Brain Encoding of Naturalistic, Continuous, and Unpredictable Tactile Events

Studies employing EEG to measure somatosensory responses have been typically optimized to compute event-related potentials in response to discrete events. However, tactile interactions involve continuous processing of nonstationary inputs that change in location, duration, and intensity. To fill this gap, this study aims to demonstrate the possibility of measuring the neural tracking of continuous and unpredictable tactile information. Twenty-seven young adults (females, 15) were continuously and passively stimulated with a random series of gentle brushes on single fingers of each hand, which were covered from view. Thus, tactile stimulations were unique for each participant and stimulated fingers. An encoding model measured the degree of synchronization between brain activity and continuous tactile input, generating a temporal response function (TRF). Brain topographies associated with the encoding of each finger stimulation showed a contralateral response at central sensors starting at 50 ms and peaking at ∼140 ms of lag, followed by a bilateral response at ∼240 ms. A series of analyses highlighted that reliable tactile TRF emerged after just 3 min of stimulation. Strikingly, topographical patterns of the TRF allowed discriminating digit lateralization across hands and digit representation within each hand. Our results demonstrated for the first time the possibility of using EEG to measure the neural tracking of a naturalistic, continuous, and unpredictable stimulation in the somatosensory domain. Crucially, this approach allows the study of brain activity following individualized, idiosyncratic tactile events to the fingers.


Introduction
EEG is the most popular method to investigate neurophysiological correlates of tactile processing.The majority of studies employing this technique have been optimized for the computation of event-related potentials (ERPs), which require the recurrence of specific events a multitude of times (Fossataro et al., 2023;Galigani et al., 2021;Hogendoorn et al., 2015;Macerollo et al., 2018;Ronga et al., 2021;Ronga et al., 2021b).Following the average of electrical activity across repetitions, the emerging ERP estimates the brain response to the onset of discrete tactile events.ERPs have an excellent signal-to-noise ratio (SNR) and can be easily compared between different experimental conditions.However, in everyday life, touch is experienced differently (Serino & Haggard, 2010).
The body-environment interaction entails continuous processing of non-stationary inputs (time-variant) that change in space (skin portion) and quality (type of touch).
Novel methodological approaches have recently emerged, allowing the investigation of electrophysiological data acquired during continuous sensory stimulation (Crosse et al., 2016;Morales & Bowers, 2022).Neuronal populations can synchronize their activity (through alignment of the phase) to temporal profiles of a continuous input (Lakatos et al., 2019;Obleser & Kayser, 2019).This neural tracking has been measured with a variety of stimuli and tasks, including visual stimulation (Bourguignon et al., 2020;Hauswald et al., 2018;King et al., 2016;Spaak et al., 2014), speech processing (Di Liberto et al., 2015;Keitel et al., 2017;Kösem et al., 2018;Zoefel & VanRullen, 2016), music (Zuk et al., 2021) and multisensory inputs (Jessen et al., 2019;Thézé et al., 2020).In the case of the forward or encoding models, the mathematical framework underpinning the neural tracking aims to predict ongoing brain activity from continuous stimulus features (e.g., Crosse et al., 2016;Lakatos et al., 2019).These analysis approaches are increasingly popular in neuroscientific research (Crosse et al., 2016(Crosse et al., , 2021)); however, they are rarely employed in the somatosensory system (Chota et al., 2023;Maallo et al., 2022), and, notably, naturalistic tactile inputs have never been employed.The possibility of investigating the association between varying somatosensory inputs and ongoing brain activity could represent a leap toward unraveling how the brain represents sensory input.
Here, we demonstrated the possibility of measuring the neural tracking of continuous and unpredictable tactile stimulation of human hand fingers.Each participant was exposed to non-standardized continuous somatosensory stimulations to fingers.To extract the timevarying information of a gentle touch, we recorded sounds generated by tactile stimulations using a miniaturized microphone fixed on a brush.We then extracted the sound envelope, which represented the information concerning the friction between the participants' skin and the brush.In this case, the sound envelope provided information regarding the onset and offset of the tactile stimulation, its dynamic, and frequency.
Since encoding for different sensory stimuli highlighted responses that are neurophysiologically interpretable like ERPs (Crosse et al., 2016), the temporal response function (TRF) derived from the forward model should resemble the somatosensory ERPs in both timing and topographical distribution.This work aimed to extend and validate the neural tracking approach in the somatosensory domain, possibly opening a new branch of study to measure how the brain encodes naturalistic tactile events.To this aim, we (i) computed the TRF at the individual level based on the specific and unique tactile stimulations each subject received; (ii) estimated a group-level TRF to assess the solidity of the measurement across subjects, despite each of them being stimulated differently; (iii) tested the minimum amount of data to reconstruct a solid and robust average TRF associated with continuous and naturalistic tactile stimulation of the hands by comparing it with noise; (iiii) ensured that we measured a pure tactile TRF, and that spurious visual and auditory input could not explain the observed effects.
Results clearly revealed a successful measurement of how the brain encodes naturalistic, continuous, and unpredictable tactile events.

