Electrophysiological Signatures of Visual Sensations Elicited by Direct Electrical Stimulation

Abstract

Direct electrical stimulation of the human cortex can produce subjective visual sensations, yet these sensations are unstable. The underlying mechanisms may stem from differences in electrophysiological activity within the distributed network outside the stimulated site. To address this problem, we recruited 69 patients who experienced visual sensations during invasive electrical stimulation while intracranial electroencephalography (iEEG) data were recorded. We found significantly flattened power spectral slopes in distributed regions involving different brain networks and decreased integrated information during elicited visual sensations compared with the non-sensation condition. Further analysis based on minimum information partitions revealed that the reconfigured network interactions primarily involved the inferior frontal cortex, posterior superior temporal sulcus, and temporoparietal junction. The flattened power spectral slope in the inferior frontal gyrus was also correlated with integrated information. Taken together, this study indicates that the altered electrophysiological signatures provide insights into the neural mechanisms underlying subjective visual sensations.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12264-025-01464-7.

Introduction

Direct electrical stimulations of the human cortex can produce visual sensations, from simple phosphenes to formed visual hallucinations [1–12]. This offers a unique opportunity to study the brain networks involved in visual sensations, which may advance efforts to improve the performance and long-term stability of visual prosthetic devices [13] and provide insights into understanding visual hallucinations in multiple brain disorders.

Visual sensations emerge not through isolated activation of focal cortical regions by intracranial electrical stimulation (iES), but rather through the coordinated engagement of distributed neural networks. Compelling empirical evidence has revealed that stimulation-induced visual phenomena (e.g., phosphenes) consistently recruit emergent activity in higher-order association areas, particularly the temporoparietal junction (TPJ) [4]. The aperiodic component of the power spectrum, commonly termed 1/f or scale-free activity, constitutes a fundamental feature of neural signals, exhibiting a characteristic power-law frequency dependence where signal power decreases systematically with increasing frequency. This broadband activity stands in marked contrast to oscillatory processes, which manifest as narrow-band peaks at specific frequencies, and is believed to originate from stochastic neural dynamics that directly reflect the underlying excitatory-inhibitory (E/I) balance within cortical circuits [14–16]. Critically, variations in this aperiodic slope serve as a sensitive electrophysiological marker of distinct conscious states: it steepens characteristically during sleep and general anesthesia relative to wakefulness [17], while exhibiting pronounced flattening under the influence of psychedelic agents like lysergic acid diethylamide [18]. The ability to computationally dissociate these aperiodic components from classical oscillatory activity in spectral analyses has provided researchers with a powerful framework for obtaining more physiologically meaningful interpretations of neural dynamics, as these distinct spectral features likely reflect separable underlying mechanisms [19].

In addition, dynamic network communication is also altered in the sensation state, which influences how information is integrated across the network. Lesion-related network mapping of sensations (e.g., hallucinations) has identified a common network that incorporates the superior temporal sulcus (STS) and cerebellum [20]. Functional neuroimaging studies have also revealed aberrant functional connectivity among different brain regions – commonly involving the superior temporal cortex – in patients with auditory hallucinations versus those without, or during a hallucinatory state versus the resting state [21]. Further, network changes between the inferior frontal cortex (IFC), STS/middle temporal gyrus, and premotor area have been associated with robot-induced presence hallucinations in healthy participants [22]. A recent study revealed that Φ^* (Phi), an information-theoretic measure of the complexity and organization of brain network dynamics [23], robustly correlates with changes in consciousness across states such as sleep, wakefulness, and general anesthesia [24]. This finding highlights Φ^* as a reliable indicator of conscious states: higher values are associated with heightened awareness and lower values are linked to diminished or absent consciousness. However, it remains unknown whether integrated information at the network level changes during visual sensations and which brain regions mediate this integration.

The current study aimed to investigate neural mechanisms of visual sensations by measuring the aperiodic slope of the power spectral density (PSD) of local field potentials and information integration (measuring Φ^*) over large-scale networks in drug-resistant epilepsy patients (n = 69) who subjectively reported visual sensations elicited by iES. The analysis of synchronously recorded intracranial electroencephalography (iEEG) data revealed both flattened aperiodic slope and decreased global information integration during visual sensations. Furthermore, the latter change arose from altered system architecture mediated by the temporoparietal junction (TPJ), IFC, dorsolateral prefrontal cortex (DLPFC), and posterior STS, with concomitant disruptions to network communication dynamics, including in the sensory cortex.

