Characterizing the Heartbeat‐Evoked Potential: A Two‐Component Model of Cardiac Signal Processing?

ABSTRACT

The heartbeat evoked potential (HEP) is a potential marker of cardiac signal integration at the cerebral level, obtained by averaging epochs time‐locked to the ECG R‐peaks. The HEP is modulated across different experimental conditions with amplitude differences arising between 200 and 600 ms post‐R‐peak over fronto‐central sites. However, substantial heterogeneity exists, and to date there is no clear characterization of the HEP. Here, we propose a two‐component model of the HEP consisting of first, an early (100–250 ms) component presumed to index primary cardiac signal integration, thus being task independent; and second a late (250–500 ms) component thought to index elaborative processes, supposed to be modulated both within and between tasks. We aimed to first delineate these components together with their frequency characteristics at rest before exploring their modulation during an emotion task and a tactile stimulation protocol, using independent datasets for exploratory and reproducibility purposes, totaling 104 participants from different cultures. Our results revealed an early (100–250 ms) fronto‐central negativity potentially associated with theta phase resetting, followed by a posterior positivity (250–500 ms). As expected, we did not observe any intra or inter‐task modulation of the early component. However, contrary to our hypothesis, the late component was not modulated by task neither. This lack of task‐related modulation in the late component contrasts with previous literature but appears robust given that our study design used multiple datasets, participants and experimental protocols. Our findings highlight the need for standardized methodologies in HEP research to improve reproducibility and enhance our understanding of cardiac‐related neural processing.

IMPACT STATEMENT

We propose a novel two‐component model for the heartbeat evoked potential (HEP) and test it using a robust methodological design comprising different tasks and both exploratory and reproducibility datasets. Our findings support the existence of distinct early and late HEP components, although their precise functional significance remains to be determined.

1. Introduction

The brain continuously processes and integrates signals arising from the body and sends back regulating signals. This bilateral communication is fundamental for human functioning. It allows allostatic processes that maintain the bodily state within the homeostatic range, but it is also implicated in higher cognitive functions. The central monitoring of the body's internal physiological variables, or interoception, is now extensively studied. Multiple studies emphasized the role of interoceptive signals in bodily self‐consciousness, exteroceptive perception, emotion representation, and decision‐making (Azzalini et al. 2019).

Cardiac interoception has been more thoroughly studied than other interoceptive systems, such as respiratory or gastric. It is a multidimensional construct that can be studied at different levels (Suksasilp and Garfinkel 2022). It is determined at the behavioral level with interoceptive accuracy, a declarative measure of interoception. It can also be measured at the cerebral, implicit level using the Heartbeat Evoked Potential (Schandry et al. 1986) (HEP). The HEP consists of a transient neural response locked to heartbeats. Typically, it is calculated as an Event‐Related Potential (ERP) time‐locked to the electrocardiogram (ECG) R‐peak. The HEP can be contaminated by the Cardiac Field Artifact (CFA) which corresponds to the diffusion of the electrical signal from the heart to the scalp, and which is by nature time‐locked to the heartbeat; this artifact needs to be accounted for in the preprocessing pipeline. Several peripheral pathways appear to contribute to the integration of the cardiac signal. Receptors that transmit the cardiac signal are mostly found in the aorta and the carotid, but others are also found in the wall of the heart and the skin. At the central level, the primary sites of cardiac signal integration are the bilateral insula, with the posterior insula considered as the primary interoceptive cortex, although the information transmitted by somatosensory receptors is processed in the somatosensory cortex (for review Azzalini et al. 2019).

The HEP is modulated across different experimental paradigms, with amplitude differences generally observed over fronto‐central sites between 200 and 600 ms post R‐peak (see meta‐analysis : Coll et al. 2021). Studies for instance investigated HEP modulations in heartbeat attention paradigms where attention is directed either towards cardiac interoceptive signals or external stimuli (Mai et al. 2018; Petzschner et al. 2019; Salamone et al. 2018; Terhaar et al. 2012), or during emotional paradigms where participants' arousal is manipulated through exposure to emotional versus neutral stimuli (Couto et al. 2015; Fukushima et al. 2011; Gentsch et al. 2019; Ito et al. 2019; Luft and Bhattacharya 2015; Marshall et al. 2018). An increase in amplitude is often reported in the condition of interest (i.e., interoception or emotion) compared to the control condition (but see Terhaar et al. 2012 for null results). However, the morphology of the HEP waveform and the latency of effects greatly differ between studies. Three types of HEP waveforms are observed across studies: (1) ECG‐like waveforms which show ECG‐like QRST complex characteristics (Petzschner et al. 2019; García‐Cordero et al. 2017; Kim et al. 2019; Perogamvros et al. 2019; Schulz et al. 2015); (2) early peak waveforms showing an often negative peak in an early time window (150–350 ms) (Salamone et al. 2018; Fukushima et al. 2011; Gentsch et al. 2019; Ito et al. 2019; Judah et al. 2018) (see Mai et al. 2018 for positive peak); (3) flat waveforms showing no characteristic deflections (Babo‐Rebelo, Richter, and Tallon‐Baudry 2016; Babo‐Rebelo, Wolpert, et al. 2016).