Participants
Since, to the best of our knowledge, this is the first study assessing neural tracking of naturalistic tactile stimulation, we estimated the sample size, allowing us to estimate reliable neural tracking as compared to noise.We focused our analysis on a central contralateral cluster of electrodes (Channel [22 25 26 27 28] for right-sided stimulation and channel [42 45 46 48 49] for left stimulation), in line with the extensive literature regarding somatosensory processing (Fossataro et al., 2023;Hogendoorn et al., 2015;Macerollo et al., 2018;Pyasik et al., 2021;Ronga et al., 2021); the frequency of interest [0.5-6Hz] was defined after the inspection of a spectral power plot (see Figure 1); the time-window for the data simulation was [0.5-200 ms] which comprises the main somatosensory responses starting from the earliest component.We ran a power analysis simulating the main contrast of interest (neural activity measured in experimental conditions vs. control condition; see 2.2) by using a cluster-based permutation test and the data of 10 participants (Wang & Zhang, 2021).The sample size estimation resulted in 27 participants.Thus, we recruited and tested in this study a total of 27 young-adult volunteers (female = 15; mean ± SE: 27.92 ± 0.36 years, range 25-32).They all had normal or corrected to normal vision and were right-handed by self-report; none reported a history of neurological disorders.Participants were naïve to the purpose of the experiment.All participants were informed about the experimental procedure and signed a written informed consent before testing.The regional ethical committee approved the study.The study protocol adhered to the guidelines of the Declaration of Helsinki (World Medical Association, 2013).