Materials and Methods

Patients

The study cohort comprised patients undergoing iEEG monitoring and iES for medically refractory epilepsy at Sanbo Brain Hospital, Capital Medical University (Beijing, China) between January 2016 and September 2020. The inclusion criteria were: (1) verbal reporting of visual sensations specifically following iES, and (2) concurrent iEEG recordings during stimulation. From this population, we identified 69 eligible patients (age range: 5–43 years, mean ± SD = 22.6 ± 9.2 years; epilepsy onset age: 0–38 years, 12.1 ± 7.8 years; 26 female). This sample included 37 newly recruited participants and 32 patients whose cortical distribution of visual sensations had been previously mapped in our prior research [12]. None of the patients had a confirmed diagnosis of mental disorders, such as schizophrenia or bipolar disorder, based on available medical records and self-reported histories. However, detailed information on substance use history, which could influence visual hallucinations, was not systematically collected. Detailed clinical information about the patients, including the location of the epileptogenic zone, etiology, and the Engel epilepsy surgery outcome, are listed in Table S1. We only analyzed sensations elicited in non-epileptic zones, identified by the neurologists (Jing Wang and Mengyang Wang). The study was approved by the Ethics Committee of the Institute of Psychology, Chinese Academy of Sciences (ID H20034). All patients, or their guardians, provided informed consent to the surgical procedure and the review of data for scientific purposes.

Intracranial Electrode Placement and Localization

Depth electrodes were implanted in patients to localize the cerebral structures involved in seizure onset and propagation. All electrodes (8–16 contacts; length: 2 mm, diameter: 0.8 mm, spaced 1.5 mm apart; Huake-Hengsheng Medical Technology Co., Ltd, Beijing, China) were placed based on clinical criteria. To precisely determine electrode locations, a semi-automated approach was used [25]. Specifically, a postoperative computed tomography (CT) was nonlinearly registered to a preoperative magnetic resonance imaging (MRI). The spatial coordinates of electrodes were then calculated based on the trajectory of detected electrode clusters in individual spaces. Stimulated site coordinates were normalized to the Montreal Neurological Institute (MNI) standard space based on the combined volumetric- and surface-based normalization between individual structural images and the MNI standard space [26]. We localized electrodes to brain areas (n = 180) and then grouped them into regions (n = 22) according to the parcellation scheme of the Human Connectome Project (HCP) atlas [27].

Intracranial Electrical Stimulation Procedure

Following spontaneous seizure recordings, patients received iES to elicit electro-clinical seizures and map eloquent cortex. A biphasic pulse of electrical stimulation (Nicolet Cortical Stimulator, WI, USA) was delivered across two adjacent electrodes at 0.3 or 0.5 ms width, 4–5 s duration, and 50 Hz frequency. Simultaneous iEEG recordings were acquired from all electrodes using a Nicolet system (128 channels; sampling rate 512 Hz; Thermo Nicolet Corp.). Stimulation was immediately terminated upon observation of clinical responses (e.g., pre-seizure auras or seizures) or EEG afterdischarges. The stimulation intensity ranged from 0.1 to 6 mA with increments of 0.1–0.5 mA. Patients were blinded to the time and location of electrical stimulation.

iEEG Data Preprocessing

Raw iEEG signals were initially processed through bipolar referencing by subtracting the signal from the adjacent contact, followed by band-pass filtering in the range of 0.5 Hz to 120 Hz using a zero-phase (forward and reverse) 4^th order Butterworth Infinite Impulse Response (IIR) filter. In addition, 50-Hz power line noise and its harmonics were effectively removed by applying a notch filter. The iES lasts for 4–5 s, with the stimulation period defined as a 4-s duration beginning from the onset of the electrical stimulation. This window was chosen based on previous research indicating that iES-induced visual sensations typically occur during the initial phase of stimulation [4]. By analyzing this period, we ensured the capture of neural events associated with visual sensations, which are transient and unlikely to persist beyond the stimulation period. This approach allowed us to focus on the genuine stimulation phase rather than a post-sensation interval. A 4-s duration was also selected as the baseline period, during which no electrical stimulation was presented. A five-cycle Morlet wavelet was applied to transform the extracted time series signals into the time-frequency domain. Subsequently, during both the stimulation and baseline periods, the time-averaged PSD was calculated for each recording contact, providing the spectral characteristics of local neural activity [28]. During the stimulation period, the contacts near the stimulating contacts (typically within a range of 40 mm) were excluded because their signals showed typical stimulus artifacts and also often exceeded the recording limits allowed by the device. The resultant signals were also visually inspected for artifacts before further analysis.

Visual Sensation and Control Trials

During the electrical stimulation phase, patients were instructed to count aloud. Following electrical stimulation, the patients were asked an open-ended question about whether any experiences were elicited, with follow-up questions as needed to clarify the experience. This task design ensured that language production was comparable between visual sensation trials and control trials, as patients were engaged in verbal tasks throughout both conditions. By maintaining consistent verbal engagement across trial types, the potential confounding effects of language production were minimized, allowing us to more accurately isolate neural activity associated with visual sensations. We classified visual sensations based on these subjective reports of the patients. If patients reported seeing images (e.g., phosphenes, shapes, objects, animals, people, and visual scenes) without physical stimuli, we classified these responses as visual sensations. Similar to the previous study, most of these sensations were simple phosphenes and shapes [1]. The patients knew that no stimuli were actually presented. Note that any disturbance or deformation of a real object (i.e., incorrect perception of external stimuli) that patients reported was classified as an illusory perception and excluded.

