Cardiovascular mechanisms of interoceptive awareness: Effects of resonance breathing

Abstract

Interoception, the ability to perceive internal bodily sensations, and heart rate variability (HRV) share common physiological pathways, including the baroreflex feedback loop. The baroreflex can be activated by resonance breathing, wherein respiration is paced at 6 times per minute (0.1 Hz), eliciting immediate physiological changes and longer-term therapeutic responses.

This registered report characterizes baroreflex functioning as a cardiac mechanism of interoception in a two-session study (n = 67). The heartbeat discrimination task was used to obtain indices of interoceptive accuracy, sensibility and metacognition. Baroreflex functioning was measured as HRV at 0.1 Hz and baroreflex sensitivity (BRS); high frequency (HF) HRV was calculated as a control. Cardiovascular indices were measured at baseline and during active and control paced breathing after which changes in interoception were measured.

The first hypothesis was that baseline baroreflex functioning would predict individual differences in interoceptive awareness. The second hypothesis was that resonance breathing would increase participants’ ability to detect their own heartbeats, and that this effect would be mediated by increases in 0.1 Hz HRV and BRS. Data were collected upon in principle acceptance of the manuscript.

We found a negative relationship of interoceptive accuracy with baseline HF HRV and BRS, and a positive relationship between metacognitive interoception and 0.1HZ HRV, BRS and HF HRV. We found that changes in 0.1 Hz HRV and BRS during resonance breathing positively correlate with increases in interoceptive accuracy.

Our results show that the extent to which breathing recruits the resonant properties of the cardiovascular system can facilitate the conscious perception of participants’ heartbeats. We interpret this as an increase in vagal afferent signaling and baroreflex functioning following resonance breathing. We put forward an alternative explanation that HRV modulation can reduce interoceptive prediction errors, facilitating the conscious perception of interoceptive signals, and consider the role of resonance breathing on mental health from an interoceptive inference perspective.

1. Introduction

Interoception is the integration and processing of internal bodily states in the brain (Cameron, 2001; Sherrington, 1948). Interoceptive awareness is the perception of these internal body states and is typically studied in relation to the cardiac axis (Garfinkel et al., 2016a). It is garnering increased research interest due to its role in emotional processing, cognitive functioning, and overall mental health (Critchley and Garfinkel, 2017, Critchley and Garfinkel, 2018; Khalsa et al., 2017; Pollatos and Ferentzi, 2018). Interoceptive deficits have been shown to contribute to mental disorders, including addiction, depression, anxiety, and autism (Garfinkel et al., 2016b; Khalsa et al., 2017; Paulus and Stein, 2010; Quattrocki and Friston, 2014; Sönmez et al., 2016; Verdejo-Garcia et al., 2012), yet methods to improve interoceptive awareness are limited.

The physiological mechanism by which interoceptive cardiac signals are experienced is presumed to rely on the firing of arterial baroreceptors at systole, when blood is ejected from the heart (Critchley and Garfinkel, 2015; Garfinkel et al., 2013). These peripheral signals are communicated to the brain through afferent pathways via the vagus nerve to provide information about the timing and intensity of each heartbeat. These signals enter the brain at the nucleus of the solitary tract (Critchley and Harrison, 2013) and are relayed to the anterior insular cortex, where interoceptive information is mapped (Critchley et al., 2004). This afferent information stream (from heart to brain), along with the efferent sympathetic and parasympathetic (vagal) information stream, comprises the baroreflex (Goldstein, 2000), a feedback loop in the brain-heart axis through which blood pressure (BP) and cardiac function regulate each other (Benarroch, 1993; Critchley and Harrison, 2013). This feedback loop plays a crucial role in the adaptation of the body to environmental changes and challenges (Bates and Buckman, 2013) and is thus a sign of cardiac health and overall physiological adaptability (Bristow et al., 1969; Lantelme et al., 2002). The baroreflex is a putative mechanism that may underlie individual differences in interoceptive awareness of heart beats.

Interoceptive awareness has been experimentally characterized using either the heartbeat discrimination task (Katkin et al., 1983; Whitehead et al., 1977), which assesses an individual’s ability to determine whether a series of auditory cues are synchronized with their heartbeats, or the heartbeat tracking task, which asks participants to count their heartbeats over different periods of time (Schandry, 1981). These tasks aim to dissociate interoceptive accuracy (objective performance on the task), sensibility (subjective confidence ratings), and metacognitive interoception or insight (the mapping of confidence onto accuracy) (Khalsa et al., 2017), which are not necessarily correlated with one another (Garfinkel et al., 2015). The tracking task, however, has recently been subjected to methodological criticisms, due to participants using estimates of their heart rates to complete the task instead of relying on their actual perception of individual heartbeats (Desmedt et al., 2018; Zamariola et al., 2018).

Eichler and Katkin (1994) found that pre-ejection period during a mental arithmetic task explained accuracy on the cardiac discrimination task in men, suggesting that high cardiovascular reactivity to stress was associated with better heartbeat detection accuracy. Baseline accuracy measures on the tracking task also correlated with cardiovascular responses (e.g., average heart rate, cardiac output, pre-ejection period) after an isometric effort (Pollatos et al., 2007), further supporting that the strength of cardiac reactions was tied to greater heart beat awareness. These measures of cardiovascular response suggest that multiple cardiovascular mechanisms may underlie interoceptive awareness. However, these measures focus on cardiac averages, which provide information about metabolic demands and general functioning. Variability indices of cardiac functioning, on the other hand, are particularly well suited to assess the ability of the system to dynamically adapt to internal and external demands (Bates et al., 2011). Heart rate variability (HRV) is measured as changes in the length of beat-to-beat intervals (RRI). It has been linked to cognitive and emotional adaptation (Lane et al., 2009) as well as mental health (i.e. Bates and Buckman, 2013; Kemp, 2016; Quintana et al., 2013). Importantly, HRV is often used as a proxy of baroreflex activity (Berntson et al., 1997; see Mccraty and Shaffer (2015) for a detailed examination of the physiological mechanisms of HRV). In particular, HRV at 0.1 Hz is a measure of peak power in the heart rate spectrum at that frequency and closely reflects baroreflex activation (deBoer et al., 1987; Vaschillo et al., 2002). In addition, BRS can be measured directly, for example, by assessing the amount of HR change in response to 1 mmHg change in blood pressure. This can be quantified using cross-spectral analysis where BP is the input and RRI is the output in a transfer function (Vaschillo et al., 2012).

Because the baroreflex mechanism is considered to be related to both interoceptive signals and HRV, the first goal of this registered report (Larson, 2016) is to examine whether interoceptive awareness across the cardiac axis is determined by individual differences in baroreflex functioning, measured as 0.1 Hz HRV and BRS. Moreover, mixed results have been found in the relationship between HF HRV, a measure of vagal activity, with interoceptive awareness using the tracking task (Herbert et al., 2010, Herbert et al., 2012); thus, HF HRV was also examined. In light of the recent methodological concerns about the tracking task, we used the discrimination task to assess cardiac interoception in this experiment.

The second aim of this registered report is to assess whether a respiratory task known to activate the baroreflex, increase baroreflex sensitivity, and maximize HRV (Lehrer et al., 2006), is able to alter interoceptive awareness. Early evidence suggested that interoceptive accuracy may be increased after a cardiac signaling biofeedback procedure (Meyerholz et al., 2019) or a “power posing” task (Weineck et al., 2019), but these studies may be limited by their use of the heart rate tracking task. Interoceptive accuracy also was susceptible to the effects of pharmacological agents, being modulated by alcohol administration (Leganes-Fonteneau et al., 2019; Leganes et al., 2021) and increased by intranasal oxytocin (Betka et al., 2018) when measured with the heartbeat discrimination task. In parallel, HRV is also modulated by acute alcohol administration (Leganes-Fonteneau et al., 2021; Leganes-Fonteneau et al., 2020a; Vaschillo et al., 2008) and increased by oxytocin (Kemp et al., 2012). While these studies demonstrate that interoceptive awareness and HRV are malleable, the pharmacological approach generates broad effects on participants that are not specific to the cardiovascular system.