Procedure
The participants underwent a (i) preliminary behavioural and an (ii) EEG experiment in a single session.The behavioural experiment always preceded the EEG.For both experiments, participants were seated in a comfortable chair and kept both hands on a table.The experimenter was seated in front of the participant.Crucially, a box covered both participants' and experimenter's hands from view (see Figure 1).Experiments were conducted in a soundproof chamber (BOXY, B-Beng s.r.l., Italy).
The preliminary behavioural experiment aimed to evaluate if the subject could perceive auditory information related to the friction between the brush and the skin.The experimenter performed 40 trials: 20 in which he stimulated his own left index finger with a single brush and 20 without stimulation.The order of the trial was fully randomized for each participant.The instructions for the subject were to guess if there was or not a stimulation.The two kinds of trials (i.e., stimulation vs. no stimulation) and the two possible responses (i.e., yes vs. no) derived four possible results: hits (i.e., stimulation and yes), omissions (i.e., stimulation and no), false alarm (i.e., no stimulation and yes) and correct rejection (i.e., no stimulation and no).The whole preliminary experiment lasted for 4 minutes.
During the EEG experiment, participants kept the same posture as in the control one, with their hands hidden from view and placed comfortably on the table.Participants were instructed to look at a fixation cross positioned at the centre-top of the box, equidistant from their hands, and to pay attention to the tactile stimulation.The main experiment had five conditions: right thumb, right pinkie, left thumb, left pinkie, and a control condition.
In the control condition, everything was kept constant except the target of the tactile stimulation.The experimenter stimulated his own left-hand thumb to rule out the possibility that auditory input (or any other visual cue) associated with brushing participants' fingers during real tactile stimulation could contribute to neural entrainment.
Only the forearms of the experimenter were hidden; thus, it is possible that during brushing, participants could entrain with other movements (shoulder, head, etc.) of the experimenter.Three trials for each condition were performed, resulting in 15 experimental trials lasting three minutes each.During each trial, the experimenter stimulated the specific finger through an ad-hoc built brush, recording the friction between the brush and the skin.Crucially, the stimulation for each trial and condition was unique (random) for each participant.The total experimental duration was 45 minutes.
Participants were encouraged to take small pauses between each trial, during which they moved their hands.The order of experimental conditions was controlled and randomized by PsychoPy3® (v2020.1.3).An external sound was used to code the beginning of each trial and ensure precise tactile stimulation/EEG synchronization.

Behavioural Data Analysis
The behavioural data collected in the preliminary experiment: Stimulation (present vs. absent) and response (yes vs. no).This resulted in a percentage ratio of hit (yes/present), missing (no/present), false alarm (yes/absent), and correct rejection (no/absent).A withinsubject paired t-test was performed to contrast hit and missing.Finally, the d', a measure to assess the sensitivity of the participants' responses, was computed by subtracting the z-score of the false alarm rate from the z-score of the hit rate.

EEG recordings and preprocessing
The EEG recordings were acquired at a sampling rate of 500 Hz using NetStation5 software and a Net Amps 400 EGI amplifier connected to 64 electrodes HydroCel Geodesic Sensor Net (Electrical Geodesics, Inc.), all signals referenced to vertex (additional channel E65/Cz).Electrode impedances were kept below 30 kΩ.
We preprocessed continuous EEG raw data offline using MATLAB (R2020b, Mathworks Inc., Natick, MA) and EEGLAB toolbox (version 264 14.1.2b,Delorme & Makeig, 2004).The preprocessing of the raw data followed these steps: (i) band-pass filtered from 1 to 40 Hz (low-pass: FIR filter, filter order: 100, window type: Hann; high-pass: FIR filter, filter order: 500, window type: Hann); (ii) downsampled to 250 Hz; (iii) Independent Component Analysis (ICA; Jung et al., 2000) to remove artefactual components related to blinks, eye movements and heartbeat.To semi-automatize this process in a data-driven manner.The ICA weights were applied to the raw data, and the components were first labelled using ICLabel (Pion-Tonachini et al., 2019).Hence, the most prototypic component for each artefact was selected.It resulted in a total of three prototypic components for eye blink, eye movements, and heartbeat, respectively, chosen by the highest probability value given by ICLabel.These three prototypic components were used as templates for the automatic identification of similar components with CORRMAP (Version 1.03; Campos Viola et al., 2009).CORRMAP automatically detected all components with a correlation coefficient exceeding a threshold of 0.8 with the prototypic ones.These identified components were then removed from the original raw data (removed components per participant mean±SE: 3,04±0,22).(iv) Noisy channel detection and spherical interpolation (Interpolated channel mean±SE: 2,37±0,22); (v) rereferenced to standard average reference; (vi) band-pass filtered from 0.5 to 6 Hz (lowpass: FIR filter, filter order: 200, window type: Hann; high-pass: FIR filter, filter order: 2000, window type: Hann); (vii) downsampled to 64Hz; (viii) data epoching according to the onset of the tactile stimulation, considering a delay between the auditory event that coded the beginning of the stimulation and its corresponding EEG marker (i.e., 89 ms), and (ix) segmented into 1 minute long trials, resulting in 9 trials per condition (N=5) per subject (N=27).The stimulation and EEG recording occurred for three minutes in each trial, so the 1-minute segmentation did not aggregate non-consecutive events (Crosse et al., 2021).Lastly, (x) EEG data corresponding to a 1-minute segment were normalized, computing the z-scores to optimize the cross-validation procedure.