We classified recordings as sensation trials when patients orally reported visual sensations during the stimulation period. There were also control trials in which patients did not report altered experiences of perception during the stimulation period. The control contacts, through which electrical stimulation was delivered but patients did not report altered perception, were located in the same electrode and as close as possible to the sensation-eliciting contacts. The stimulus intensity was matched between sensation and control trials (P = 0.389). Note that we only considered the shared recording electrodes between sensation and control trials.

Calculation of the Slope OF The Aperiodic Power Spectrum

We utilized the FOOOF toolbox [19] to extract the aperiodic slope. Within a broad frequency range (e.g., 2–120 Hz), a knee can occur, characterized by two distinct 1/f regimes, each with its own aperiodic slopes, separated by a “bend” at a knee frequency [29]. Since action potentials and firing rate changes can alter gamma and high gamma power at frequencies as low as 50 Hz [30–32], a frequency range of 2–45 Hz was selected, in line with prior studies in both EEG and iEEG studies [33–35]. Moreover, the positive correlation between the E/I ratio and PSD slope diminished with increasing frequency; the PSD slope below 50 Hz had a higher correlation with the E/I ratio [15]. The other parameters were as follows: aperiodic mode, fixed; peak width limits, [1–8]; maximum number of peaks, 6; minimum peak amplitude exceeding the aperiodic fit, 0.05; default settings otherwise. The FOOOF toolbox parameterized the periodic components of PSDs as Gaussian peaks and removed their effects on the power spectrum before estimating the aperiodic components of PSDs; this can better estimate the aperiodic slope from neural signals. The exponent value [i.e., (−1)$\times$ slope] output from the FOOOF toolbox was used to measure the aperiodic slope. We also removed outliers using the three-sigma rule. Given prior evidence of divergent aperiodic slope characteristics across frequency bands [14, 36], we conducted comparative analyses in the 2–19 Hz and 20–45 Hz ranges to elucidate these frequency-dependent dynamics.

Statistical Analysis

To analyze changes in the aperiodic slope associated with visual sensations elicited by electrical stimulation, we used linear mixed-effects models to account for the hierarchical structure of our data. For each of the 22 brain regions defined by the HCP atlas, we fitted a model [exponent ~1 + trial_type × stimulation_phase + (1|patients:recording_contacts) + (1| patients)] that included fixed effects for trial type (sensation vs control), stimulation phase (baseline vs stimulation period), and their interaction, while controlling for random effects of patients and electrode recording contacts nested within patients.

Electrode recording contacts were mapped to specific HCP brain regions [27] based on their MNI coordinates using a spherical search algorithm with a 4-mm radius. Contacts were classified according to the most frequently occurring atlas label within this search volume, applying a minimum threshold of 20 voxels for reliable classification. The same methodological approach was applied to assign contacts to the Yeo-7 functional networks [37].

For each brain region, we assessed the significance of the main effects (trial type and stimulation phase) and their interaction using the Satterthwaite approximation for degrees of freedom. When significant interaction effects emerged (P <0.05, Bonferroni-corrected for multiple comparisons across regions), we applied post hoc analyses using estimated marginal means to evaluate: (1) sensation-control trial differences within each stimulation phase (baseline and stimulation periods), and (2) baseline-stimulation period differences within each trial type (sensation and control trials). This analytical approach enabled the identification of visual sensation-specific electrophysiological signatures while accounting for potential electrical stimulation artifacts. We applied identical analytical procedures to examine functional networks defined by the Yeo-7 functional networks [37].

Analyzing Information Integration over the Brain

In this study, we quantified the information integrated into a given system using integrated information (Φ), proposed by Integrated Information Theory (IIT) as related to consciousness level [23]. Due to computational constraints, we utilized Φ^*, a practical surrogate derived from mutual information, which provides a feasible method for assessing Φ in large systems and is particularly amenable to experimental data. This approach has been validated in several prior studies [24, 38, 39].

IIT proposes that integrated information should be calculated between the least interdependent parts of a system, and the partition of the system delineating these parts is called the minimum information partition (MIP). This partitioning reveals the critical structural characteristics of neural systems, where interdependencies within regions on the same side are relatively strong, while those between regions on different sides are weaker [40].

Based on this framework, we defined two critical behaviors, “split” and “merge”, to elucidate how brain regions interact under different conditions. A “split” occurs when a pair of brain regions, initially on the same side of the MIP during a baseline period, move to different sides during the electrical stimulation period. Conversely, a “merge” happens when a pair of brain regions, initially on different sides of the MIP, come to reside on the same side during stimulation. These dynamics are essential for understanding the reconfiguration of neural networks in response to visual sensation.