Biofeedback techniques based on resonance breathing (Song and Lehrer, 2003; Vaschillo et al., 2002) are non-invasive and easy to learn and use outside of clinical settings. Rhythmical breathing at a frequency of 0.1 Hz (6 times per minute), as opposed to natural arrhythmical breathing (12–20 times per minute), triggers a series of physiological reactions that increase the amplitude of heart rate oscillations compared to resting state. The complex physiological processes by which this occurs include the coupling of phase in oscillations between breathing, HR, and BP that exhibit the resonance characteristics of the cardiovascular system and engage the baroreflex mechanism (Lehrer and Gevirtz, 2014). This generates drastic and immediate increases in BRS (Lehrer et al., 2003) and HRV, centered primarily at 0.1 Hz (Lehrer and Eddie, 2013). These changes are considered healthful as they strengthen the adaptive range and flexibility of the cardiovascular system in the face of challenges (Vaschillo et al., 2006). Resonance breathing also has proven clinical benefit in various mental and physical health conditions (Eddie et al., 2014; Goessl et al., 2017; Hassett et al., 2007; Karavidas et al., 2007; Reiner, 2008; Siepmann et al., 2008; Sutarto et al., 2012; Windhorst, 2007). In this registered report, we proposed to manipulate the cardiovascular system with resonance breathing to increase 0.1 Hz HRV and BRS, and predicted that this also would increase interoceptive awareness.

Given that cardiac interoceptive awareness purportedly represents the integration of afferent baroreceptor firing in the brain (Critchley and Garfinkel, 2015), an increase in BRS theoretically should increase the perception of individual heartbeats. This is because increased BRS implies that the cardiovascular system is more attuned to changes in blood pressure after each cardiac contraction, via the neural mediation accomplished by the baroreflex loop. As a result of the cardiac resonance paced breathing task, we expect an increase in BRS (Lehrer et al., 2003) reflecting more efficient neural processing of the afferent signals from the heart that originate from baroreceptor firing.

1.1. Hypotheses

• Our first goal was to examine whether cardiovascular functioning can predict individual differences in interoceptive awareness. Given the shared putative mechanism of HRV and cardiac interoceptive signals, we hypothesized that 0.1 Hz HRV and BRS would positively correlate with interoceptive awareness. We examined as well the equivalences between 0.1 Hz HRV and BRS in their power to predict interoceptive awareness.

• Our second goal was to modulate the cardiovascular system with a resonance breathing task. We expected acute increases in interoceptive awareness after a task that involves breathing 6 times per minute (6P-resonance breathing) compared to a control breathing task (VP-variable breathing). We further predicted that changes in cardiac perception would be mediated by the extent to which resonance breathing increases 0.1 Hz HRV and BRS.

2. Methods

2.1. Participants

We aimed to recruit a sample of 70 participants for this experiment from the students at and the general population around Rutgers University – New Brunswick (New Jersey, USA; see Power analysis section). We conducted a total of 163 experimental sessions in 96 participants; however, 29 participants only completed the first session due to the Covid-19 outbreak and were excluded from all analyses. Data from 67 individuals were available for analysis.

Participants were eligible if they did not report any serious health issue, cardiac abnormality or mental health disorder. Participants under pharmacological treatment (except birth control, Rebelo et al. (2011)) were excluded to avoid any interactions with the experiment. Participants with a Body Mass Index (BMI) 20% over- or under-weight (normal BMI = [18.5–24.9]) were also excluded (no exclusions). Considering previous literature on the decreased effects of paced breathing biofeedback interventions on HRV in participants older than 40 (Lehrer et al., 2006), only participants between 18 and 40 years of age were recruited.

2.2. Tasks and questionnaires

2.2.1. Physiological data recording

ECG data were recorded during the Vanilla (Physio_B1 and Physio_B2) and breathing tasks (Physio_6P and Physio_VP). A PowerLab 16/30 Acquisition System (ADInstruments, Colorado Springs, CO, US) was used to collect ECG with electrodes placed on their left and right arms and left ankle, and respiration using chest and abdominal straps. A finger cuff measured beat-to-beat BP (Finometer MIDI - Finapres, Amsterdam, The Netherlands). Data were obtained at a 2000-Hz sampling rate. Physiological data were processed for analysis using WinCPRS software (Absolute Alien Oy, Turku, Finland). Ectopic heartbeats and movement artifacts were manually edited without excluding data segments.

2.2.2. Physiological tasks

The Vanilla task is a low-demand baseline task used to measure ECG and BP at t₀. For 5 min, participants observed in the screen a series of colored squares presented one at a time for 10 s each. During this task, participants were requested to sit still and silently count how many “blue” squares they saw, although responses were irrelevant.

During the breathing task, participants observed a pacer on the screen that moved up and down according to the frequency of the breathing condition; physiological data were recorded throughout the task. Participants were instructed to time their breath to the pacer. In the 6P condition, the pacer moved up for 5 s and down for 5 s (6 breaths per minute). In the VP condition, the pacer moved at progressively variable frequencies, from a minimum of 10 to a maximum of 15 times per minute. Participants were free to breath-in and out through their noses or mouths.

2.2.3. Interoceptive discrimination

Participants’ ability to integrate interoceptive and exteroceptive information as an index of cardiac interoceptive awareness was evaluated with the discrimination task (Katkin et al., 1981). For each trial of this task, participants hear 10 consecutive tones (100 ms, 440 Hz) triggered by their heartbeats. On half of the trials the tones occur 200 ms after the R-wave, and on the other half, a phasic delay of 500 ms is inserted between the R-wave and the tone, corresponding to maximum and minimum synchronicity judgements respectively (Wiens and Palmer, 2001). At the end of each trial participants have to indicate whether the tone is synchronized or not synchronized with their heartbeat and how confident they are of their response using a visual analog scale (0–100).

Real-time data from the ECG is required for the discrimination task. An event trigger was set through Labchart 8 each time a voltage above 4 V was detected on the ECG channel, relaying a signal to the analog output of Powerlab that was then transmitted to a National Instruments USB-6001 data acquisition device for its integration on Matlab. This method for presenting stimuli timed with heart beats is more accurate than previous techniques based on r-wave prediction estimates from mean or median RRI on preceding heartbeats (i.e. Pfeifer et al., 2017).

2.2.4. Breath-modulated interoception task

After completing a 5-min block of paced breathing (6P or VP), participants performed 20 trials of the discrimination task. Every 2 trials, a 30-s period of paced breathing was included to potentiate the effects of resonance breathing on the cardiac system, out of which we obtained measures of breath-modulated interoception, see Fig. 1B.

Fig. 1.

A – Representation of the paced breathing tasks at different frequencies, B – Procedure for the breath-modulated interoception task, C – Experimental procedure for sessions 1 and 2. Physio_XX indicates the measurement of ECG data and the name of the associated variable.

