above, then filtered from 0.5 to 2 Hz and epoched to the time of random stimulation delivery. Timeseries were averaged across trials. The time-locked averages were visually inspected and compared to the 95% baseline confidence interval computed across the entire night. Subjects without a visually confirmed nAEP were excluded from the nAEP analysis. Intracranial EEG channels were labeled with the AALv3 atlas,¹⁶⁹ as above, and averaged within regions, then averaged across subjects, then subject to a principal components analysis.

Estimation of optimal target phase for PTAS

Scalp EEG data from the Discovery cohort were imported into MNE-Python³⁰ and preprocessed as above, then epoched from −7000 ms to +7000 ms relative to each auditory stimulation as marked in the trigger channel. Epochs with artifacts in the Fpz or Cz channels were removed from the analysis. To compute the true phase of the endogenous slow oscillation at time of the auditory stimuli, the Fpz-Cz bipolar timeseries was bandpass filtered at 0.5–2 Hz with a two-pass 2^nd order Bessel filter, and phase was computed from the Hilbert transform of the band-limited signal. Separately, mean power across all scalp EEG channels in the slow oscillation frequency band (0.5–2 Hz) was computed by windowing with discrete prolate spheroidal sequences (DPSS) tapers between 500 and 2000 ms, and 2000 to 5000 ms post-auditory stimulation. For each trial, we thus obtain the phase at stimulation and the resultant EEG power post-stimulation.

Phase-aligned average timeseries

The SO phase at time of stimulation in the Discovery cohort, computed as above, was used to dichotomize trials into either the “Up” phase bin comprising phase values from 7π/6 to π/6, or into the “Down” phase bin consisting of the remainder phases (Figure 4A). Within each bin, stimulation markers were realigned to a common phase to facilitate averaging and comparison of mean effects.¹⁷⁴ Briefly, the stimulation marker was phase-aligned forward or backwards to the closest sample to the mean phase of each bin. Average evoked timeseries were then computed on epochs derived from the phase-aligned stimulation markers. Differences between Up and Down phase bins were evaluated through permuting bin labels and adjusted for FDR at q < 0.05. Separately, whole scalp EEG was averaged within each phase bin, then averaged across subjects.

Time-locked average timeseries

EEG acquired from the Experimental cohort were imported into MNE-Python,³⁰ as above, and filtered from 0.5 to 2 Hz, and epoched to the time of auditory stimulation delivery (PTAS condition) or time at which stimulation would have been delivered if not withheld (muted condition). Timeseries were averaged across trials, then across participants.

Fpz-Cz bipolar timeseries

In the Fpz-Cz bipolar channel, significant differences between the PTAS and MUTED conditions were evaluated by randomly permuting the condition labels (N = 5000), then thresholding at q < 0.05 with FDR correction.

Whole scalp analysis

EEG data from all scalp channels was averaged across stimulations, then across subjects. The timing of the first two post-stimulation peaks was selected and significant differences between PTAS and MUTED conditions were evaluated at both time points by randomly permuting condition labels (N = 5000), then thresholding at q < 0.05 with FDR correction to identify channels exhibiting significant differences between conditions.

Time-locked power

The Fpz-Cz timeseries within each bin were then averaged across all subjects in the Experimental cohort. Separately, the time and frequency resolved power in the Fpz-Cz bipolar channel was computed with the Morlet wavelet transform, then differences between condition estimated using a linear mixed effects model fitted to each time-frequency element (Equation 1).

$$\text{Power}\sim \text{Phase}\text{Bin}+\left(1\right|\text{Subject})$$ (Equation 1)

Intracranial power analysis

Slow oscillation power (0.5–2 Hz) between 0 and 4 s was computed with a time-frequency decomposition as above for every intracranial electrode. Electrodes were sorted into regions of the AALv3.¹⁶⁹ Power differences between condition within each AAL region was estimated with a linear mixed effects model fitted to each region (Equation 1).