Our analysis focused on both the values of Φ^*, indicating the level of integrated information, and on the system’s partitioning to elucidate changes in the underlying architecture of neural systems. The analytical process, detailed in Supplementary Fig. 1, comprised four main parts. Initially, for each recording session, we calculated Φ^* and identified the MIP for various subsystems across different brain regions at varying lag times. Subsequently, we determined the optimal lag time that maximized Φ^*, and then proceeded to analyze changes in Φ^* values during visual sensations. Finally, we examined the dynamic changes in the MIP, particularly focusing on the “merge” and “split” behaviors of brain region pairs under different stimulation conditions, either eliciting visual sensations or during control scenarios.

Analyzing the Relationship Between Aperiodic Slopes and Φ^* Values

To test whether the changed local aperiodic slopes were associated with the Φ^* values, Spearman’s rank correlations were used. For each patient, we calculated the changes in aperiodic slope for each region [IFC, IPL (inferior parietal lobe), MTComp (MT complex and neighboring visual areas), MTL (medial temporal lobe), VenV (ventral stream visual area), and AudA (auditory association cortex)], which demonstrated significant changes during visual sensations. The statistical significance of the correlations was tested based on the permutation and determined by a Bonferroni-corrected significance threshold of P <0.05. Similar analyses were also carried out for the remaining regions in the HCP22 atlas.

Data Availability

The MNI coordinates and current intensity for stimulation that elicited visual sensations are provided in a table (Supplementary file 1.xlsx). The electrophysiological data that support the findings of this study are available from the corresponding author upon reasonable request.

Results

Intracranial Electrode Distribution Across Patients

Six hundred and forty-one stimulation contacts elicited visual sensations (Fig. 1A, red; Supplementary file 1.xlsx for the MNI coordinates and current intensity) and 553 control contacts did not induce sensations after electrical stimulation (Fig. 1A, green). There were 5,612 recording contacts for sensation trials and 5,862 recording contacts for control trials, in which 5,353 contacts were shared between the two trial types. These shared recording contacts from 69 patients were distributed across a wide range of cortical areas (Fig. 1B) and were used for further electrophysiological analysis.

Fig. 1.

Distribution of stimulation and recording contacts. A Distribution of stimulation contacts that induce visual sensation (red) and control contacts (green). B Distribution of recording contacts shared by sensation and control trials. C The response ratio of sensation contacts among 22 regions of the HCP atlas. The ratio near each bar denotes the number of responsive contacts eliciting sensations vs the total number of contacts in each region. PriV: primary visual cortex; EarV: early visual areas; DorV: dorsal stream visual areas; VenV: ventral stream visual areas; MTComp: MT+ complex and neighboring visual areas; SenM: somatosensory and motor area; ParC: paracentral lobular and middle cingulate cortex; PreM: premotor area; PosOp: posterior operculum; EarA: early auditory area; AudA: auditory association cortex; InsFop: insular and frontal operculum; MTL: medial temporal lobe; LTL: lateral temporal lobe; TPJ: temporo-parieto-occipital junction; SPG: superior parietal gyrus; IPL: inferior parietal lobe; PCC: posterior cingulate cortex; mPFC: anterior cingulate cortex and medial prefrontal cortex; OFC: orbital and polar frontal cortex; IFC: inferior frontal cortex; DLPFC: dorsolateral prefrontal cortex.

Each responsive contact that elicited visual sensations was assigned to a region in the HCP22 atlas [27] if the center of the bipolar stimulation contacts was <5 mm away from that region. The primary and early visual cortex (PriV and EarV) hosted the highest number of stimulation contacts. In addition, the dorsal stream visual area (DorV), VenV, posterior cingulate cortex (PCC), MTComp, MTL, and TPJ each had 5% or more sensation-inducing stimulation contacts (Fig. 1C). To validate the replicability of induced visual sensations, seven patients were delivered stimulations by increasing the current density (<6 mA) after reporting visual sensations (145 repetitions). The repetition rate was as high as 91.8%.

The Aperiodic Slope Flattens During Visual Sensations

To determine the aperiodic slope of the PSD during elicited sensation, the iEEG recording data from each patient were characterized based on the four identified brain states: baseline and stimulation period for sensation trials, where patients reported visual sensations during electrical stimulation; and baseline and stimulation period for control trials, where no altered sensations were reported (Fig. 2).

Fig. 2.

Electrical stimulation and selection criterion of sensation and control contacts. The schematic electrode has 15 contacts (length: 2 mm, spacing: 1.5 mm). Stimulation contacts eliciting visual sensation are marked in red (sensation contacts), corresponding to sensation trials. Nearby stimulation contacts that do not elicit perceptual changes are marked in green (control contacts), corresponding to control trials. Stimulation contacts marked in black are not used in this study. For both sensation and control trials, the stimulation period is defined as a 4-s period during electrical stimulation, and the baseline is defined as a 4-s period before any electrical stimulation is given.