2.3. Procedure

Participants were screened to assess their eligibility for the study. Those who were eligible completed 2 sessions at the Cardiac Neuroscience Laboratory, at the Center of Alcohol Studies, Rutgers University. This study was approved by the university’s IRB upon acceptance of the first stage of the manuscript registration. They received compensation ($7/h) for participation or were given student credits (2/h). At the first session, participants provided signed informed consent and completed basic demographic information confirming their eligibility. Their weight and height were measured to calculate BMI. Considering the effect that fitness level can have on the cardiovascular system (Ogoh et al., 2003), participants also completed a brief physical activity assessment (Marshall et al., 2005) consisting of 2 items measuring their weekly frequency of vigorous and moderate physical activity. A measure of fitness level was derived, scored from 0 to 8. Next, participants were seated and connected to the ECG equipment. At the first session, participants completed a baseline assessment of interoception comprised of 20 trials of interoceptive discrimination task (10 min) and the Vanilla task (5 min). They then completed either the resonance breathing task, which paces breath to 6 times per minute (6P), or the inactive control task (variable paced breathing, VP) for 5 min. Finally, participants completed the breath-modulated interoception task (~20 min). All tasks were completed in the same order (see Fig. 1C for a description of the experimental procedure).

In the second session, participants completed the Vanilla task, the breathing task not completed during the first session, and the breath-modulated interoception task. The order of paced breathing task (6P vs. VP) was counterbalanced. The second session took place at least one week after the first one, and was scheduled as close as possible to the same time of the day.

2.4. Data extraction

2.4.1. Interoceptive discrimination

Three indices were obtained: accuracy, sensibility and metacognitive interoception. Interoceptive accuracy was obtained using Type-1 d-prime scores (d′), calculating the difference between participants’ normalized hit rate and normalized false alarm rate (z(Hits) - z(False Alarms)) to account for responses biases (Sokol-Hessner et al., 2015). Data corrections were applied in case of hit rates or false alarm rates equal to 1 (subtracting 0.001) or 0 (adding 0.001). Sensibility was computed as the average confidence on each of the trials. Metacognitive interoception is the relationship between accuracy and confidence; an Area Under the Receiver Operating Characteristics Curve (AUROC) (Green and Swets, 1966; Hajian-Tilaki, 2013) was used to provide an index of the degree to which confidence scores predicted accuracy, while accounting for individual differences in confidence levels.

2.4.2. ECG data

WinCPRS software (Absolute Alien Oy, Turku, Finland) provided measures of beat-beat BP, HRV, Stroke Volume (SV – calculated using the ModelFlow methodology for cardiac output measurement (Bogert and van Lieshout, 2005)), and Pulse Transit Time (PTT - the interval between an R-wave measured on the ECG and the corresponding apex of the pulse wave on the finger). HRV was computed as the peak power in the resonance frequency (0.1 Hz in ms²; 0.075–0.108 Hz) and total power in the low (LF in ms²/Hz; 0.005–0.15 Hz) and high (HF; 0.15–0.5 Hz) frequency ranges. Transfer functions [TF(BP-RRI), TF(BP-SV), and TF(BP-PTT)] were calculated with systolic BP as the input and RRI, SV, or PTT as the outputs. HR-, SV-, and VT-BRS gain were assessed as average values of TF(BP-RRI), TF(BP-SV), and TF(BP-PTT), respectively, in the low frequency range where BP-RRI, BP-SV, and BP-PTT coherence was >0.5 (Vaschillo et al., 2012).

The primary physiological variables of interest in the registered analysis were HR-BRS and 0.1 Hz HRV, as these variables are most susceptible to the effects of 6P breathing and are the best proxies for baroreflex activation. HF HRV was used to capture vagal influences on HRV. The rest of the variables were computed for possible non-planned examinations of other cardiovascular mechanisms driving interoceptive awareness.

2.5. Data analysis

2.5.1. Proposed analyses

Prior to analysis, respiration data were examined. Participants with a peak respiration power outside the resonance breathing frequency range (0.1 Hz; 0.075–0.108 Hz) during the 6P task were excluded, as that implied the 6P task had not been performed correctly; this examination resulted in no exclusions.

Any technical errors or incidents with participants not complying with the tasks were registered, and their data excluded. Out of 67 subjects with a complete data set, two participants with technical errors in the collection of ECG data and one participant with missing interoception data at baseline were excluded. As shown in Fig. 2, the differences in observable effect sizes with the predicted vs. obtained sample sizes are negligible.

Fig. 2.

Power curves for each simulated dataset with N:50, 60 and 70. The curves display the power estimates for the main effect of Task (VP vs. 6P) and the interaction of PhysioΔ and Task over breath-modulated interoception.

According to the pre-registered analysis pipeline, participants with a negative d′ score on the Baseline discrimination task were to be excluded because of the difficulty in interpreting negative d′ scores (Sokol-Hessner et al., 2015). Interim baseline checks, however, revealed a large percentage of participants with negative d′ scores. Based on previous data, we expected ~12% of participants to show negative d′ scores, but, in the present sample, over 30% had negative d′ scores.

Data collection were suspended due to the Covid-19 pandemic. We attempted to resume collection in September 2020, but, for safety reasons, could not recruit a large enough sample within a reasonable time frame to obtain 70 participants with d′ scores above 0. We pre-registered a deviation from the analysis pipeline prior to finishing data collection and running any analyses. We specified that we would present the original analyses, only with participants with d′ ≥ 0 (n = 38, 26 participants excluded with d′ < 0), and a separate analysis setting the cut-off at d′ ≥ −2 (n = 63, one participant excluded with d′ < −4) (https://osf.io/vzdgs/registrations). Upon further consideration, the exclusion of all participants with d′ < 0 is not supported by signal detection theory, as negative d′ scores can reflect normal variability in responses around chance level.

2.5.2. Baseline predictions

Physiological variables (0.1 Hz HRV, HF HRV and BRS) were log10 transformed. Interoceptive variables were examined for skewness and transformed as needed. A regression analysis within the general linear modeling (PROC GLM; SAS v9.4) framework (Tatsuoka and Lohnes, 1988) was used to independently examine the 3 baseline interoceptive measures (accuracy, sensibility, metacognitive interoception) as DVs, with 0.1 Hz HRV, HF HRV and BRS during the vanilla task (Physio_B1) as predictors. Each initial model was examined for influential data points using Cook’s D and the PRESS statistic. If any observations were determined to be unduly influential for the regression model, they were removed from the analysis and the regression was rerun.

For each significant predictor, independent correlations with baseline interoceptive measures were performed. The differences between correlation coefficients were examined to compare the predictive power of each physiological variable on interoception. The t-statistic for the difference between correlations was computed (Steiger, 1980) out of which a SE of the mean difference was obtained; Bayes Factors (BF) (Dienes, 2014) further examined the sensitivity of the differences between correlations. Using data from Leganes-Fonteneau et al. (2019) on the relationship between interoception and subjective feelings of light-headedness, r(33) = 0.553, p = .001, and stimulation, r(33) = 0.043, p = .811, a prior was computed as the difference between correlations (0.510). BF modeling H1 with a Uniform from 0 (no correlation) to 0.510 were obtained.

As also specified in the pre-registration of the deviation from the original pipeline, BF also examined the sensitivity of all correlational data for baseline predictions, modeling H1 with a Uniform from 0 (no correlation) to 0.3. A BF above 3 provides evidence for H1 (significant differences between correlations), a BF below 1/3 provides sensitive evidence for H0 (no difference between correlations), and a BF = [1/3, 3] reveals that the comparison is insensitive.

2.5.3. Effects of resonance breathing

Linear mixed modeling (SAS PROC MIXED) was used to assess breath-modulated interoception (after the breathing task) as a function of baseline interoception measures, the breathing task performed (6P = 1, VP = 0), and the difference scores (PhysioΔ) of physiological responses observed during 6P and VP (Physio_6p – Physio_VP). The analyses were performed independently for 0.1 Hz HRV, HF HRV and BRS. BF examined the sensitivity of the main effect of the breathing task on each interoceptive measure, modeling H1 according to previous data (Leganes et al., 2021).