Tactile Envelope Extraction
The sound corresponding to the tactile stimulations was registered online through a custom-built brush with a miniaturized condensation microphone (RØDE, Lavalier GO).
For each condition, sounds of the tactile stimulations were first concatenated, leading to 12 min of recording and segmented into one-minute trials, obtaining 12 trials for each condition and subject (N=27).We extracted the tactile envelope for each trial by computing (i) the absolute value of the Hilbert transform of the stimulations, (ii) applying a low-pass filter (3rd-order Butterworth filter with 6 Hz as cut-off frequency), (iii) downsampling the signal to 64 Hz (i.e., matching the downsampling of the EEG data, Mirkovic et al., 2015) and (iv) normalizing them by dividing for the maximum value.

Temporal response function estimation
The temporal response function estimation was performed with the mTRF toolbox (Crosse et al., 2016).This TRF approach allows the building of an encoding model to predict an unknown EEG response from the stimuli features (i.e., the auditory envelope of touch).This approach has already been employed to evaluate the neural tracking of speech sound, linguistic, and visual properties (King et al., 2016;Plass et al., 2020;Thézé et al., 2020).This mathematical function exploits ridge regression.The TRF is a mathematical filter that describes the linear relation of the brain response to the stimuli features in a pre-specified number of time points associated with the stimuli occurrence (Crosse et al., 2016).The ridge regression is used to avoid the problem of overfitting the model (i.e., building a model that perfectly predicts the data used in the training set but fails to predict unseen data).This function relies on the ridge parameter, which prevents overfitting by penalizing the model weights to keep the model as generalizable as possible.To avoid model overfitting, the optimal regularization parameter was empirically identified, employing a leave-one-out cross-validation procedure, choosing a predetermined selection of ridge values (i.e., from 10 -6 to 10 8 ) for time lags from -400 to 600 ms.To choose the regularization parameter, the mean squared error (MSE) between the real and the predicted EEG response for each participant and condition was evaluated, and critically, the parameter yielding the lowest MSE on the majority of test datasets was selected.The optimal parameter (λ=10^3) was selected and kept constant across participants and experimental conditions to allow a correct generalization of the model at the group level.We computed independent encoding models for each EEG channel (N=64) in each participant for each condition.We employed the data in a window starting from -400 to 600 ms.The tactile envelope was extracted below 6 Hz since these frequencies were the most represented in the tactile stimulation (see Figure 1).The encoding models were trained to employ the tactile envelope as the predictor, using a leave-one-out cross-validation procedure (Crosse et al., 2016).Critically, the TRF model for each participant is the average of each model computed for every left-out trial.Finally, the grand average TRF models were computed by averaging TRFs across participants within conditions.The main advantage of the TRF model is that encoding model weights are physiologically interpretable (Crosse et al., 2016;Haufe et al., 2014).These weights, computed for each EEG channel, allowed us to explore the topography of the TRFs (similar to the ERP response) regarding both the amplitude and directionality of the data.