Using linear mixed-effects models, we analyzed the changes in aperiodic slope across brain regions delineated by the HCP22 atlas [27]. To identify regions specifically associated with visual sensations, we focused on regions showing significant interaction effects between trial type and stimulation phase, as well as significant differences in two critical post hoc comparisons: between the stimulation period and baseline in sensation trials, and between sensation trials and control trials during the stimulation period. This approach allowed us to isolate electrophysiological changes specifically linked to the experience of visual sensations. Twelve brain regions met these criteria: PriV, EarV, DorV, VenV, MTComp, AudA, insular and frontal operculum (InsFop), MTL, lateral temporal lobe (LTL), IPL, PCC, and IFC. In all these regions, aperiodic slopes became significantly flatter during the stimulation period compared to baseline in sensation trials, and were significantly flatter in sensation trials compared to control trials during stimulation (Fig. 3A). We have also illustrated the anatomical locations of these regions on brain atlas maps in Fig. 3B for reference.

Fig. 3.

Changes in aperiodic slope during visual sensations. A Twelve brain regions identified from the HCP22 atlas showing significant interaction effects between trial type (sensation trial and control trial) and stimulation phase (baseline period and stimulation period), with significantly flatter aperiodic slopes during stimulation compared to baseline in sensation trials and significantly flatter slopes in sensation trials compared to control trials during stimulation. Bar plots show estimated marginal means derived from linear mixed-effects models (± SEM) for each condition. B Anatomical locations of these 12 regions displayed on brain atlas maps. C Five functional networks from the Yeo-7 parcellation meeting the same significance criteria show similar patterns of aperiodic slope flattening during visual sensations. DA: dorsal attention network; DMN: default mode network; FP: frontoparietal network; LIM: limbic network; SM: sensorimotor network; VA: ventral attention network; VIS: visual network. Legend in panel A applies to both panels A and C. (linear mixed-effects models; *P <0.05, **P <0.01, ***P <0.001).

Similar linear mixed-effects analyses were applied to the functional networks defined by the Yeo-7 parcellation [37]. Using the same criteria to identify networks specifically associated with visual sensations, five networks emerged: visual network, dorsal attention network, ventral attention network, frontoparietal network, and default mode network (Fig. 3C). In all these networks, aperiodic slopes became significantly flatter during the stimulation period compared to baseline in sensation trials, and were significantly flatter in sensation trials compared to control trials during stimulation. These findings at both regional and network levels consistently demonstrate that visual sensations elicited by electrical stimulation are associated with a flattening of the aperiodic slope in a distributed but specific set of brain regions and networks.

Further analyses of frequency-specific aperiodic activity revealed both shared and distinct patterns across low- and high-frequency ranges. In the low-frequency range (2–19 Hz), we found a significant interaction between trial type and stimulation phase in five regions: PriV, MTComp, MTL, IPL, and PCC. These regions exhibited a consistent flattening of the aperiodic slope during stimulation compared to baseline in sensation trials, as well as flatter slopes in sensation trials relative to control trials during the stimulation phase. This pattern aligned with the results from our original 2–45 Hz analysis and was further supported by Yeo-7 network analysis, which identified consistent effects in four key networks: the visual network, dorsal attention network, frontoparietal network, and default mode network.

In contrast, the higher frequency range (20–45 Hz) showed a more restricted pattern of activity, with only the superior parietal gyrus (SPG) exhibiting a significant interaction between trial type and stimulation phase. Notably, this region was not identified in the lower frequency band or the original 2–45 Hz analysis, suggesting its unique role in higher-frequency neural dynamics. Yeo-7 network analysis in this frequency range revealed significant effects in three networks: the visual network, frontoparietal network, and default mode network, representing a subset of the networks identified in the lower frequency range. Together, these findings highlight both overlapping and distinct neural substrates underlying low- and high-frequency aperiodic activity, with broader involvement of cortical regions and networks in the lower frequency range and more specialized contributions in the higher frequency range.

Decreased Information Integration During Visual Sensations

The measurement Φ^*, based on the mutual information between brain regions, quantifies the information integrated into a given system. It is important to emphasize that we calculate information integrated based on brain regions rather than individual recording contacts as nodes to make the results more compelling for further analysis. Despite the variability in electrode implantation sites among patients, there is a significant overlap in the brain regions involved in the information calculation (Supplementary File 2.xlsx). We found that, in sensation trials, the mean Φ^* during electrical stimulation was significantly lower than that during the baseline period (P <0.001, n = 68, Wilcoxon test). In contrast, in control trials, there was no significant difference in the mean Φ^* between the baseline and electrical stimulation periods (P = 0.34, n = 68, Wilcoxon test) (Fig. 4A). In addition, when comparing Φ^* during the baseline period between sensation and control trials, no significant difference was detected (P = 0.52, n = 68, Wilcoxon test). However, a significant difference in Φ^* was found between sensation and control trials during electrical stimulation (P = 0.014, n = 68, Wilcoxon test). This indicated a reduction in global information integration among brain regions during visual sensations.