A “top-down” approach (West et al., 2014) was used to build each of the final linear mixed models by testing two different models: one that contained all main effects plus the PhysioΔ * Paced Breathing task interaction, which was the primary interaction of interest, and a second exploratory model that contained all possible two- and three-way interactions. Each model was examined for influential data points using Cook’s D and the PRESS statistic, and influential data points were removed. F values were used to determine significance of interactions and main effects. Post-hoc tests were conducted using complex contrast tests for any combination of statistically significant interactions and main effects (see Appendix A1 for a detailed methodological description of the analysis).

An exploratory analysis added fitness level as a covariate to the main model, with the expectation that the model would remain significant accounting for fitness level. In case fitness levels reduced the significance of the model, we introduced fitness level as an interaction term in the model to examine its effects.

2.6. Power analysis

A power analysis determined the adequate sample size for this experiment. Simulated data based on the hierarchical linear modeling statistical framework (Raudenbush and Bryk, 2002) was obtained using R. All variables were generated with a normal distribution (Mean = 0, SD = 1) to yield a standardized solution for the final mixed model. The relationship between breath-modulated interoception and the breathing task (dummy coded, VP vs. 6P) was specified at level 1 with a simulated error distribution of 0.5. This generated a slope corresponding to a moderate effect size (r_effect ≈ 0.4) (Rosnow et al., 1996). The subject level predictors of baseline interoception and PhysioΔ on breath-modulated interoception were specified at level 2, generated from a multivariate normal distribution consisting of a variance of 1 for both predictors and a covariance of 0.1 between them.

We generated a moderate relationship between breath-modulated interoception and baseline interoception (r ≈ 0.50). For 6P, we generated a moderate relationship between breath-modulated interoception and PhysioΔ (r ≈ 0.40). For VP, we specified no relationship between breath-modulated interoception and PhysioΔ. Using these parameters, we obtained three datasets with N: 50, 60 and 70, see Appendix A2 for detailed variance-covariance matrices.

The power for the independent variables was assessed using the SAS MIXED procedure (Littell et al., 2006). A power curve as a function of effect size was generated for Task and the Task × PhysioΔ interaction for each dataset (see Fig. 2), and effect size estimates at power = 0.90 were obtained. For N = 50: task = 0.38, PhysioΔ × Task = 0.51; for N = 60: task = 0.35, PhysioΔ × Task = 0.45; for N = 70: Task = 0.32, PhysioΔ × Task = 0.43. It was decided therefore to recruit 70 participants.

2.7. Reliability analyses

The reliability of physiological (0.1 Hz HRV, HF HRV, BRS) and baseline interoception measures (accuracy, sensibility and metacognitive interoception) were examined. Pearson’s correlations were performed on physiological variables obtained during the baseline Vanilla task in sessions 1 and 2. For interoceptive measures, Pearson’s correlations were performed between baseline interoception and interoception measured after the VP breathing task.

3. Results

3.1. Baseline predictions

For participants with a d′ score at baseline ≥0, initial quality checks identified different influential data points for each model (see Appendix A2). Any missing data on interoception was handled using case-wise deletion. For interoceptive accuracy, the regression analysis including all three physiological indices was non-significant, F(3,36) = 2.40, p = .0857. Statistical analyses showed a very low tolerance in the model (<0.42) due to the very high multi-collinearity of the BRS and HF HRV predictors (r = 0.75).

For interoceptive sensibility, the regression analysis was non-significant, F(3,36) = 1.22, p = .317.

For metacognitive interoception, the regression analysis was significant, F(3,37) = 4.14, p = .013. 0.1 Hz HRV was a significant predictor of metacognitive interoception, B = 0.105, t(37) = 2.07, p = .046. The rest of the predictors were non-significant, but again there was strong multi-collinearity between predictors and the low tolerance in the model (<0.35).

3.2. Deviation from proposed analyses due to multicollinearity

In light of the strong multi-collinearity affecting all models, we independently examined the relationship between each physiological variable and baseline interoception measures using independent Pearson’s correlations, see Table 1 and Fig. 3. Interoceptive accuracy negatively correlated with BRS and HF HRV, whereas metacognitive interoception positively correlated with 0.1 Hz HRV and BRS. The rest of the results were insensitive.

Table 1.

Correlations between physiological measures at baseline and indices of interoceptive awareness.

		0.1 Hz HRV	BRS	HF HRV
Interoceptive Accuracy	r(37)	−0.30853	−0.40986	−0.32835
	p	0.0632	0.0118	0.0472
	BF U[0,0.3]	3.19	7.26	3.41
Interoceptive Sensibility	r(37)	0.09345	0.22795	0.04694
	p	0.5822	0.1748	0.79
	BF U[0,0.3]	0.99	2.13	0.79
Metacognitive interoception	r(38)	0.45207	0.40514	0.26017
	p	0.0044	0.0116	0.1147
	BF U[0,0.3]	9.73	7.05	2.61

Results presented only for participants with d′ ≥ 0 at baseline, n = 38.

Bold marks Bayes Factors providing sensitive evidence for the alternative hypothesis.

Fig. 3.

Correlations between each interoceptive awareness measure and physiological predictors at baseline. Log HRV indices reflect peak power at 0.1 Hz and total power in the HF range of the RRI power spectrum. BRS is indexed as s/mmHg. + indicates a significant correlation. Results presented only for participants with d′ ≥ 0 at baseline.

The examination of the relative strength of significant predictors showed a sensitively null difference between BRS and HF HRV as predictors of interoceptive accuracy, BF_U[0,0.51] = 0.143. The rest of the comparisons were insensitive, BF > 1.011.

3.3. Registered deviation to include d′ ≥ −2

In light of the large number of participants with d′ < 0, we repeated all analyses including participants with d′ ≥ −2 as a pre-registered deviation.

For interoceptive accuracy, the pattern of results was parallel to those for the originally proposed analyses, except that the regression was significant, F(3,60) = 3.31, p = .026. Statistical analyses again showed a low tolerance in the model (<0.31) due to the very high multi-collinearity of the BRS and HF HRV predictors (r = 0.75). In this case, however, BRS was the only significant negative predictor, B = −0.467, t(61) = −2.15, p = .036.

For interoceptive sensibility, the regression analysis was non-significant, F(3,60) = 0.80, p = .499.

For metacognitive interoception, the regression analysis was significant, F(3,61) = 5.24, p = .002. 0.1 Hz HRV was a significant predictor of metacognitive interoception, B = 0.423, t(61) = 2.96, p = .005. The rest of the predictors were non-significant, but again results are unclear due to the strong multi-collinearity between predictors and the low tolerance in the model (<0.35).

We re-examined the relationship between physiological variables and interoception using independent Pearson’s correlations, see Table 2 and Fig. 4. As before, interoceptive accuracy negatively correlated with BRS and HF HRV. In this larger sample, metacognitive interoception positively correlated with HF HRV as well as 0.1 Hz HRV and BRS. The rest of the results were insensitive.

Table 2.

Correlations between physiological measures at baseline and indices of interoceptive awareness.

		0.1 Hz HRV	BRS	HF HRV
Interoceptive Accuracy	r(61)	−0.037	−0.333	−0.247
	p	0.773	0.009	0.055
	BF U[0,0.3]	0.66	12.25	4.28
Interoceptive Sensibility	r(61)	0.088	0.185	0.116
	p	0.499	0.644	0.607
	BF U[0,0.3]	0.97	2.22	1.21
Metacognitive interoception	r(60)	0.452	0.303	0.276
	p	0.0002	0.017	0.030
	BF U[0,0.3]	50.96	8.15	5.90

Results presented for participants with d′ ≥ −2 at baseline, n = 63.

Bold marks Bayes Factors providing sensitive evidence for the alternative hypothesis.

Fig. 4.

Correlations between each interoceptive awareness measure and physiological predictors at baseline. Log HRV indices reflect peak power at 0.1 Hz and total power in the HF range of the RRI power spectrum. BRS is indexed as s/mmHg. + indicates a significant correlation. Results presented for participants with d′ ≥ −2 at baseline.