Statistical analysis
First, for each condition, we evaluated the existence of somatosensory entrainment against the noise of the data.The spatiotemporal profiles of each experimental condition TRF (tactile TRF) were contrasted to a null effect (null TRF) obtained by cross-validating models iteratively (100 iterations) fitting TRFs on randomly mismatched pairings of somatosensory envelopes and EEG responses and evaluating the models on original data.This method was done separately for each participant and then grand averaged to obtain a null TRF model for each condition, which was employed as a baseline in the analysis.
Statistical comparisons were conducted employing paired t-test with a cluster-based permutation approach (Maris & Oostenveld, 2007, Monte Carlo method 1000 permutations, the cluster-based p-value was 0.025, accounting for two-tailed testing) across 0 to 400 ms time-lags and all electrodes.Cluster-based permutation statistics allowed us to control for the multiple comparisons problem associated with testing across time lags and sensors.The summed t-statistics associated with the comparison of interest are clustered in time and space and contrasted versus the summed t-statistics associated with the comparison following a random partition obtained by shuffling the experimental condition labels (i.e., experimental TRF vs null TRF).
Subsequently, to test the robustness of the somatosensory entrainment, we contrasted the spatiotemporal profiles of tactile TRF (i.e., real touch) against the control condition (control TRF).Namely, we contrasted each experimental condition (i.e., left thumb, left pinkie, right thumb, and right pinkie) versus the control condition in which the experimenter's hand was stimulated (see Method, 2.2 procedure).These contrasts allowed us to firmly exclude the possibility that spurious non-somatosensory cues (e.g., auditory and visual information) could explain the tactile TRFs measured in each experimental condition.As previously reported, in the preliminary behavioral experiment, we measured participants' ability to discriminate by sounds, whether the experimenter was touching his own hand with the brush.However, we could not exclude the possibility that sounds below the threshold could be informative.Similarly, while participants could not see their hands, nor the experimenters', we ensured that no other visual cue could lead to spurious entrainment in synchrony with the tactile one.To this end, cluster-based permutations were performed between tactile TRFs at each condition and the control condition control TRF (Maris & Oostenveld, 2007, Monte Carlo method 1000 permutations, the cluster-based p-value was 0.025, accounting for two-tailed testing) across 0 to 400 ms time-lags and all electrodes.
Lastly, we evaluated the consistency of the somatosensory entrainment, investigating the minimum tactile stimulation time necessary to obtain a reliable tactile TRF greater than the null TRF.The spatiotemporal profiles of each experimental condition (tactile TRF) were contrasted to a null effect (null TRF).Crucially, we built eight different TRF models starting from 2 and reaching 9 minutes of stimulation.After evaluating the normality of the distribution, a series of paired t-tests were performed between each model tactile TRF and the corresponding null TRF, using as critical alpha 0.006 (Bonferroni corrected, eight multiple tests).

Behavioral data analysis
The behavioral data collected in the preliminary experiment resulted in a percentage of hit and omission that entered a paired t-test crucially hit, and omission did not differ (mean ± SE: Hit 0.51 ± 0.01; Omission 0.49 ± 0.01; p-value = 0.70).Subsequently, the d', a measure of sensitivity, was computed to assess if participants were able to distinguish trials in which the stimulation was performed compared to when it was not performed, considering that a d' near to 0 indicates the absence of sensitivity (d' = 1.6494E-15).

EEG data analysis
Two different sets of cluster-based permutated paired t-tests were performed.A preliminary analysis contrasted each tactile neural tracking versus the correspondent null effect to ensure a reliable statistical existence of the tactile TRF.A secondary analysis contrasting each experimental condition TRF versus the control condition, in which the experimenter was stimulating their own hand hidden from participants' view.This further control analysis was performed to prove that tactile TRF could not be explained by spurious auditory or visual information.
Tactile entrainment vs Null.As a first analysis, within each experimental condition (i.e., left thumb, left pinkie, right thumb, and right pinkie), the tactile TRF was contrasted with a corresponding null TRF at the whole brain level to capture the clusters of significant electrodes within the 0-400 ms time-lags.A clear tactile TRF emerged in each experimental condition, and this response was statistically different from the null TRF.Tactile entrainment vs control.Subsequently, the tactile TRF of each condition was contrasted with the control TRF (of the control condition).A clear tactile response function was significant when controlling for auditory and visual spurious activity.Each experimental condition differed significantly from the control condition TRF, revealing one positive response at the central contralateral cluster (p-values< critical alpha: 0.025) and a second negative response (p-values< critical alpha: 0.025; See Figure 2).Critically, also with this contrast, two defined peaks emerged.Their statistical significance, topographical, and temporal distribution were strictly similar to the previous one, with a central contralateral positive response (50 -170 ms) peaking around 140 ms and a central bilateral negativity (200-300 ms) peaking around 245 ms (see Figure 3).Minimum data to measure tactile entrainment.Eventually, we contrasted tactile TRF and null TRF estimated by eight different models obtained using different lengths of tactile stimulation (2-9 minutes).Crucially, the contrast between tactile TRF and null-TRF was significant for each experimental condition, starting from 3 minutes of tactile stimulation (p-values< critical alpha: 0.0065).