Despite the wide use of the HEP as a marker of the cerebral processing of cardiac signals, very few studies aimed to characterize it (but see Coll et al. 2021; Park et al. 2018; Kern et al. 2013). Some studies described it as a positive deflection over anterior sites from 250 to 600 ms (Montoya et al. 1993), others as a biphasic potential consisting of a negative deflection at around 200–280 ms followed by a positive deflection at around 350 ms (Immanuel et al. 2014) (or with inverted polarity in an intracranial study (Kern et al. 2013)). Additionally, Park et al. (2018) reported an increased Inter‐Trial Phase Consistency (ITPC) peaking in the theta band (4 Hz) between 100 and 250 ms following the R‐peak. Notably, the increase in ITPC (Park et al. 2018) occurred in the absence of an increase in power, suggesting that the HEP is generated by phase resetting of the theta rhythm (Sauseng et al. 2007). Furthermore, Kim and Jeong (2019) observed a selective increase in theta phase connectivity following heartbeats at rest, emphasizing the role of the theta rhythm in cardiac information processing (Kern et al. 2013). Yet, despite these few studies, the HEP, its components and how they are modulated by different experimental settings, are still not well understood and characterized. A clear characterization of the HEP would have theoretical implications by enhancing our understanding of cardiac signal cerebral processing and integration, but also help resolve discrepancies in the HEP literature by providing more information for future studies. Hence, we aimed to provide a comprehensive description of the HEP using the same analysis strategies in different datasets, manipulating experimental conditions both across and within subjects.

Two models could explain the HEP and its modulations. First, it can be hypothesized that the HEP is a unitary response to heartbeats (i.e., it consists of a single component). In this model, the HEP may be generated by theta phase resetting (Park et al. 2018), leading to the biphasic component observed by Kern et al. (2013) and Immanuel et al. (2014). Under this (henceforth “single‐component”) model, the HEP amplitude is directly modulated by experimental conditions, and its topography is task‐ and time‐invariant. However, the discrepancy between the early modulation of ITPC (Park et al. 2018) and the later modulation reported in ERP studies (Coll et al. 2021) led us to propose a “two‐component model” of the HEP. We hypothesized that both components would be observed during all tasks, irrespective of conditions. First, we expected to find an early component (100–250 ms), potentially generated by theta phase resetting (Park et al. 2018). This component is thought to reflect primary cardiac signal integration within interoceptive regions (i.e., insula, amygdala) and would remain invariant across tasks and conditions. To further characterize its frequency dynamics, we examined ITPC and power modulations in the theta range at rest, expecting a heartbeat‐induced ITPC increase without a concomitant power change (Park et al. 2018; Sauseng et al. 2007). Importantly, we validated the presence of theta activity in the signal (Donoghue et al. 2021). We then expected to observe a second component (250–500 ms), corresponding to a time window when task‐induced modulations are usually described in HEP studies (Coll et al. 2021; Montoya et al. 1993). This latter component is considered a nonphase‐locked evoked response, indexing the elaborative processing of cardiac information in a task‐dependent manner. This component likely reflects the integration of cardiac information within task‐specific networks, where the cardiac signal undergoes further processing depending on the task or situation at hand. We hypothesized that both its amplitude and topography would vary according to task demands in terms of interoceptive processing, reflecting the engagement of distinct neural networks. To test these hypotheses, we analyzed multiple datasets with different tasks and conditions. First, we characterized the HEP components during resting state, to uncover whether each component can be observed with no ongoing task. Then, we explored whether these components are modulated by different cognitive and sensory contexts. Specifically, we used an emotion dataset during which participants had to watch emotional or neutral videos, and a sensory dataset during which participants were stimulated with tactile stimuli either strokes by a social or nonsocial object or tapping by a mechanical device. Finally, we replicated our findings in an independent dataset where the same participants completed all three protocols (resting state, emotion, and tactile stimulation).

2. Methods

2.1. Participants, Material and Procedure

Our hypotheses were investigated across different datasets. For clarity, we categorized them into two groups: the “exploratory” datasets group and the “reproducibility” datasets group. Each dataset group comprised three datasets: resting state, emotion, and tactile stimulation datasets. The exploratory resting‐state and emotion datasets were obtained from open‐access repositories on OpenNeuro (https://openneuro.org/), while the exploratory tactile stimulation dataset was retrieved from a previous study conducted in our lab (Guidotti et al. 2023). Finally, the reproducibility dataset was acquired specifically for this study. Detailed participant characteristics are displayed in the Table 1.

TABLE 1.

Participant's characteristics in each dataset. The table includes the reference from which the data were retrieved, the sample size (N), the mean age (+ standard deviation) and the male‐to‐female ratio.

Dataset	References	Final N	Age mean (SD)	Sex ratio (M/F)
Exploratory dataset
Rest	Pavlov et al. (2022)	30	20.2 (3.34)	3/27
Emotion	Mishra et al. (2021)	30	23.23 (1.45)	28/2
Tactile stimulation	Guidotti et al. (2023)	21	22.95 (2.60)	11/10
Reproducibility dataset
Rest	Current study	22	25.68 (3.94)	4/18
Emotion	Current study	23	25.91 (4.01)	4/19
Tactile stimulation	Current study	19	25.68 (4.15)	2/17

Note: Age differed between the exploratory and reproducibility rest (t(41.68) = −5.51, p < 0.001, d = 1.58, BF₁₀ = 10,740), emotion (t(26.89) = −3.29, p < 0.01, d = 1, BF₁₀ = 18.78), and tactile stimulation (t(30.29) = −2.67, p = 0.01, d = 0.86, BF₁₀ = 4.59) datasets; participants in the reproducibility dataset were older than in the exploratory datasets. The sex ratio also differed between the exploratory and reproducibility emotion (χ ²(1) = 28.29, p < 0.001) and tactile stimulation (χ ²(1) = 6.17, p = 0.012) datasets, but no difference was observed for rest (χ ²(1) = 0.196, p = 0.66).

2.1.1. Exploratory Datasets

The exploratory resting‐state dataset, from Pavlov et al. (2022), included 30 participants. Participants completed 4 min of eye‐closed resting state while electrophysiological data were recorded with a 64 Electroencephalography (EEG) + 2 ECG channels (ActiCHamp, BrainProducts, Germany). Data were sampled at 1000 Hz, with FCz as the reference electrode and Fpz as ground. It was recorded and made available with the primary aim to help investigate physiological processes implicated in working memory, as the dataset also includes a digit span task (not analyzed here).