The examination of the relative strength of significant predictors showed a sensitively null difference between BRS and HF HRV as predictors of interoceptive accuracy, BF_U[0,0.51] = 0.178. The rest of the comparisons were insensitive, BF > 0.43.

3.4. Effects of resonance breathing

For breath-modulated interoceptive accuracy, we found higher scores in the VP condition compared to the 6P condition, F(1,31) = 8.03, p = .0080, η² = 0.214, BF_U[0,1] = 10.46. Visual inspection of the data suggests that this may be driven by a few individuals who showed a strong response to the control VP task (see Fig. 5). We found no significant main effect of baseline accuracy on breath-modulated interoceptive accuracy, regardless of the breathing condition, F(1,31) = 1.61, p = .211, η² = 0.049. Results for the linear mixed models examining the interaction between the breathing task and PhysioΔ are presented on Table 3. In the 6P condition, we found a significant positive relationship with 0.1 Hz HRV PhysioΔ (Fig. 6A).

Fig. 5.

Individual interoceptive accuracy scores during baseline, 6P and VP. Results presented only for participants with d′ ≥ 0 at baseline.

Table 3.

Results of linear mixed models examining the effect of paced breathing on interoceptive accuracy, interoceptive sensibility and metacognitive interoception as a function of the effects on PhysioΔ.

PhysioΔ	df	F	p	η2
Interoceptive accuracy
0.1 Hz HRV	1, 31	8.03	0.008	0.205
	B	t	p
6P	0.777	2.6	0.014
VP	−0.342	−1.08	0.287
BRS	1, 29	3.27	0.081	0.101
	B	t	p
6P	2.79	2.17	0.039
VP	−0.005	0.01	0.996
HF HRV	1, 31	4.35	0.045	0.123
	B	t	p
6P	0.709	1.92	0.644
VP	−0.005	−0.02	0.985
Interoceptive sensibility
0.1 Hz HRV	1, 30	7.87	0.009	0.208
	B	t	p
6P	7.057	1.91	0.066
VP	−6.662	−2.18	0.037
BRS	1, 30	0	0.996	0
	B	t	p
6P	−4.135	−0.38	0.709
VP	−4.065	−0.73	0.471
HF HRV	1, 31	1.73	0.198	0.053
	B	t	p
6P	6.802	2.22	0.034
VP	1.262	0.4	0.691
Metacognitive interoception
0.1 Hz HRV	1, 33	0	0.967	0.028
	B	t	p
6P	−0.063	−1.01	0.32
VP	−0.067	−1.25	0.218
BRS	1, 33	2.38	0.133	0.67
	B	t	p
6P	−0.364	−2.49	0.018
VP	−0.037	−0.22	0.824
HF HRV	1, 31	0.56	0.46	0.018
	B	t	p
6P	−0.006	−0.11	0.911
VP	0.055	0.95	0.35

Results presented only for participants with d′ ≥ 0 at baseline.

Bold marks p-value below <0.05.

Fig. 6.

Simple slope correlations between breath-modulated interoceptive accuracy and 0.1 Hz HRV PhysioΔ (A), and between breath-modulated interoceptive sensibility and 0.1 Hz HRV PhysioΔ (B). + indicates a significant correlation. Results presented only for participants with d′ ≥ 0 at baseline.

For breath-modulated interoceptive sensibility, we found a significant main effect of interoceptive sensibility at baseline, F(1,30) = 120.01, p < .0001, η² = 0.800, but found no sensitive main effect of paced breathing task, F(1,30) = 3.21, p = .083, η² = 0.096, BF_U[0,7] = 2.56. The linear mixed model revealed a negative relationship between 0.1 Hz HRV PhysioΔ and interoceptive sensibility during the VP condition (Fig. 6B).

For breath-modulated meta-cognitive interoception, we found no sensitive main effect of paced breathing task, F(1,33) = 0.72, p = .401, η² = 0.021, BF_U[0,0.7] = 0.42, nor a main effect of for metacognitive interoception at baseline, F(1,33) = 0.18, p = .672, η² = 0.005, or significant interactions in the linear mixed models (see Table 3).

Exploratory two- and three-way interactions were non-significant, ps > .07. See Table 4 for raw HRV and BRS values in different conditions.

Table 4.

Raw 0.1 Hz HRV peak, HF-HRV power and BRS in different conditions.

	6P		Baseline 6P		VP		Baseline VP
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
0.1 Hz-HRV (ms2)	780,254.00	507,649.21	32,581.90	39,997.78	29,551.66	49,727.86	45,269.45	63,430.39
BRS (ms/mmHg)	15.30	7.22	12.23	6.37	11.93	7.49	13.27	6.6
HF-HRV (ms2/Hz)	1136.61	1804.45	1402.11	3488.11	3088.47	4739.24	1303.42	1930.45

Results presented only for participants with d′ ≥ 0 at baseline.

3.5. Registered deviation to include d′ ≥ −2

For breath-modulated interoceptive accuracy, we found higher scores in the VP condition compared to the 6P condition, F(1,53) = 7.70, p = .0076, η² = 0.126, BF_U[0,1] = 15.37, although again, visual inspection of the data suggests that this may be driven by a few individuals who showed a strong response to the control VP task (see Fig. 7). We found this time a significant main effect of baseline accuracy on breath-modulated interoceptive accuracy, F(1,53) = 4.17, p = .046, η² = 0.072. Results for the linear mixed models examining the interaction between the breathing task and PhysioΔ are presented on Table 5. In the 6P condition, we found again a significant positive relationship with 0.1 Hz HRV PhysioΔ, and also this time with BRS PhysioΔ (Fig. 8A–B). Exploratory analyses observed a significant 3-way interaction using 0.1 Hz PhysioΔ as a covariate, F(1,47) = 5.41, p = .024. As observed in Fig. 9, the positive relationship between 0.1 Hz PhysioΔ and breath-modulated interoceptive accuracy is progressively stronger for participants who have higher interoceptive accuracy at baseline.

Fig. 7.

Individual interoceptive accuracy scores during baseline, 6P and VP. Results presented for participants with d′ ≥ −2 at baseline.

Table 5.

Results of linear mixed models examining the effect of breathing on interoceptive accuracy, interoceptive sensibility and metacognitive interoception as a function of the effects on PhysioΔ.

PhysioΔ	df	F	p	η2
Interoceptive accuracy
0.1 Hz HRV	1, 53	7.67	0.008	0.126
	B	t	p
6P	0.527	2.34	0.023
VP	−0.294	−1.65	0.104
BRS	1, 52	5.22	0.026	0.091
	B	t	p
6P	2.102	2.13	0.038
VP	−1.038	−1.26	0.213
HF HRV	1, 52	0.97	0.33	0.018
	B	t	p
6P	0.149	0.52	0.605
VP	−0.177	−0.81	0.422
Interoceptive sensibility
0.1 Hz HRV	1, 55	4.34	0.042	0.073
	B	t	p
6P	5.28	2.01	0.049
VP	−0.78	−0.78	0.438
BRS	1, 54	0.02	0.876	0
	B	t	p
6P	−3.865	−0.49	0.626
VP	−2.405	−0.49	0.624
HF HRV	1, 54	4.45	0.039	0.076
	B	t	p
6P	5.759	2.81	0.007
VP	0.658	0.51	0.615
Metacognitive interoception
0.1 Hz HRV	1, 53	0.27	0.61	0.005
	B	t	p
6P	−0.046	−0.94	0.35
VP	−0.05	−0.47	0.64
BRS	1, 55	2.17	0.146	0.038
	B	t	p
6P	−0.24	−2.05	0.045
VP	0.006	0.05	0.962
HF HRV	1, 53	0.97	0.33	0.018
	B	t	p
6P	−0.027	−0.61	0.546
VP	0.096	2.15	0.037

Results presented for participants with d′ ≥ −2 at baseline.