Discussion
Neural tracking is an essential mechanism subserving an efficient interaction with the external world.It was reliably measured in the visual (Bourguignon et al., 2020;Hauswald et al., 2018;King et al., 2016;Spaak et al., 2014) and auditory (Di Liberto et al., 2015;Keitel et al., 2017;Kösem et al., 2018;Zoefel & VanRullen, 2016;Zuk et al., 2021) systems following continuous stimulation.Here, we extended the measurement of neural tracking to the somatosensory domain, employing a continuous and naturalistic tactile stimulation of the hands.Notably, each participant received unique and random stimulation with a series of gentle brushes.We modeled the sound generated by tactile stimulation, namely the recording of the friction between the skin and the miniaturized microphone inserted in a brush.In defining the tactile envelope, the frequency range with the greater energy (i.e., 0-6 Hz, see Figure 1) was chosen to have a comprehensive descriptor of the data with maximal information.
First, we estimated tactile neural tracking at the individual level employing a temporal response function (TRF).Critically, this is done with a leave-one-out cross-validation, validating each model at the single-subject level, considering the intrinsic variability of each participant and each one of the stimulations performed.Second, we estimated a group-level TRF and tested its existence by comparing it with a null TRF obtained by the random permutation of the data.Third, we calculated the minimum amount of data to estimate a robust group level TRF associated with continuous somatosensory stimulation for each finger; three minutes of passive tactile stimulation were sufficient to obtain a group level TRF significantly different from the null TRF, calculated with the same amount of data.Finally, contrasting the TRF obtained by the stimulation of single fingers of the hand vs. a control TRF, in which we controlled for possible auditory and visual confounds, a clear tactile response function emerged, confirming the measurement of pure tactile entrainment with continuous tactile stimulation.
Only a few preliminary observations employed continuous tactile stimulation while recording the EEG (Chota et al., 2023;Fu & Riecke, 2023).Most of the studies investigating the somatosensory system still rely on somatosensory evoked potentials (SEPs) and averaged responses across several trials (Fossataro et al., 2023a;Hogendoorn et al., 2015;Macerollo et al., 2018;Pyasik et al., 2021;Ronga et al., 2021).Among the few studies that investigated neural responses to continuous tactile input, Chota and colleagues investigated the role of phase-locked β bursts following a vibrotactile artificial stimulation determined by the somatosensory system resonance properties.Crucially, when the somatosensory input was mapped continuously (Chota et al., 2023;Fu & Riecke, 2023), these studies measured steady-state somatosensory evoked potentials (SSSEPs) that required a highly standardized series of events with a specific predefined frequency of stimulation.Here, we measured how the brain encodes dynamic somatosensory events, as in tactile interactions.Despite employing unique and different stimulations for each condition and participant, we obtained a robust and reliable tactile TRF.This result was possible because the tactile TRF was estimated for everyone by modeling the envelope of the tactile information describing the time series and the dynamics of tactile events.Critically, each stimulation differed randomly in several parameters, like the onset/offset of each single brush, the duration and intensity of each brush, and the locus of stimulation on the finger.In other words, the variability of the stimulation was modeled and used to gain generalizability of findings.
To assess the reliability of this protocol, it was essential to evaluate the consistency in terms of timing and topography with previous studies employing somatosensory stimuli.
Noteworthy is that by measuring the encoding of finger brushing, we modelled slow tactile events (sound envelope below 6 Hz) on slow rhythmic brain activity (below 6 Hz).This prevented us from measuring faster tactile response functions.Classical clinical SEP studies highlight early components through the stimulation of the median nerve, leading to a complex chain of early responses (i.e., P14, N20, P22, N24, and N30) starting before 30 ms after the stimulus onset (Valeriani et al., 2001;Vibell et al., 2023).However, these early subcortical components are visible following the average across several hundred trials (Vibell et al., 2023) and are not measurable following skin stimulations (Zeller et al., 2015).Conversely, the cortical SEPs generally include the N25, P45, N80, P1/P100, N1/N140, P2/P200, N2 (about 250 ms after touch onset), and P3 components, with early contralateral topography followed by later bilateral responses (Cardini et al., 2012;Nuwer, 1998;Sambo & Forster, 2008).Our results are in line with well-established cortical SEPs and highlighted in the TRF a clear first contralateral positive response at central electrodes, which started at about 50 ms of lag; this response was compatible with early somatosensory processing components highlighted in ERP studies (Shimojo et al., 2000;Zeller et al., 2015).This response peaked at around 160 ms and was followed by a central bilateral negativity peaking at about 260 ms.Both the timing and topography of these responses align with the SEPs literature (Eimer & Forster, 2003;Sambo & Forster, 2008;Shimojo et al., 2000;Zeller et al., 2015;Vibell et al., 2023), suggesting that the method unveiled consistent neural tracking of tactile stimulation.