Data for the exploratory emotion dataset were retrieved from Mishra et al. (2021). Participants (N = 30) watched emotional videos while electrophysiological data were recorded using a 128 EEG and two ECG channels system (HydroGel Geodesic Sensor Net; Electrical Geodesics Inc.; United States of America) with a 250 Hz sampling rate. Only electrodes corresponding to standard 64 channels systems were included in subsequent analyses. Participants watched a total of 11 emotional or neutral 1‐min videos, and provided ratings for valence, arousal, liking and dominance for each video. Video categorization into neutral, positive, and negative categories was based on participants' valence ratings. This dataset was recorded and made available in order to help investigate the physiological processes associated with memory processing using ecological (video) stimuli. It also aimed to provide data for the Indian population which is less represented in the literature than the western individuals.

Data from the exploratory tactile stimulation dataset retrieved from Guidotti et al. (2023) included 21 participants. First, during a simple stimulation block targeting the Aβ fibers, discrete (< 50 ms) tactile stimuli were delivered to the right hand using tactors (vibrotactile stimulation device) placed on the participants' hands. This block lasted 12 min and included 240 stimuli with an interstimulus interval (ISI) of 1250 (±250) ms for 200 trials or 11,000 (±1000) ms for 40 trials. Next, during the complex stimulation block, participants received stroking stimulations on a 5 cm area of the forearm at a rate of 5 cm/s, hence targeting the C‐tactile fibers (Löken et al. 2009). Stimulation was delivered either by the experimenter's finger (social condition) or using an object (nonsocial condition). A total of 120 stimulations were delivered, with either a short ISI (650 ± 200 ms; 100 stimulations) or a long ISI (11,000 ± 1000 ms; 20 stimulations). Electrophysiological data were recorded using a 64 EEG + 2 ECG channels active system (Biosemi, Netherlands) at a 1024 Hz sampling rate referenced to the common mode sense (CMS) electrode. The original study from Guidotti et al. (2023) aimed to investigate the somatosensory evoked potential response according to different types of tactile stimulation, with different types of skin (glabrous or hairy) and different types of nerves stimulated.

2.1.2. Reproducibility Datasets

Twenty‐three participants were recruited for the current study to assess the reproducibility of the exploratory datasets, which comprised resting state, emotion, and tactile protocols. The sample size was determined from a power analysis using the meta‐analytic sample size observed by Coll et al. (2021) in the arousal condition as it matched our emotion condition. The power analysis was run for an effect size of g = 0.72, a significance level of α = 0.05 and a power β = 0.8 in a paired t‐test configuration as it was the statistical framework used by Coll et al. (2021). The analysis returned a minimum required sample size of N = 17. As our analyses rely on within‐subject ANOVAs and given the lack of information needed to convert g into ANOVA‐specific effect sizes (e.g., correlation between conditions), we increased the sample size to N > 20 to ensure sufficient power and account for methodological differences. Each participant signed an informed consent form, and the protocol received approval from the Ethics Committee (PROSCEA 2017/23; ID RCB: 2017‐A00756‐47). Participants were seated in front of a 1920 × 1080 pixels screen with a 60 Hz frame rate. First, participants completed a resting state protocol with their eyes open. A fixation cross was displayed on the screen for 5 min, and no behavioral output was required. Then, participants partook in the emotional protocol in which they were presented with positive, neutral, and negative videos with a duration of 236.05 s. One hundred and twenty videos were chosen from the DEVO database (Baraly et al. 2020) based on their valence and arousal ratings. The positive videos consisted of the videos with the highest valence and arousal values. The negative condition comprised videos with the lowest valence and highest arousal values. Finally, videos with a valence score between 4 and 6 were used to create the neutral videos. Participants were instructed to attentively watch each video and subsequently rate its perceived valence and arousal judgment on a 1–9 Likert scale.

Finally, participants underwent the tactile stimulation protocol. One block comprised nonrandomized social and nonsocial tactile stimulations (complex block), while another block comprised discrete stimulations (simple block). In the complex block, participants were asked to place their arm inside a stimulation box to prevent them from directly seeing the object used for stimulation. Tactile stimulations were delivered by the experimenter using either his finger (social touch) or an object (nonsocial touch). Fifty social and 50 nonsocial stimulations were performed sequentially on the participants' arm. Stimulations consisted of 1‐s‐long strokes on the participant's forearm with 3–5 s Gaussian random ISI. For the simple stimulations block, 100 discrete mechanical stimulations were delivered by vibrotactile tactors on the participants' arm or hand (ISI: 1000–1885 ms). The final number of participants varied across tasks, as data from some participants were excluded from the final analyses due to technical issues. Electrophysiological data were recorded using a 64 EEG + 2 ECG active Biosemi system at a 1024 Hz sampling rate referenced to the CMS electrode.

2.2. Electrophysiological Data Processing

2.2.1. Preprocessing

All datasets underwent the same preprocessing steps. Processing was performed in Python using the MNE software (Gramfort et al. 2013) and custom scripts. The data were filtered between 0.1 and 40 Hz (Tanner et al. 2015) and resampled to either 500 or 512 Hz, depending on the characteristics of the recording system (except for the exploratory emotion dataset, which was kept at a sampling rate of 250 Hz, as it was the recording sampling frequency). Data were referenced offline to the average reference. Bad data segments were identified and rejected based on visual inspection. Blink and saccade artifacts were corrected using independent component analysis (ICA) with the extended infomax algorithm (Lee et al. 1999). Artifactual components were rejected based on visual inspection. This procedure was applied to all datasets except for the eyes‐closed rest data from the exploratory rest dataset that did not contain any artifacts. Independent components (ICs) were fitted on a 2 Hz high‐pass‐filtered version of the original raw data to prevent interference from slow drifts during the estimation process (Winkler et al. 2015). The corrected signal was reconstructed by applying the IC weights to the original data. A current source density (CSD) transform (i.e., the Laplacian of the scalp voltage) using spherical splines (Perrin et al. 1989) was then applied to correct for the CFA, as implemented in previous studies (Pollatos et al. 2005). We ran a supplementary analysis, which also demonstrated that the CSD transform was better suited for CFA correction in our dataset (see supplementary analysis 1: Appendix S1). All follow‐up analyses were then done on the CSD‐converted data.