Bold marks p-value below <0.05.

Fig. 8.

Simple slope correlations between breath-modulated interoceptive accuracy and 0.1 Hz HRV PhysioΔ (A), BRS (B) and between breath-modulated interoceptive sensibility and 0.1 Hz HRV PhysioΔ (C) and HF HRV (D). + indicates a significant correlation. Results presented for participants with d′ ≥ −2 at baseline.

Fig. 9.

Relationship between breath-modulated interoceptive accuracy and 0.1 Hz PhysioΔ as a function of baseline interoceptive accuracy. The relationship is stronger for participants with higher baseline interoceptive accuracy and non-existent for those with low baseline interoceptive accuracy. Results presented for participants with d′ ≥ −2 at baseline.

For breath-modulated interoceptive sensibility, we found a significant main effect of interoceptive sensibility at baseline, F(1,55) = 183.50, p < .0001, η² = 0.769, but found no sensitive main effect of paced breathing task, F(1,55) = 3.10, p = .084, η² = 0.053, BF_U[0,7] = 2.99. The linear mixed model revealed this time in the 6P condition a significant positive relationship with 0.1.Hz HRV and HF HRV PhysioΔ (Fig. 8C–D), although these findings were not observed in the pre-registered analysis pipeline. Exploratory analyses observed a significant 3-way interaction using BRS PhysioΔ as a covariate, F(1,48) = 7.74, p = .008, that revealed no theoretically relevant effects.

For breath-modulated meta-cognitive interoception, we found no sensitive main effect of paced breathing task, F(1,53) = 0.93, p = .340, η² = 0.017, BF_U[0,0.7] = 0.39, nor a main effect of metacognitive interoception at baseline, F(1,53) = 0.27, p = .607, η² = 0.005, or significant interactions in the linear mixed models. Exploratory analyses observed a significant 2-way interaction between metacognitive interoception at baseline and paced breathing task, F(1,49) = 5.12, p = .028, that revealed no theoretically relevant effects.

Finally, we did not observe any changes in the significance of the linear mixed models when including physical activity as a covariate. See Table 6 for raw HRV and BRS values in different conditions.

Table 6.

Raw 0.1 Hz HRV peak, HF-HRV power and BRS in different conditions.

	6P		Baseline 6P		VP		Baseline VP
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
0.1 Hz-HRV (ms2)	735,021.86	508,306.25	29,735.75	37,944.98	24,366.82	40,196.25	43,653.31	82,495.89
BRS (ms/mmHg)	15.07	7.05	11.95	6.17	11.90	7.79	14.32	11.38
HF-HRV (ms2/Hz)	978.00	1501.47	1323.90	2878.74	3093.38	4613.74	1654.90	3675.84

Results presented only for participants with d′ ≥ −2 at baseline.

3.6. Reliability analyses

For participants with d′ ≥ 0, the relationship between baseline physiology at session 1 and 2 showed a good test-retest reliability for 0.1 Hz HRV, r(38) = 0.662, p < .0001, BRS, r(38) = 0.680, p < .0001, and HF HRV r(38) = 0.748, p < .0001. The relationship between baseline interoception measures and interoception measures after VP, test-retest reliability was high for interoceptive sensibility, r(37) = 0.831, p < .0001, but low for interoceptive accuracy, r(37) = −0.264, p = .114, and metacognitive interoception, r(37) = 0.033, p = .846.

In parallel, for participants with d′ ≥ −2, the relationship between baseline physiology at session 1 and 2 showed a good test-retest reliability for 0.1 Hz HRV, r(63) = 0.662, p < .0001, BRS, r(63) = 0.565, p < .0001, and HF HRV r(63) = 0.708, p < .0001. For the relationship between baseline interoception measures and interoception measures after VP, test-retest reliability was high for interoceptive sensibility, r(63) = 0.812, p < .0001, but low for interoceptive accuracy, r(63) = 0.040, p = .758, and metacognitive interoception, r(63) = 0.020, p = .880. Although in both samples the test-retest reliability appears to be variable for different dimensions of interoceptive awareness, the significant main effect of baseline accuracy on breath-modulated accuracy supports moderate test-retest reliability when accounting for experimental covariates.

4. Discussion

Please note that this discussion focuses on the larger, d′ ≥ −2 sample results.

In this registered report, our first goal was to examine whether cardiovascular functioning could predict individual differences in interoceptive awareness, as measured with the cardiac discrimination task. We proposed a specific role for the baroreflex mechanism on interoception and thus predicted that 0.1 Hz HRV and BRS, but not HF HRV, would be positively correlated with three measures of interoceptive awareness: accuracy, sensibility, and metacognitive interoception. Contrary to our hypotheses, we found that baseline interoceptive accuracy negatively, rather than positively, correlated with BRS and HF HRV, and was not sensitively related to 0.1 Hz HRV. Also in contrast to our predictions, interoceptive sensibility was unrelated to cardiovascular physiology. On the other hand, as predicted, metacognitive interoception positively correlated with physiology, but this relationship was not specific to 0.1 Hz HRV and BRS as we expected. The high multi-collinearity of cardiovascular measures disrupted the ability to distinguish which specific cardiovascular mechanisms correlate with interoception.

The second goal was to determine whether interoceptive awareness could be modified with an experimental manipulation that activates the baroreflex mechanism, called resonance breathing or 6-breaths-per-minute breathing (6P). Contrary to our prediction, we did not find a positive main effect of 6P on performance in the heartbeat discrimination task, in line with recent null-findings (Rominger et al., 2021). However, the examination of concomitant physiological changes shows, as predicted, that increases in 0.1 Hz HRV and BRS after the 6P task correlated with increases in interoceptive accuracy. In other words, the extent to which the breathing task engages the resonant properties of the cardiovascular system and activates the baroreflex improves participants’ ability to feel their own heart. We also found that increases in 0.1 Hz HRV after the 6P task correlated with interoceptive sensibility; this positive relationship was also unexpectedly observed with HF HRV. There were no significant relationships between physiological changes during resonance breathing and metacognitive interoception.

Taken together, the results partially support our hypotheses, but the complex nature of the relationships and the pattern of results across physiological indices and interoceptive awareness measures raises additional questions beyond those explained by our originally proposed framework. Below, we interpret the baseline and breathing-modulated interoception analyses separately within the proposed interpretation. In addition, we put forth an alternate interpretation grounded in interoceptive inference perspectives that integrates the results of the breathing-modulated interoception task with the negative correlations between interoceptive accuracy and physiology that were observed at baseline.

4.1. Interoceptive awareness and resonance breathing: Increased afferent signaling

To the best of our knowledge, this is the first report directly examining how baroreflex functioning shapes interoceptive awareness. Multiple seminal reports have pointed towards baroreceptor firing as a source of cardiac interoceptive signals (Critchley and Garfinkel, 2015; Garfinkel et al., 2013), however experimental evidence for the direct involvement of baroreceptor functioning in interoceptive awareness is limited. The present study provides such evidence by assessing changes in interoceptive awareness following a task that directly affects the baroreflex. We interpret changes in 0.1 Hz HRV and BRS during resonance breathing as evidence of direct modulation of viscero-afferent cardiac signaling. Resonance breathing increases the amplitude of heart rate oscillations specifically at 0.1 Hz (Song and Lehrer, 2003; Vaschillo et al., 2002); these oscillations reflect the afferent components of the baroreflex loop (Reyes del Paso et al., 2013) and such physiological mechanism could have improved interoceptive accuracy by facilitating the detection of afferent cardiac signals. Resonance breathing also increases the sensitivity of the baroreflex (Bates et al., 2019; Lehrer et al., 2003), observable as a greater blood pressure response to a given change in heart rate. This provides evidence that baroreceptors become more sensitive to changes in heart rate during and for some period following resonance breathing, which, in this study, resulted in improved interoceptive accuracy. This perspective aligns with the observation that resonance breathing improves the strength of heartbeat evoked potentials (MacKinnon et al., 2013) as a source of increased vagal information that is then processed in the insula, the interoceptive cortical hub (Lehrer and Gevirtz, 2014). More generally, it suggests that interoceptive accuracy can be improved by altering viscero-afferent cardiac signaling.