It is important to stress that the encoding model of the TRF was statistically tested with a data-driven approach, controlling for the multiple comparisons problem and contrasting TRFs with null effects (Maris & Oostenveld, 2007).In canonical ERP research, the brain responses are usually averaged and contrasted between experimental conditions.Here, the approach provides control by variability since the entire tactile information is modeled, and the response of interest can be objectively identified in time (time lags) and space (scalp) domains, strictly when it exceeds the estimation of noise or control conditions, such as the null TRF and the control TRF.In addition to its effectiveness, this paradigm strength also relies on the limited amount of time required to obtain a reliable tactile response function.Indeed, despite idiosyncratic stimulations (every trial was different for every finger and participant), we measured a significant and stable TRF already with three minutes of stimulation.This indication can drastically shorten the experimental paradigms employing tactile stimulation, allowing the contrast of several experimental conditions within a relatively short experimental session, especially in populations such as infants (Ronga et al., 2021), people with autistic spectrum disorders (Voos et al., 2013) or clinical and psychiatric population (Keizer et al., 2022).
Results clearly demonstrated a pure tactile TRF.In addition to the contrast with the null effect, the tactile TRF was contrasted with the control TRF.The control condition was identical to the experimental one regarding auditory and visual cues, such as the noise associated with brushing or the experimenter's movements.This allowed for control of potential auditory and visual inputs collinear to the tactile stimulation.However, since the experimenter stimulated his own hands, the participants did not have any tactile stimulation.Therefore, contrasting the tactile TRF with the control TRF ensured estimating actual tactile entrainment, avoiding spurious effects associated with other input processing.
This study opens new venues for future work employing similar analytical approaches to evaluate different populations as well as experimental manipulations of somatosensory processing.With an ontogenetic perspective, it will be essential to investigate how the tactile entrainment emerges by describing its developmental trajectory, opening the possibility of using this measure as a marker of typical and atypical development.Studies in the animal domain could provide the phylogenesis of this tactile response function.
This method could additionally help characterize and further distinguish the discriminative touch from the affective touch (McGlone et al., 2014;Singh et al., 2014) or tickle, a specific somatosensory experience that has intrinsic value as a mediator of the social aspect of touch both in the phylogenetic and ontogenetic development, employing and evaluating online and real-life tactile interactions among individuals.Due to its intrinsic properties, it is not easy to assess how the human brain elaborates affective touch (i.e., a continuous tactile stimulation with a preferential velocity of around 1-10 cm/s that activates the CT-fibers).Indeed, it is impossible to reproduce this affective component of touch with highly standardized laboratory stimuli, while it is possible to employ a naturalistic stimulation.Currently, most EEG studies investigate the early and late components of late positive potential (LPP) or modulations of brain activity in specific frequency bands (McGlone et al., 2014;Portnova et al., 2020;Schirmer & Gunter, 2017;Singh et al., 2014;von Mohr et al., 2018).The tactile TRF should be able to highlight and characterize when and how the affective touch is processed in the brain and its difference compared to the discriminative touch simply by changing stimulation properties.Finally, the current approach could be used to stimulate different body parts and evaluate eventual differences in neural tracking associated with different body districts.Our results pave the way for a more efficient and naturalistic study of the somatosensory system, evaluating how it works in situations more affine at how the brain elaborates tactile information in real-time daily interactions.