R‐peak events were identified using the built‐in function from the MNE software, and CSD‐converted data were epoched from −3 s to +3 s around the R‐peaks. These time windows were defined by including buffer periods before and after the intervals of interest, in order to prevent contamination of the subsequent time‐frequency representations (TFR) by edge effects (Cohen 2014). CSD‐converted epochs, henceforth referred to as epochs, underwent a linear detrend to correct for slow drifts. To further improve data quality, epochs with the 5% highest peak‐to‐peak amplitudes were rejected. The number of epochs statistics for each dataset and conditions is provided in Table S1. For the emotion and tactile stimulation datasets, epochs were classified into their corresponding conditions based on their co‐occurrence with experimental stimuli, with an upper limit of 0.25 s before the subsequent event onset. R‐peak epochs that did not co‐occur with experimental stimuli were discarded.

2.2.2. Event‐Related Potential

Prior to averaging, epochs underwent baseline correction (−0.15 to −0.05 s before R‐peak) and 30 Hz low‐pass filter. These additional steps reduced voltage offset variability across channels and facilitated ERP smoothing for subsequent peak detection. For each dataset, epochs were then averaged per conditions if any. ERPs pooled across conditions within a dataset were entered in a waveform analysis, to evaluate spatio‐temporal modulations. Component identification was done on the condition‐specific ERPs using either peak or mean amplitude. To identify the early component, we used a custom algorithm to detect the peak over FCz (or Cz in the exploratory rest dataset, as FCz was the reference channel in this dataset) as the point within the 50–250 ms time window where the signal derivative crossed zero with maximum absolute amplitude. Amplitude and latency at this point were then extracted for further analyses. A time window starting at 50 ms post R‐peak was used to allow the detection of individual peaks occurring just before 100 ms. For the late component, we applied a time window of interest (TOI) approach by extracting the mean amplitude between 250 and 500 ms over left (CP3, P3) and right centroparietal sites (CP4, P4), with amplitude averaged across the electrodes from each site.

2.2.3. Time‐Frequency Representation (TFR)

This analysis aimed to replicate the theta phase reset described in the literature (Park et al. 2018), hence it was uniquely performed for the exploratory resting state dataset and replicated in the reproducibility resting state dataset. The epochs were decomposed in the time‐frequency domain within the 2–20 Hz frequency range using Morlet wavelets, as described by Tallon‐Baudry et al. (1997). The analytic signal was computed by convolving 5 cycles of complex Morlet wavelets with the signal using fast Fourier transform. To prevent edge effects from contaminating the time window of interest (−200 to 500 ms), buffer periods of 2700 and 2500 ms were appended before and after the period of interest, respectively. These specific lengths were chosen to allow the inclusion of more than three cycles at the lowest frequency studied (2 Hz) (Cohen 2014). Power was extracted by averaging the analytic signal's magnitude over epochs, while the ITPC was extracted by calculating the length of the average vector from the phase angles distribution over epochs. Both power and ITPC were then baseline‐corrected by subtracting the mean from the −200 to −100 ms period, allowing for the investigation of event‐related changes in frequency dynamics. We then used irregular resampling auto‐spectral analysis (IRASA) (Gerster et al. 2022; Wen and Liu 2016) to verify if any results in the TF analysis, mainly in the theta band, reflect periodic activity (Donoghue et al. 2021). The IRASA algorithm aims to separate the oscillatory (or periodic) activity from the aperiodic signal by returning power spectral density (PSD) spectrums for periodic and aperiodic activity. Oscillations hence appear as peaks in the periodic spectrum. Epochs were fed into an IRASA algorithm adapted from Cole et al. (2019). PSD spectrums were computed using the Welch method (Welch 1967) with 3 s time windows on resampled data with h factors from 1.1 to 1.95 with a step of 0.05. Then, peaks were detected in the periodic spectrum between 2 and 8 Hz using the same algorithm as described for ERP peak detection, with the modification that only positive maximum values were considered for peak detection.

2.3. Statistical Analyses

2.3.1. Waveform Analysis

Identification of ERP deflections corresponding to the hypothesized components was conducted on the exploratory rest, emotion, and tactile stimulation datasets, with conditions pooled within each dataset. This analysis was subsequently replicated on the reproducibility dataset for the rest task only. A spatio‐temporal waveform analysis was performed on each electrode and at each time point to characterize the ERP deflections between 0 and 500 ms. Components were estimated using one‐sample t‐tests and cluster‐based permutation testing. Clusters were formed using a threshold computed as the 95th percentile t‐value from a t‐distribution with N−1 degrees of freedom. Cluster Monte Carlo p‐values were computed by comparing the cluster mass to a null distribution estimated from 10,000 permutations with random sign flips.

2.3.2. TFR Analysis

Power and ITPC were extracted and averaged from a time‐frequency window of interest between 100 and 250 ms and between 4 and 7 Hz over Cz. This specific electrode was selected to match the site used for the investigation of the early component, as it is hypothesized to be generated by theta phase resetting. Significance for heartbeat‐related modulations was established using one‐sample t‐tests against zero, and evidence towards the alternative or null hypotheses was quantified by calculating corresponding Bayes factors.

2.3.3. CFA Diffusion Control Analysis

Additionally, a control analysis was performed on the exploratory rest dataset to ensure that the early component peak was not due to diffusion from the ECG T‐wave. To do that, the peak amplitude in the 50–250 ms time window was extracted from the EEG and ECG signals. A correlation analysis was then conducted between the individual peak amplitude from EEG and ECG. As the R‐peak could still spread towards the time window of interest, we also correlated the ECG R‐peak amplitude with the HEP early component amplitude. We also investigated ECG amplitude differences between conditions as a supplementary analysis (see supplementary analysis 3: Appendix S1 for detailed methods and results). Hypothesis‐based analyses were then conducted on the exploratory emotion and tactile stimulation datasets and replicated on the matching reproducibility datasets. These analyses focused on the intra‐task modulation of the early and late components.