Fig. 4.

Changes in Φ^* and integrated structure during visual sensations. A Comparing Φ^* between stimulation and baseline periods for sensation and control trials. In sensation trials, Φ^* during electrical stimulation is significantly decreased compared with that during the baseline (Wilcoxon test, n = 68; ***P <0.001). In control trials, no significant difference existed. B Integrated brain regions during normal perceptual experience. The normalized number of pairs of brain regions whose relationship to the MIP is unchanged during the stimulation period compared to the baseline for control trials. In the upper triangle, each pair of brain regions belongs to the same side of the MIP during the baseline and stimulation; in the lower triangle, each pair of brain regions belongs to different sides of the MIP during the baseline and stimulation. C The definition of “merge” and “split” connections (A, B, C, D, and E represent five different brain regions). For example, the MIP during the baseline is AB | CDE, which changes to ACD | BE during the stimulation period. Then the pairs of brain regions AB, CE, and DE are classified as “split”, since they belong to the same partition during baseline but change to different partitions during the stimulation period. The pairs of brain regions AC, AD, and BE are classified as “merge”, since they first belonged to different partitions but changed to the same partition during the stimulation period. Other partitions of brain regions are unchanged during the stimulation period compared to the baseline and are not used for further analysis. D Changes of the integrated structure during visual sensations. E Hierarchical clustering of the “merge” and “split” connection matrix. The 22 brain regions form 3 clusters corresponding to networks incorporating the TPJ, IFC/DLPFC, or other regions, indicated by the red dashed line. The color-coded values in the connection (D) and matrix (E) are equal. Positive and negative entries indicate region pairs that tended to merge and split during visual sensations, respectively.

Changes of Integrated Structures During Visual Sensations

A key step in calculating Φ^* involves determining the MIP for a given system of brain regions, where a cut in the system causes the least information loss. This “weakest link” in the system demarcates brain regions forming integrated structures, i.e., regions associated on the same side of the MIP, which have causes and effects on each other. To characterize integrated brain regions during normal perceptual experience, we focused on control trials and measured the tendency of regions to be on the same side of the MIP during both the baseline and stimulation periods as well as a tendency to be on different sides of the MIP during both periods. The brain regions integrated during normal experience (the same side of the MIP indicated by warm colors in Fig. 4B) are sensory areas, InsFop, and MTL. The sensory areas are typically higher-order and located in the somatosensory (posterior operculum, PosOp), auditory (EarA and AudA), and visual (MTComp and LTL) cortex. The association of these brain regions into an integrated network structure is consistent with our normal multi-modal perceptual experience (supported by somatosensory, auditory, visual, and insular areas) and drawing on memory (MTL).

We hypothesized that sensations may occur when there is a reconfiguration of brain regions into atypical integrated structures. To test this hypothesis, we measured the tendency of paired brain regions to be associated with the same side of the MIP (i.e., Merge) or to be dissociated to the different side of the MIP (i.e., Split) when visual sensations were elicited by electrical stimulation (Fig. 4C) compared with the baseline. We found that three regions, the IFC, DLPFC, and TPJ, merged with multiple other regions. Specifically, the DLPFC/IFC tended to merge with SPG and EarA. In contrast, the TPJ tended to merge with VenV, PriV, MTComp, paracentral lobular, and mid-cingulate cortex (ParC). Notably, both the IFC and TPJ tended to merge with the MTL, PosOp, PCC, LTL, and IPL. We also found that the AudA tended to split with the ParC, PCC, PosOp, IPL, LTL, MTL, EarA, VenV, PriV, EarV, and MTComp (Fig. 4D). The hierarchical clustering analysis further confirmed 3 clusters corresponding to networks incorporating the TPJ, IFC/DLPFC, or other regions (Fig. 5E). Taken together, during visual sensations, visual areas were predominantly split from other sensory areas, particularly auditory (Fig. 4D, E), which may enable a “disconnected” visual experience. At the same time, these visual areas (MTComp, VenV, and PriV) merged with the TPJ, a region implicated in awareness [41, 42] and reorienting attention [43], whereas other sensory areas were merged with the IFC.

Fig. 5.

The relationship between the changes in the aperiodic slope of the IFC and the Φ^* values during visual sensations. A The stimulation period to the baseline period for the sensation trials. B The stimulation period for the sensation trials compared to that of the control trials (Spearman’s rank correlations; *P <0.05, **P <0.01).