Fig. 1 (
Fig. 1 (A) Experimental set-up.The participant's hands are kept comfortably on the table, covered from view by a wooden box.With an ad-hoc built brush, the experimenter stimulated the participant's fingers (e.g., the right thumb).(B) Data processing and performed model.(C) Power spectrum (0.5 -8 Hz) of the audio envelope of touch for each experimental condition.(D) tactile TRF results of right thumb stimulation.The grand average TRF (N=27) is depicted in purple, with single participants in light grey.The topographies show the spatial distribution of the grand averaged TRFs of the first positive (~140ms) and negative (~245ms) responses in two 45 ms time windows centred on each peak.

For
each experimental condition, one positive cluster (p-values< critical alpha: 0.025) and one negative cluster (p-values< critical alpha: 0.025) were identified on the contralateral central electrodes.Specifically, our results highlighted a central and contralateral positive response (50 -170 ms) peaking around 140 ms and a central bilateral negativity (200-300 ms) peaking around 245 ms (see Figure2).

Fig. 2 .
Fig. 2. Tactile TRFs vs null TRFs.Upper panel: the TRF corresponding to each experimental condition (colored lines) and the null-TRF (in light grey).The null TRF is obtained by fitting the model on randomly mismatched pairs of stimulations (sound envelope of touch) and EEG responses.The TRFs and the null-TRFs here represent the average across central contralateral electrodes with respect to the finger stimulation.For all finger stimulations, the topographies reveal a first positive central contralateral activity (~140ms) followed by a second negative bilateral response (~245ms).Lower panel: results of the cluster-based permutation tests (critical alpha 0.025) between the experimental TRF and null-TRF, across 0 to 400 ms time-lags and all sensors.For each experimental condition, a first positive response (p-values< critical alpha: 0.025) and a second negative response (p-values< critical alpha: 0.025) emerged.

Fig. 3 .
Fig. 3. Tactile TRFs vs Control TRFs.Upper panel: the TRF corresponding to each experimental condition (colored lines) and the control TRF (in black).The control TRF is obtained by keeping visual and auditory spurious activity constant but without the somatosensory component of the stimulation (i.e., the experimenter's left thumb is stimulated and not the participant's).The tactile TRFs and the control-TRFs here represent the average across central contralateral electrodes with respect to finger stimulation.The topographies reveal a first positive contralateral peak (~140ms) followed by a second negative bilateral response (~245ms).Lower panel: results of the clusterbased permutation paired t-tests (critical alpha 0.025) between the experimental TRF and control TRF across 0 to 400 ms time-lags and all sensors.For each experimental condition, a positive response (p-values< critical alpha: 0.025) and a second negative response (p-values< critical alpha: 0.025) emerged.