2.3.4. Early Component

Latency and amplitude of the early component were analyzed with a repeated measures ANOVA with condition (positive vs. negative vs. neutral for the emotion dataset, or social vs. nonsocial vs. simple for the tactile stimulation dataset) as a within‐subjects factor.

2.3.5. Late Component

The mean amplitude derived from the TOIs was analyzed with a repeated measures ANOVA with condition and hemisphere (left vs. right) as within‐subject factors.

2.3.6. Exploratory Analyses

Exploratory analyses were conducted to investigate the effect of conditions outside the TOIs for the exploratory datasets. For these analyses, cluster‐based permutation testing was applied to condition contrasts (paired t‐tests) within the 50–500 ms time‐window, using the same parameters as those employed in the waveform analysis.

2.3.7. Inter‐Task Analysis

Finally, hypothesis‐driven analyses were performed to assess inter‐task modulation for each component within each dataset group independently. For exploratory datasets, an ANOVA was conducted with Task (rest, emotion, tactile stimulation) as a between‐subjects factor, using amplitude values of the early component averaged across conditions. A similar analysis was performed to analyze the late component averaged across conditions, with the additional inclusion of Hemisphere (left, right) as a within‐subject factor. These analyses were replicated on the reproducibility datasets, where both Task and Hemisphere factors were considered within‐subjects.

ANOVAs were performed using JASP (JASP Team 2024). Statistical significance was assessed through frequentist p‐value, and evidence in favor of the alternative (and inversely the null) hypothesis was quantified using Bayes Factor (BF₁₀), interpreted according to the scale proposed by Lee and Wagenmakers (2014) (see Table S2). The BF₁₀ quantifies the relative increase in evidence for the alternative model over the null model, assuming both were equally probable a priori. For example, a BF₁₀ value of X means that the observed data are X times more likely under the alternative hypothesis than under the null hypothesis. Hence, BF₁₀ values higher than 1 describe evidence towards the alternative hypothesis, and values inferior to 1 quantify evidence towards the null hypothesis. For factorial ANOVA models, inclusion Bayes factors were used to characterize evidence towards the inclusion of each predictor. Models with each factor combination are computed, and the results are then averaged over all the resulting models. The inclusion BF quantifies the change of evidence towards the inclusion of a predictor with the observation of data (van den Bergh et al. 2020). We used the default priors for ANOVA models, as proposed by Rouder et al. (2012). Cluster‐based permutation analyses were conducted using built‐in functions from MNE (Gramfort et al. 2013). Correlation analyses from the control analysis were performed using the Pingouin python package (Vallat 2018) providing both frequentist p‐values and Bayesian BF₁₀.

3. Results

3.1. Rest

3.1.1. Waveform Analysis

The spatio‐temporal waveform analysis revealed three significant clusters in the exploratory dataset (Figure 1):

1. Early negative deflection (20–300 ms) over central sites (p = 0.0026).

2. Late positive deflection (200–500 ms) over centro‐posterior sites (p = 0.0002).

3. Late negative deflection (200–450 ms) over anterior sites (p = 0.0219).

FIGURE 1.

HEP topographical maps for the early and late components for the exploratory rest, emotion and tactile stimulation datasets. Data were averaged over the time windows corresponding to each component (early: 100–250 ms; late: 250–500 ms). The early component manifests as a fronto‐central negativity, and the late component appears as a centro‐posterior negativity. Scales differ for the early and late components. Highlighted electrodes correspond to significant electrodes from the waveform analyses, that is, cluster‐based permutation analysis against zero.

Waveforms illustrating the early negative and late positive deflection corresponding to the two components of the HEP are illustrated in Figure 2. In the reproducibility dataset, a single significant cluster was identified, capturing early (0–350 ms) negative deflections over central, anterior, and posterior sites (p = 0.0001).

FIGURE 2.

Illustration of the early and late components on the exploratory rest dataset HEP ± 95% confidence intervals (shaded areas). The electrode displayed is Cz. The early component corresponds to the negative peak between 100 and 250 ms post R‐peak. The late component corresponds to the latter positivity arising after 300 ms here.

3.1.2. CFA Diffusion Control Analyses

Results from the control analysis revealed no correlation between the HEP early component and the ECG T‐wave (r = 0.12, p = 0.52, BF₁₀ = 0.277) with moderate evidence towards an absence of correlation. Similarly, no correlation was observed between the HEP early component and the ECG R‐peak: r = 0.11, p = 0.54, BF₁₀ = 0.27.

3.1.3. TFR Analyses

Results from TFR analyses revealed a significant increase in ITPC relative to baseline in the exploratory dataset (t(29) = 2.91, p = 0.006, d = 0.53, BF₁₀ = 6.21). This effect was replicated in the reproducibility dataset (t(21) = 2.84, p = 0.01, d = 0.86, BF₁₀ = 5.13), providing strong evidence for phase‐resetting mechanisms. In contrast, no significant modulation of power dynamics was observed in either the exploratory dataset (t(29) = −0.38; p = 0.71, d = 0.07, BF₁₀ = 0.21) or in the reproducibility dataset (t(21) = −0.26, p = 0.8, d = 0.05, BF₁₀ = 0.23), with moderate evidence against the presence of a power effect. ITPC and power activity are represented in Figure 3.

FIGURE 3.

(A) ITPC and (B) Power results at electrode Cz. Values are expressed on a baseline‐normalized scale (proportion).

With IRASA, a theta peak was detected in 28/30 participants from the exploratory dataset, with frequencies ranging from 2.3 to 7.3 Hz (mean = 4.39 Hz). In the reproducibility dataset, a theta peak was detected in 21/22 participants, with a similar frequency range (mean = 4.49 Hz).