Correlation Between Regional Aperiodic Slopes and Φ^* Values

During both sensation trials (stimulation vs baseline) and stimulation period (sensation trials vs control trials), the changes in the aperiodic slope of the IFC and the Φ^* values specifically exhibited the strongest correlation (r = 0.464, P = 0.029; r = 0.656, P = 0.0012) (Fig. 5). In contrast, during both control trials (stimulation vs baseline) and the baseline period (sensation trials vs control trials), the IFC region showed no significant correlation between the changes in aperiodic slope and the Φ^* values (r = 0.204, P = 0.87; r = −0.079, P > 0.999). These findings suggest that, during the sensation period, the flattening of the aperiodic slope in the IFC region is linked to a decrease in Φ^* values, and this association is independent of potential interference from electrical stimulation. The relationship between the changes in aperiodic slope and the Φ^* values in other brain regions of the HCP22 atlas under the four contrasting conditions can be found in Tables S2–5.

Rationale for Metric Selection

Initial spectral analysis of visual sensations revealed a distinctive "seesaw" pattern, characterized by decreased power in lower frequencies and increased power in higher frequencies. The aperiodic slope effectively captured these subtle spectral variations, suggesting an altered E/I balance that traditional oscillatory analyses failed to detect. Furthermore, Φ^* revealed decreased information integration across brain networks, linking these spectral changes to disruptions in consciousness. Compared to time-frequency decomposition, the aperiodic slope and Φ^* demonstrated nuanced spectral and network dynamics. These metrics offer mechanistic insights into visual sensations.

Discussion

Combining electrical stimulation with iEEG recordings, our findings showed: (1) flatter power spectral slopes in distributed regions during stimulation-elicited visual sensations, particularly in IFC, IPL, MTComp, MTL, VenV, and AudA; (2) significantly decreased Φ^* values, indicating impaired information integration during visual sensations; (3) significant correlations between IFC spectral slope changes and Φ^* values during sensations; and (4) atypical integrated network structures during visual sensations, reflecting brain network reorganization that deviated from control patterns. This included reconfigured interactions among IFC/DLPFC, posterior STS (primarily AudA), and TPJ (which was associated with visual areas during sensations), along with functional decoupling between the visual cortex and multiple sensorimotor regions.

Recent studies have sought to elucidate why certain neural stimulations evoke sensations while others do not, identifying several potential neural signatures. These include enhanced local functional responses [9], increased long-range connectivity [44], and elevated gamma-band activity in distal cortical areas [4, 45]. Our findings contribute to this body of work by demonstrating that the aperiodic slope (1/f background activity) may serve as an electrophysiological marker of the local E/I balance – a fundamental regulator of neural circuit function that has been extensively characterized in both computational models and empirical studies [14–16]. The flattening of spectral slopes during sensory experiences, suggesting elevated E/I ratios [15, 19], provides mesoscale physiological evidence supporting E/I imbalance models of hallucination generation [46]. Notably, while gamma oscillations have been causally implicated in iES-induced visual sensations [4], our exclusion of high-frequency bands (>45 Hz) to avoid contamination by stimulation artifacts [30–32] suggests that gamma-mediated perceptual phenomena may involve mechanisms distinct from those underlying aperiodic slope alterations. This dissociation highlights the importance of considering multiple, potentially independent neural signatures of sensation. Regions exhibiting flatter spectral slopes during visual sensations showed remarkable overlap with structural abnormalities identified in Parkinson's disease patients with visual hallucinations [47]. Crucially, these same regions demonstrated atypical functional integration specifically during hallucinatory experiences, suggesting they may constitute a common neural substrate for pathological and stimulation-induced perceptual phenomena.

While we noted broader cortical and network involvement in the lower frequency range (2–19 Hz) and more specialized contributions in the higher frequency range (20–45 Hz), a consistent flattening of aperiodic slopes during visual sensations was evident across both frequency bands. These similar aperiodic activity patterns suggest shared neural mechanisms involving E/I balance [14–16], potentially reflecting E/I ratio alterations implicated in hallucination-related neural circuit dysfunction [46]. This interpretation is further supported by evidence that E/I imbalance manifests as broad power spectrum changes [14], which may contribute to the neurophysiology of hallucinatory experiences [46]. Collectively, these findings indicate that although low- and high-frequency aperiodic activity have distinct characteristics, their common dependence on E/I balance provides a unifying mechanism for integrating neural dynamics across frequency bands in hallucination-related processes [46].