Effect of PTAS on IEDs

IEDs were detected automatically using an established algorithm¹¹⁶ from the scalp EEG recordings. In brief, amplitude of broadband (0.1–80 Hz) and band-limited (25–80 Hz) EEG timeseries were computed from the Hilbert transform and z-scored and thresholded at a Z score of 3. Suprathreshold periods between 2.5 and 200 ms in band-limited amplitude accompanied by a concomitant increase in broad-band amplitude were considered to be IEDs. Epileptiform discharges were excluded from this analysis if they occurred during seizures as marked in the clinical EEG recording by an expert epileptologist.

A linear mixed effects model (Equation 2) was used to estimate the difference in the number of IEDs present within each 30 s epoch used for sleep staging.

$$\text{IED}\text{count}\sim \text{Sleep}\text{stage}\ast \text{Experimental}\text{mode}+\left(1\right|\text{Subject})$$ (Equation 2)

Similar linear mixed effects models (Equation 2) were used to estimate differences in the number of spikes occurring within specific time windows (−7 to −4 s, 0 to 3 s, 2 to 5 s, and 4 to 7 s) relative to the onset of auditory stimulation between PTAS and MUTED conditions in the Experimental cohort.

Effect of PTAS on neurocognition

Data from the GNG and PVT computer-based tests, collected as above, were used to estimate the effect of PTAS on neurocognition, specifically in the domains of response inhibition and vigilance using data from the Experimental cohort. Differences in response inhibition, indexed by accuracy in withholding a button press response to the presentation of a NoGo stimuli, were evaluated with a linear mixed effects model (Equation 3).

$$\text{Accuracy}\sim \text{Experiment}\text{Mode}+\left(1\right|\text{Subject})$$ (Equation 3)

Changes to response inhibition due to day-to-day fluctuations in vigilance and alertness were controlled for with the vigilance condition in the GNG, and with the psychomotor vigilance task. Vigilance as indexed by reaction time and accuracy in GNG, and vigilance as indexed by mean reaction time and accuracy at 330 ms and 500 ms thresholds on the PVT were non-significant between experimental and withheld nights, even when accounting for the effects of current dose and changes to anti-seizure medication.

Visualizations

Data figures were generated using matplotlib, seaborn, MNE’s visualization tools, surfplot, and VTK, all in Python 3.12.

Quantification and statistical analyses

Unless otherwise stated, two-sided tests were used with a nominal α = 0.05. Linear mixed-effects (LME) models were fit in R 4.4.1 (lme4 v1.1–35.5) accessed via rpy2/pymer4 from Python 3.12; fixed-effect p-values and 95% CIs were obtained using pymer4’s implementation (Satterthwaite df), with models estimated by REML for continuous outcomes. For time–frequency and regional analyses we fit LME models at each element using the structures specified in the relevant subsections. Unless otherwise specified, we controlled for multiple comparisons using the Benjamini–Hochberg FDR at q < 0.05 within each family of tests (time points/channels for scalp, time–frequency pixels for TF maps, and AAL regions for intracranial analyses). Additional details can be found in figure captions and detailed methods above.

Circular analyses used pycircstat; non-uniformity of phase distributions were assessed with Rayleigh’s test, and circular means were reported for preferred phases in early (0.5–2 s) and late (2–5 s) post-stimulus windows.

For time-domain comparisons (Fpz-Cz and whole-scalp), inference was based on Monte-Carlo label-permutation tests (5,000 permutations), constraining permutations within subject to preserve dependence; resulting p-values were FDR-adjusted at q < 0.05. Peak-based whole-scalp tests were performed at prespecified post-stimulus peaks.

Spike counts were analyzed with linear mixed-effects models including sleep stage and experimental mode as fixed effects and subject as a random intercept. For neurocognitive outcomes, Go/NoGo accuracy was logit-transformed and modeled with an LME including experimental mode as a fixed effect and subject as a random intercept, with vigilance measures (Go/NoGo vigilance condition and PVT metrics) analyzed separately to verify that effects were not attributable to arousal.

Additional resources

The TWave algorithm is open source and freely available at https://github.com/gmilab/ptas_benchmarks. The code for our adapted PVT task is open source and freely available at https://github.com/gmilab/webpvt.

Published: January 8, 2026