3.2. Emotion

3.2.1. Waveform Analysis

The spatio‐temporal waveform analysis identified two significant clusters:

1. Early negative deflection (100–250 ms) over central sites (p = 0.0001).

2. Late positive deflection (175–450 ms) over left posterior sites (p = 0.0004).

3.2.2. Early Component

The ANOVA revealed no effect of condition on amplitude (F(2, 56) = 0.69, p = 0.50, BF₁₀ = 0.179). The data were 5.59 times more likely under the null hypothesis. This result was replicated on the reproducibility emotion dataset (F(2, 44) = 0.391, p = 0.68, BF₁₀ = 0.161). Regarding latency, a marginal but inconclusive effect was observed (F(2,56) = 3.01, p = 0.057, BF₁₀ = 1.1). However, in the reproducibility dataset, the data were about 8.26 times more likely under the null hypothesis (F(2, 44) = 0.019, p = 0.98, BF₁₀ = 0.121), further supporting the absence of effect.

3.2.3. Late Component

No significant effects were found for condition (F(2, 56) = 1.065, p = 0.35, BF₁₀ = 0.234), Hemisphere (F(1, 28) = 0.001, p = 0.97, BF₁₀ = 0.141) or their interaction (F(2, 56) = 0.927, p = 0.40, BF₁₀ = 0.059). These results were replicated on the reproducibility dataset: condition (F(2, 44) = 1.068, p = 0.352, BF₁₀ = 0.162), Hemisphere (F(1, 22) = 1.198, p = 0.28, BF₁₀ = 0.393), interaction (F(2, 44) = 0.417, p = 0.662, BF₁₀ = 0.047). Waveforms for the HEP for the three conditions are displayed in the Figure 4.

FIGURE 4.

HEP waveform ± 95% confidence intervals (shaded areas). HEP waveforms for the emotion exploratory dataset over three electrodes: FCz (top), CP3 (middle), and CP4 (bottom)—the electrodes used for the early and late components analyses. Waveforms are shown for each condition: neutral (dark blue), positive (light green), and negative (dark green).

3.2.4. Exploratory Analysis

Exploratory cluster‐based permutation analyses for the positive vs. neutral contrast and for the negative vs. neutral contrast revealed no significant clusters in the exploratory dataset.

3.3. Tactile Stimulation

3.3.1. Waveform Analysis

The spatio‐temporal waveform analysis identified two significant clusters:

1. Early negative deflection (50–300 ms) over central sites (p = 0.0155).

2. Positive deflection extending from right anterior sites (50–200 ms) to posterior sites (170–500 ms; p = 0.0003).

3.3.2. Early Component

In the exploratory dataset, no effect of condition was observed on amplitude (F(2, 34) = 0.96, p = 0.39, BF₁₀ = 0.285), with data being 3.51 times more likely under the null hypothesis. This result was replicated in the reproducibility dataset (F(2, 36) = 0.152, p = 0.859, BF₁₀ = 0.154). For latency, the exploratory dataset suggested a marginal effect of condition (F(2, 34) = 2.63, p = 0.087, BF₁₀ = 1.04), although the Bayes factor conveyed almost no evidence toward it. However, in the reproducibility dataset, an effect of condition on latency was observed (F(2, 36) = 3.33, p = 0.047, BF₁₀ = 1.569), with a higher latency for nonsocial vs. social stimulations (t = 2.562, p = 0.044, BF₁₀ = 4.817).

3.3.3. Late Component

In the exploratory dataset, no effect of condition (F(2, 34) = 1.099, p = 0.35, BF₁₀ = 0.248) or interaction between condition and Hemisphere (F(2, 34) = 0.996, p = 0.38, BF₁₀ = 0.229) was found. However, a main effect of hemisphere was detected (F(1, 17) = 6.736, p = 0.019, BF₁₀ = 1.658) with higher amplitude in the right vs. left hemisphere. Lack of condition effect (F(2, 36) = 0.546, p = 0.584, BF₁₀ = 0.144) and interaction (F(2, 36) = 0.174, p = 0.841, BF₁₀ = 0.032) were replicated on the reproducibility dataset. However, we did not find a main effect of hemisphere (F(1, 18) = 0.399, p = 0.536, BF₁₀ = 0.283) on the reproducibility dataset. Waveforms for the HEP for the three conditions are displayed in the Figure 5.

FIGURE 5.

HEP waveform ± 95% confidence intervals (shaded areas). HEP waveform for the tactile stimulation exploratory dataset over three electrodes: FCz (top), CP3 (middle), and CP4 (bottom)—the electrodes used for the early and late components analyses. Waveforms are shown for each condition: simple (dark blue), social (light green), and nonsocial (dark green).

3.3.4. Exploratory Analyses

Cluster‐based permutation analyses on the social vs. nonsocial, the social vs. simple and the nonsocial vs. simple contrasts revealed no significant clusters in the exploratory dataset.

3.4. Inter‐Task Modulation

3.4.1. Early Component

No effect of task was observed on the early component in either the exploratory datasets (F(2, 74) = 2.07, p = 0.133, BF₁₀ = 0.589; inconclusive BF) or the reproducibility dataset (F(2,36) = 0.675, p = 0.52, BF₁₀ = 0.231; BF in favor of the null hypothesis). The data were about 4.33 times more likely under the null hypothesis in the reproducibility dataset.

3.4.2. Late Component

In the exploratory dataset, the late component was not affected by task (F(2, 74) = 0.732, p = 0.48, BF₁₀ = 0.194). A marginal effect of hemisphere was detected with weak evidence towards the inclusion of this factor (F(1, 74) = 3.258, p = 0.075, BF₁₀ = 0.53). No interaction between hemisphere and task (F(2, 74) = 1.137, p = 0.326, BF₁₀ = 0.266) was found. In the reproducibility dataset, a marginal effect of task was observed (F(2, 36) = 3.026, p = 0.061, BF₁₀ = 0.628), but the Bayes factor remained inconclusive. No main effect of hemisphere with inconclusive evidence (F(1, 18) = 0.74, p = 0.40, BF₁₀ = 0.559) and no interaction between task and hemisphere (F(2, 36) = 0.043, p = 0.96, BF₁₀ = 0.144) were found.