Compared with non-sensations, the finding of lower Φ^* during visual sensations indicated decreased neural integration between brain regions, consistent with a reduced level of consciousness according to integrated information theory [23]. At both the functional and structural levels, aberrant connectivity decreases in large-scale networks have been found in people with hallucinations compared to those without hallucinations [21, 22]. In particular, the disconnection hypothesis of schizophrenia [48] was recently linked to hallucinations [22, 49, 50]. Consistent with these previous studies, our finding of reduced global information integration indicated that network communication architecture is reconfigured during stimulation-induced sensations. Spectral slope flattening correlated with elevated baseline firing rates driven by greater local positive excitatory feedback [50]. The correlation between the aperiodic slope of IFC and Φ^* values during visual sensations might support the hypothesis that the “noisy” neural activity in the inferior frontal gyrus destabilizes network dynamics based on the dynamic network communication hypothesis [51, 52].

The iES modulates cortical excitability and influences visual awareness by directly activating precisely tuned neurons, particularly within the primary visual cortex. These topographically organized columnar neurons mediate spatially-selective excitation [53, 54]. The hierarchical architecture of these regions, characterized by robust feed-forward projections, amplifies iES-evoked activity and propagates it through successive visual processing streams [55], including MTComp and VenV. These visual areas become functionally disconnected from other sensory regions, particularly auditory areas, which may facilitate a "disconnected" visual experience. Simultaneously, they merge with the TPJ, a region implicated in awareness [41, 42] and attentional reorienting [43]. This network reconfiguration suggests that iES simultaneously isolates visual processing while enhancing access to consciousness systems [56]. The posterior STS serves as a critical multisensory integrator in this process, where its disrupted connectivity can lead to impaired differentiation between internally generated and external sensations, consistent with findings in auditory-verbal hallucinations [51] and agency disturbances [52].

The IFC emerges as a key regulatory node in this process. Its established roles in inhibitory control and attention allocation, combined with our findings of flattened power spectral slopes and altered sensorimotor connectivity, suggest compromised top-down regulation of sensory processing. While the IFC has been proposed as an "ignition" center for global workspace formation [57], our data challenge this interpretation. The lack of direct IFC-visual integration during sensations contradicts the classical ignition models. Instead, we find that IFC hyperactivation may propagate maladaptive activity patterns without proper sensory integration. Rather, our data support an alternative formulation wherein excessive IFC activation facilitates the dissemination of dysfunctional neural activity in the absence of appropriate sensory assimilation. These findings establish the IFC's dual clinical relevance, serving not only as a diagnostic indicator of pathological conditions but also as a potential therapeutic focus for neuromodulatory interventions targeting hallucinatory phenomena.

Notably, patients may realize that the visual responses elicited by iES may not appear in the environment [58]. These phenomena, although more diverse, may be similar to the minor hallucinations with completely preserved insight in Parkinson’s and Lewy body diseases [59]. Our investigation of neural mechanisms of elicited sensations may provide guidance for future studies to identify the dysfunction of neural circuits leading to more severe and persistent forms of hallucination in psychotic disorders.

Although our study provides potential neural mechanisms underlying visual sensations, there are caveats related to the interpretation of our findings. First, the location of electrodes was clinically determined, and the sparse sampling cannot guarantee fine coverage across the whole brain [60]. Thus, it is possible that other critical regions, outside our electrode coverage, contributed to eliciting visual sensations. Second, the contents of the visual sensations were diverse, from phosphenes, shapes, and colors to complex human figures or faces, animals, and scenes [1, 12]. Evidence from psychotic disorders indicates that the relative frequency of simple and complex hallucinations differs across disorders, suggesting that simple and complex sensations may result from different neural mechanisms [61]. However, previous functional mapping studies have revealed similar mechanisms by comparing simple to complex visual sensations [1, 12]. Here, relatively few elicited complex visual sensations, combined with the computational requirement of a threshold number of electrodes within each area, prevented us from calculating all measurements for complex visual sensations only. Third, the lack of behavioral metrics, particularly subjective ratings of sensation clarity, complexity, and duration, constrains our interpretation of how the recorded neural network alterations relate to the qualitative aspects of visual experience. Future investigations are needed to integrate standardized behavioral assessments with neural recordings to systematically examine the correspondence between network dynamics and subjective perceptual characteristics. Finally, although patients were instructed to report subjective experiences after both visual sensation and control trials, the potential influence of the motivation to communicate with the clinician may introduce confounding effects. Future research will focus on distinguishing these effects through the implementation of more sophisticated experimental designs.

Conclusions

Altogether, we combined iES with iEEG to reveal the electrophysiological signatures of aperiodic components and network reorganization during visual sensations. We found a flatter power spectral slope, along with disrupted network organization and integrated structures, when patients perceived visual sensations during electrical stimulation. Elucidating these electrophysiological mechanisms enables us to control them and use them, potentially developing better prosthetic devices that restore vision by direct brain stimulation. It also provides guidance for future studies to identify the dysfunction of network communication and integrated structures associated with hallucinations in mental disorders. We propose the IFC as a potential target area for hallucination interventions that might lead to greater improvements in clinical outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Conflict of interest

The authors declare that they have no conflict of interest.