We also observed that increases in 0.1 Hz HRV and HF HRV correlate with changes in interoceptive sensibility. 0.1 Hz and HF HRV indices are thought to reflect different underlying physiological mechanisms (i.e., baroreflex and vagal, respectively, Bates and Buckman, 2013) and 0.1 Hz HRV is thought to more accurately reflect afferent communication (Reyes del Paso et al., 2013). The absence of a relationship with BRS draws into question the possibility that the physiological mechanism responsible for the increase in interoceptive accuracy also shapes increased perceptions of confidence. Rather, the observation that changes in both measures of HRV are linked to an individual’s interoceptive confidence may suggest a more general role for cardiovascular dynamics at play. Resonance breathing has a variety of cognitive implications, particularly regarding task performance and stress (i.e. Charles and Nixon, 2019); and this could account for the fact that the cardiac discrimination task may not sufficiently differentiate interoceptive confidence from self-confidence. This hypothesis warrants further testing.

4.2. Interoceptive inference perspective: a post-hoc interpretation

An important caveat to the above interpretation is that BRS and HF HRV were negatively correlated with interoceptive accuracy at baseline, and accuracy was not sensitively related to 0.1 Hz HRV. We expected 0.1 Hz HRV and BRS to positively predict interoceptive abilities, reflecting higher viscero-afferent signaling, and no correlation with HF HRV. When considered together with the results from the breathing-modulated interoception task, this could support a role for predictive coding and an interoceptive inference perspective (Seth, 2013; Seth and Friston, 2016). This perspective suggests that neural systems seek to establish an accurate representation of exteroceptive, proprioceptive and interoceptive states through Bayesian inference processes.

In the context of interoceptive inference, neurally encoded probability distributions about physiological states serve as a model or “belief” about the internal condition of the body. These predictions are updated by continuous afferent physiological signals generating interoceptive prediction errors (Barrett, 2017) that in turn modify the model or “beliefs” about bodily states (Friston, 2018; Petzschner et al., 2021). Importantly, Bayesian models take into account the reliability of prediction errors, understood as the inverse variance of the signal, and only highly reliable prediction errors can successfully update prior expectations and resolve uncertainty (Stephan et al., 2016).

Applying this framework to our findings, it follows that lower HRV (i.e., less variability in heart rate) equates with a less “noisy” and more predictable stream of cardiovascular afferent information. With smaller and more reliable prediction errors, participants with lower HRV then may be more effective at generating an accurate model. Since the accuracy of predictive models and the extent to which they can suppress prediction errors is hypothesized to affect the conscious processing of cognitive content (Hohwy, 2014; Seth et al., 2016), a more accurate model of interoceptive states due to lower HRV would reduce interoceptive prediction errors and in turn enable the conscious perception of cardiac signals, as measured through interoceptive accuracy.

This interpretation is in line with recent evidence showing that lower HRV facilitates task performance in conditions of attentional uncertainty (Corcoran et al., 2021) and the integration of cardiac signals in response inhibition (Rae et al., 2018). Accordingly, cardiac stabilization would increase the predictability of subsequent baroreceptor firing, facilitating the integration of interoceptive signals in cognitive task performance. We posit that cardiac stabilization may not only facilitate the processing of exteroceptive information, but also interoceptive cardiac signals. This view is particularly relevant for the heartbeat discrimination task, which is based on the coupling of exteroceptive (the beep) and interoceptive signals; more stable RRIs and reduced variability in HR could facilitate the integration of both signals through smaller interoceptive prediction errors (Corcoran et al., 2021).

Interoceptive inference perspectives can also explain the correlation between changes in 0.1 Hz HRV during resonance breathing and increases in interoceptive accuracy. In this case, we argue that although resonance breathing increases HRV, it does so only within a very narrow bandwidth (Vaschillo et al., 2011). Thus, this does not represent a broad spectral increase in HRV that would entail a higher degree of afferent signal “noise”, but rather reflects a massive and targeted increase in the amplitude of cardiac oscillations at 0.1 Hz, tightening cardiovascular and respiratory dynamics within a predictable oscillatory pattern (see Fig. 10 for a visualization of the effects of resonance breathing on oscillatory RRI patterns). This reorganization of HRV increases the predictability of afferent baroreceptor signaling, which would then improve interoceptive accuracy. If the extent to which interoceptive prediction errors can effectively update priors depends on the variance of the signal, arranging RRIs within a predictable oscillatory pattern would enact an increase in the predictability of said prediction errors, resulting in an improved model for cardiac states and an increased conscious perception of interoceptive signals.

Fig. 10.

Representation of the effects of resonance breathing on cardiovascular dynamics across frequency and time domains.

Importantly, this view is not at odds with the finding that the effects of resonance breathing on BRS mediate increases in interoceptive accuracy. Multiple redundant mechanisms exist within the cardiovascular system to allow the maintenance of systems critical for survival (Vaschillo et al., 2012), and similar mechanisms may be at play here. Previous research has also shown that resonance breathing increases the strength of afferent vagal communication, as measured through heartbeat evoked potentials (MacKinnon et al., 2013). Moreover, animal (Corcoran et al., 2018) and human research (Johannknecht and Kayser, 2021) highlights the cross-modal integration of respiratory oscillations with the active sampling of sensory stimuli. It is possible that the neural oscillations generated by resonance breathing further interact with the perception of heartbeats, as oscillatory brain patterns would synchronize with the resonant properties of the cardiovascular system. The re-organization of RRI across a predictable oscillatory pattern together with the increased strength of the signal and the introduction of cross-modal respiratory oscillations would further reduce interoceptive prediction errors. We therefore hypothesize multiple pathways that ensure the integration of viscero-afferent responses in the generation of interoceptive experiences.

4.3. Interoceptive inference as a conceptual framework for resonance breathing

The general consensus in the field is that resonance breathing positively influences cognition, emotion and mental health through the improvement of regulatory feedback loops, the increase in afferent vagal signaling, and/or the improvement in cerebral blood flow (Lehrer and Gevirtz, 2014). There is, however, limited evidence as to whether resonance breathing instigates transient or long-term residual changes in the cardiovascular system (Lehrer et al., 2003); this raises questions about the mechanism by which resonance breathing exerts its therapeutic effects. We tentatively propose that one mechanism by which resonance breathing could promote health is through the systematic and effective reduction of interoceptive prediction errors. This hypothesis, if demonstrated, has substantial therapeutic implications.

Faulty integration of interoceptive prediction errors may contribute to a variety of mental health issues, such as addiction, anxiety, alexithymia, or autism (Khalsa et al., 2017; Seth and Friston, 2016). Individuals with these conditions might present hyper-precise priors about their interoceptive state and a difficulty in adjusting those beliefs in the presence of novel afferent information that generates interoceptive prediction errors (Paulus et al., 2019). Interestingly, these populations also traditionally display reduced indices of HRV, possibly representing an allostatic down-regulation strategy to improve the predictability of afferent cardiac signals and to reduce their interference with exteroceptive processing (Ottaviani, 2018; Peters et al., 2017), enabling them to cope with the effects of environmental uncertainty. We argue that resonance breathing provides a structured and predictable framework of HRV oscillations that reduces interoceptive prediction errors and facilitates interoceptive inference processes. In doing so, the integration of interoceptive and exteroceptive information is bolstered and the individual becomes more physically and psychologically adaptable. Importantly, resonance breathing is easy to learn, simple to perform in almost any context, and requires no expensive equipment. The number of health conditions in which interoceptive dysfunction is implicated and the scalability of resonance breathing strong support further research in this area.