4. Discussion

This study aimed to better characterize the HEP components at rest and across different tasks by exploring their intra‐ and inter‐task modulations and ensuring replication across independent datasets. The spatial–temporal waveform analysis at rest revealed two components in the HEP: an early fronto‐central negative component (100–250 ms), corresponding to a negative peak in the ERP waveform, and a later and more posterior positive component (250–500 ms). These two components were consistently observed across tasks and datasets. The presence of an increased ITPC, in the absence of a power increase, suggests that the HEP is generated through phase resetting (Sauseng et al. 2007), and was in line with previous findings (Park et al. 2018). Moreover, theta periodic activity was also observed in more than 90% of participants, suggesting the presence of theta activity. Moreover, we ensured using control analyses that the amplitude of the early peak was not associated with the amplitude of the ECG T‐wave or R‐peak, suggesting that the early peak did not reflect the CFA. Moreover, no modulation of the ECG amplitude across conditions was observed (see supplementary analysis 3: Appendix S1). To assess the functional significance of these two components, we examined their modulation by intra‐task conditions, hypothesizing that only the late component would be sensitive to task demands. As expected, the early component remained unaffected by task. However, contrary to our expectations, the late component also showed no modulation across tasks, analysis and dataset used. This consistent absence of task‐related effects suggests that the late component is not influenced by cognitive or perceptual factors inherent to the experimental paradigms. Importantly, the datasets used were collected across culturally diverse populations (France, India, and Russia), generalizing the results beyond the Western Educated Industrialized Rich & Democratic (WEIRD; Henrich et al. 2010) population. This is of importance since studies emphasized cultural differences in the expression of emotions and interoceptive awareness across different cultures (for review: Ma‐Kellams 2014), which could have influenced HEP modulations. Although all results were replicated across datasets, the absence of modulation in either dataset does not count as evidence towards an absence of cultural difference in HEP modulation, as this hypothesis was not explicitly investigated in this study. The exploratory and reproducibility datasets have imbalanced sex distribution; to control for potential confounding factors the analysis was redone with sex as a factor (see supplementary analysis 2: Appendix S1). This analysis provided results similar to the ones observed without taking sex into account, suggesting that the imbalanced sex ratios between datasets did not influence the results.

4.1. Competing Models for the HEP

The HEP waveform observed in these studies matched the early peak waveform type, consistent with some previous studies (Mai et al. 2018; Salamone et al. 2018; Fukushima et al. 2011; Gentsch et al. 2019; Ito et al. 2019; Judah et al. 2018). The exploratory waveform analysis emphasized an early negative peak at fronto‐central areas and a later more sustained positivity at more posterior sites. This suggests different sources for the two components in line with our hypothesis of a two‐component model. However, the lack of task modulation of the late component does not allow us to uncover whether the two components are associated with different functions. Moreover, the absence of interaction between task and hemisphere in the inter‐task analysis suggests that the late component's topography is the same across tasks. These elements are contradictory to the two‐component model but do not necessarily constitute evidence towards the single‐component model as both components are still associated with different brain sources. Results from the literature also seem to fit more with the two‐component model as condition effects are most of the time reported in a time window corresponding to the late component (Coll et al. 2021; but see Fukushima et al. 2011; Judah et al. 2018) who show amplitude modulation during an early negative peak. However, results are constrained by the time window used for analysis that often were defined to avoid CFA influence (Coll et al. 2021), biasing results towards late time windows. Another reason for the absence of HEP modulation could be related to the experimental paradigms used in this study. The emotion paradigm has been widely used in the literature and has been shown to be well suited for inducing HEP modulations. There is limited evidence of tactile‐induced interoceptive modulation. However, a modulation of the HEP amplitude in a late time window has been described during a tactile stimulation protocol (Al et al. 2020). The absence of HEP modulation could also be due to the fact that the primary goal of the datasets was not to investigate the HEP. Hence, the interoceptive component of the task was not emphasized during instructions. Yet, this would suggest that HEP modulations are dependent on a certain context that yields implicit or explicit awareness of the subject of the study.

4.2. Variability in Methods

The HEP literature is characterized by a large variability in preprocessing methods. This may contribute to the large heterogeneity of waveforms and latency for the HEP effects. One source of variability relates to CFA correction. However, the CFA correction used does not seem to be related to the type of waveform observed: (1) ECG‐like waveforms were observed across studies using ICA (García‐Cordero et al. 2017; Kim et al. 2019) or no correction (Petzschner et al. 2019; Perogamvros et al. 2019; Schulz et al. 2015); (2) early peak waveforms were observed using ICA (Mai et al. 2018; Salamone et al. 2018; Gentsch et al. 2019); scaled EEG subtraction (Fukushima et al. 2011; Judah et al. 2018); and (3) flat waveforms were observed with ICA corrections (Babo‐Rebelo, Richter, and Tallon‐Baudry 2016; Babo‐Rebelo, Wolpert, et al. 2016). Although ICA is a largely used and very efficient method for eye‐related artifact correction (Hoffmann and Falkenstein 2008), it may not be optimal for CFA correction in the context of the HEP at least in scalp EEG studies. Indeed, heartbeat‐related ICs cannot be reliably identified in all individuals, leading to inconsistencies in artifact rejection criteria. The supplementary analysis 1: Appendix S1 revealed that a CFA‐related ICA component was retrieved in 9.09% of participants in the replication rest dataset. The identified components may also capture a mixture of CFA and brain activity, which could lead to an alteration of the HEP waveform. An alternative approach is CSD transformation, which may correct for the CFA while preserving neural signals. This is suggested by the absence of correlation between the ECG T‐wave and HEP early component amplitude in our study. The supplementary analysis 1: Appendix S1 also revealed that in the context of the current study, only the CSD transformation allows a reduction of the CFA; the ICA did not prove sufficient to suppress the CFA. Moreover, the CSD transform constitutes a reference‐free representation of EEG data (Kayser and Tenke 2015). The adoption of CSD transformation for CFA correction would contribute to reducing the heterogeneity due to the different reference electrodes used across studies. However, the CSD remains a data transformation technique, so its use may contribute to the lack of replication of the effects described in the literature in our study. It is also important to note that no existing method seems to fully eliminate the CFA, highlighting the need for methodological standardization in future research. It is also to note that more elaborated methods using neural networks to correct for the CFA exist. It should be evaluated whether these methods allow for a correction of the CFA while preserving the HEP, as this was tested against R‐peak locked exteroceptive stimulation.