5. Limitations

As disclosed in the results section, we partially deviated from the registered data analysis pipeline. Due to Covid-19, we were not able to test as many participants as we had planned. In addition, we did not anticipate that excluding all participants with d′ < 0 at baseline would constitute such a stringent cut-off. Notwithstanding, we pre-registered the change in inclusion criteria before performing the analyses and results are mostly equivalent between models that included participants with d′ ≥ 0 or d′ ≥ −2 at baseline. We also disclosed a post-hoc deviation from the registered analysis pipeline due to the extreme multicollinearity of the physiological predictors; in hindsight, we should have predicted this collinearity during Stage 1. These changes are clearly labelled as unplanned. As we only assessed three physiological indices, risk of alpha inflation from multiple testing was modest.

The explanation we have delineated to explain the inverse relationship between HRV and interoceptive accuracy does not reconcile with the results related to metacognitive interoception. At baseline, metacognition was correlated with all three physiological indices; however, it responded to neither breathing task. There is limited experimental research related to metacognitive processing of interoception; this requires further research especially considering the results obtained for interoceptive sensibility.

The task used in this experiment has been subject to methodological criticism (Brener and Ring, 2016), and despite finding that interoceptive accuracy at baseline significantly predicted performance after breathing, test-retest reliability was low. This may have been due in part to heterogeneity in individuals’ responses to paced breathing; test-retest reliability is best assessed under stationary conditions. A method of constant stimulus would have produced a more precise characterization of interoceptive accuracy, but its lengthy application (~1 h) makes it difficult to apply in this experimental context. Future research should seek to replicate the observed effects using a more reliable heart rate discrimination task, such as a novel task developed after this report was pre-registered (Legrand and Allen, 2021). In addition, other tasks that are less susceptible to subjective inferences of cardiac awareness may more directly examine the role of cardiac signaling in emotion (Leganes et al., 2021; Leganes-Fonteneau et al., 2020b) and cognition (Rae et al., 2018) in response to resonance breathing. Nonetheless, our results support a strong physiological basis for the cardiac discrimination task.

Finally, the theoretical interpretation presented in the discussion regarding the interoceptive inference framework was not anticipated in the introduction and is a post-hoc contribution based upon the study’s novelty and its unexpected results. We believe that it provides a heuristic interpretation of our results and fits well with the accumulating evidence towards the interoceptive inference framework as a whole. It also may present an opportunity to further our understanding of the therapeutic effects of resonance breathing. This hypothesis, as already mentioned, needs to be further examined empirically using other measures of interoception.

6. Conclusion

This registered report brings some of the first evidence for the involvement of baroreflex functioning in cardiac interoceptive processes. The finding that it is possible to modulate interoceptive awareness using a simple paced breathing task opens the possibility to examine the clinical utility of resonance breathing in mental health disorders characterized by interoceptive deficits. It may also serve as a new tool to modulate interoceptive responses non-invasively in experimental settings. We put forth the idea that the well-replicated benefits of resonance breathing on cognition, emotion and mental health can be understood, at least in part, through an interoceptive inference perspective. This hypothesis bridges the two seemingly disconnected experimental and theoretical frameworks of interoception and basic psychophysiology and strongly encourages expanded integrated examinations.

Appendix A1

Regarding the detection of outliers, any Cook’s D_i > 1 were automatically considered an influential data point. For values of Cook’s D < 1, a criteria of $D_i>\frac{4}{(n-k-1)}$ , where i = individual subject index, n = total number of subjects, and k = number of regressors in the model, was used to flag influential data points (Fox, 2008). The PRESS statistic was graphically examined using an index plot to determine if there were any unduly large values relative to the entire sample, especially values in the top 5% (Fox, 2008)

The linear mixed model was built using a “top-down” strategy as delineated by West et al. (2014). The first step involved specifying a linear mixed model with all variables and their interactions of interest. In the present case, there were two different models of interest. The first model was a parsimonious model with three main effects of the Paced Breathing Condition (6P vs. VB), Baseline interoception, and the difference scores (PhysioΔ) in physiological responding (0.1 Hz HRV and BRS) observed during VP and 6P (Physio_6p – Physio_VP). The interaction of (PhysioΔ * Paced Breathing Condition) was also specified. In a more comprehensive exploratory model, all possible interactions among the three main effects were also included. The Paced Breathing Condition was specified in the model as a continuous dummy coded variable in order to facilitate the testing of the interactions of the Paced Breathing Condition with other variables and to more easily model its variance and covariance with other effects in the model.

The second step involved specifying the random effects in the model and the associated variance covariance (G) matrix, i.e., G-side variance covariance modeling. Subject was the nesting variable. The intercept was specified as a random effect in the initial version of the model. Models were tested specifying Paced Breathing Condition and/or PhysioΔ and/or Baseline Interception as random effects. For such models, variance components were initially specified as the structure for G. It should be noted that a linear mixed model with random effects may not have been able to successfully converge on a solution for a number of reasons, for instance, a given slope parameter may not have enough variance present to model as a random effect. In such a case, the variable was specified only as a fixed effect. If no regression parameters, including the intercept, could be modeled as random, then a marginal effects model was specified (see next step). If a model with two or more random effects could be specified, then a number of different structures for G were tested using REML-based likelihood ratio tests as recommended by West et al. (2014). The structures that were considered were: variance components (VC, the default for SAS PROC MIXED), compound symmetry (CS), heterogeneous CS (CSH), and unstructured (UN).

The third step involved investigating and specifying the structure of the variance covariance matrix of the residuals (R), i.e., R-side covariance modeling. In the present case, there were only two repeated sessions, so the only structures for R that were considered were the relatively simpler ones of VC, CS, CSH, and UN. If the previous step created a model with no random effects, then the present step created a marginal model, i.e., a population averaged model, in which only R-side covariance modeling is conducted for the repeated measures across time for each subject.

The fourth and last step involved using statistical tests and indices of fit to determine whether the mixed model could be reduced. Goodness of fit for the models will compared against each other, using fit statistics (AIC, AICC, BIC; the smaller the better) and the likelihood ratio test.

Once a final model was determined, then post-hoc analyses on individual group means and/or regression slopes were conducted using complex contrast tests. For instance, if the interaction of (PhysioΔ * Paced Breathing Condition) was found to be the only significant interaction, the individual slopes of interoception as a function of PhysioΔ for each paced breathing condition were computed, followed by a statistical test of the difference between the two slopes. Effect sizes were computed for each effect.

Appendix A2

Data were simulated for N: 50, 60 and 70, producing normalized datasets with a normal distribution (Mean = 0, SD = 1).

The lineplots present the simulated main effect of Task (VP vs. 6P) on breath-modulated interoception with a simulated error distribution of 0.5, see Fig. A1.

The variance covariance matrices show interrelationships among breath-modulated interoception, Baseline interoception and PhysioΔ for each task. Note that PhysioΔ is defined as a difference score in physiological responsiveness between the 6P and VP conditions, see Fig. A2.

Fig. A1.

Lineplot for each sample size, N: 50, 60 and 70. Effect of Task (0 = VP vs. 1 = 6P) on breath-modulated interoception.

Fig. A2.

Variance covariance matrix for each sample size, N:50, 60 and 70 showing the interrelationships between Baseline interoception PhysioΔ on breath-modulated interoception for each task.