Other sources of variability in preprocessing concern the use of a baseline correction and the ECG event used as HEP onset. The use of a baseline correction is controversial in the HEP literature due to the cyclical nature of cardiac activity. While the majority of authors reviewed here reported applying a baseline correction (Coll et al. 2021), some authors advocate against its use (Babo Rebelo 2017). There is also variability in terms of the ECG event used as HEP onset. While most studies used the R‐peak, others used the ECG T‐wave (Park et al. 2014).

Additionally, the statistical framework used to analyze HEP data varies considerably. Indeed, some studies (Park et al. 2018) use surrogate heartbeat analyses to investigate whether the observed effects are locked to the heartbeats. These surrogate analyses can control for confounding activity in the context where a concomitant task is performed by the participants (Steinfath et al. 2024), and that task events are—by design—time‐locked to heartbeats (Park and Blanke 2019). Yet they do not address the same research questions as traditional null hypothesis testing. While regular cluster‐based permutation analyses directly test for the presence of an effect by building their null distribution by permuting condition labels, surrogate analyses assess whether observed clusters (in terms of their mass) could be generated by nonheartbeat locked data. In the context of our study, one may think that retaining only R‐peaks occurring within a defined time window around experimental stimuli could result in confusion between the HEP and the ERP evoked by the external stimulation. However, the jitter between stimulus onsets and R‐peaks, resulting in stimuli not being time‐locked to ECG R‐peaks, effectively prevents task‐related activity from contaminating the HEP. This is an argument against the use of surrogate analysis in this precise context. Finally, the number of epochs used for ERP calculation influences ERP reliability and statistical power, with recommendations regarding the minimum number of epochs depending on the ERP component considered (e.g., 150 epochs for an MMN, 36 for a P300 and 40 for an N400; Steinfath et al. 2024; Park and Blanke 2019). The lack of recommendations for the study of HEP also contributes to variability in results. The implications of these methodological differences warrant further investigations. Future studies should investigate more systematically the influence of sources of variability (e.g., baseline, number of epochs, ECG events considered, etc.) on HEP reliability.

5. Perspectives

Although results from this study provide partial support for the two‐components model, we are still unable to conclude concerning the functional significance of the late component. To further explore the functional mechanisms allowing interoception under the two‐components model, future studies may also move beyond simple modulation studies toward more functionally interpretable studies. Indeed, phase resetting has been proposed to promote information integration between brain areas (Canavier 2015). Hence, theta phase resetting in primary interoceptive areas may trigger synchronization within an extended, task‐specific processing network via theta phase synchronization. This hypothesis can be assessed by investigating whether theta phase connectivity between the regions associated with the early component and the ones associated with the late component is present and whether it is modulated across conditions. Information exchange can also be assessed using information theory‐derived directed measures. Some studies already applied connectivity analyses to the HEP, showing increased connectivity between frontal and posterior regions after interoception feedback (García‐Cordero et al. 2017), increased connectivity from the anterior insula to the anterior cingulate cortex during the perception of sad faces (Kim et al. 2019), and an increase in theta phase connectivity following heartbeats at rest (Kim and Jeong 2019), highlighting the existence of an extended cardiac information processing network. An additional issue in the HEP literature is the existence of multiple HEP waveforms, which do not seem to relate to the CFA correction methods or preprocessing steps undertaken. Future studies may investigate which factors lead to different HEP waveforms.

6. Conclusion

In conclusion, this study supports the two‐component model of the HEP, comprising an early component that may reflect primary cardiac signal integration, and a late component whose functional significance remains unclear. While the absence of task modulation on the late component contradicts prior findings the robustness of this result is reinforced by the methodological rigor of our approach, drawing on multiple datasets, experimental designs, and participants from various cultures. These findings highlight the need for standardized methodologies in HEP research to improve reproducibility and, enhance our understanding of cardiac‐related neural processing. Future research should aim to move from exploratory studies to more theoretically informed studies encompassing the different components described in the literature.

Author Contributions

Raphaël Gautier: conceptualization, data curation, formal analysis, investigation, writing – original draft. Marianne Latinus: conceptualization, supervision, writing – review and editing. Frederic Briend: conceptualization, supervision, writing – review and editing.

Funding

This work was supported by the EXcellence Center in Autism and neurodevelopmental disorders—Tours, CHRU de Tours, 2 boulevard Tonnellé, 37044, Tours Cedex 9, France.

Ethics Statement

The protocol received approval from the Ethics Committee (PROSCEA 2017/23; ID RCB: 2017‐A00756‐47).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Appendix S1: psyp70206‐sup‐0001‐Supinfo.docx.

Gautier, R. , Latinus M., and Briend F.. 2025. “Characterizing the Heartbeat‐Evoked Potential: A Two‐Component Model of Cardiac Signal Processing?.” Psychophysiology 62, no. 12: e70206. 10.1111/psyp.70206.

Data Availability Statement

The data that support the findings of this study are openly available in HEP_Characterization at https://gin.g‐node.org/Rgautier/HEP_Characterization.git.