Characterizing Interoceptive Differences in Autism: A Systematic Review and Meta-analysis of Case-control Studies

Abstract

Interoception, the body’s perception of its own internal states, is thought to be altered in autism, though results have been inconsistent. The current study systematically reviewed and meta-analyzed the extant literature comparing interoceptive outcomes between autistic (AUT) and neurotypical (NT) individuals, determining which domains of interoception demonstrate robust between-group differences. A three-level Bayesian meta-analysis compared heartbeat counting performance, heartbeat discrimination performance, heartbeat counting confidence ratings, and self-reported interoceptive attention between AUT and NT groups (15 studies; n_AUT=467, n_NT=478). Autistic participants showed significantly reduced heartbeat counting performance (g=−0.333, CrI_95% [−0.535, −0.138]) and higher confidence in their heartbeat counting abilities (g=0.430, CrI_95% [0.123, 0.750]), but groups were equivalent on other meta-analyzed outcomes. Implications for future interoception research in autism are discussed.

Individuals on the autism spectrum often experience differences related to sensory perception and reactivity, now considered to be a core feature of the condition (American Psychiatric Association, 2013; for reviews, see Ben-Sasson et al., 2009, 2019; Hazen et al., 2014; Proff et al., 2022). One sensory modality of emerging interest in autism research is interoception, the body’s perception of its own internal states (Bonaz et al., 2021; Chen et al., 2021; DuBois et al., 2016; Murphy et al., 2017; Proff et al., 2022; Quattrocki & Friston, 2014). Interoceptive signals such as heartbeat frequency, gut distention, and blood osmolality originate from within the body (Craig, 2002; Kleckner et al., 2017; Tsakiris & de Preester, 2019), and the body’s perception of these signals may play a key role in helping the brain regulate many vital processes such as digestion, heart rate, and respiration (DuBois et al., 2016; Khalsa et al., 2018; Quattrocki & Friston, 2014). As a result, interoception is often characterized as the “body to brain axis” and is theorized to play an important role in both homeostasis and emotional regulation (Adolfi et al., 2017; Garfinkel & Critchley, 2013; Seth, 2013).

Altered interoception has been implicated in the pathophysiology of psychiatric disorders and in the symptomatic expression of developmental, neurodegenerative, and neurological conditions, including autism (Bonaz et al., 2021; Brewer et al., 2021; Khalsa et al., 2018). In addition, reduced interoceptive ability has been hypothesized to serve as a pathophysiologic basis of alexithymia, a personality trait characterized by difficulty identifying emotional states in oneself or others (Brewer et al., 2016; Murphy, Catmur, et al., 2018; Trevisan et al., 2019). As levels of alexithymia are often substantially elevated in the autistic population (Kinnaird et al., 2019; Z. J. Williams & Gotham, 2021), interoceptive difficulties may be particularly relevant to understanding the socio-emotional differences present in many autistic individuals (Trevisan et al., 2021). Furthermore, interoceptive training has recently been tested as a novel intervention modality for reducing anxiety in autistic adults, with an initial study demonstrating modest benefits over a control intervention (Quadt et al., 2021). Given the many potential roles of interoception in autism and associated neuropsychiatric conditions, a deeper understanding of interoceptive differences in autism can potentially provide novel insights into several clinically relevant features of the condition.

When examining the ways in which autism is associated with interoceptive differences, studies to date have provided inconsistent and sometimes contradictory results (Suzman et al., 2021). Whereas some investigators have found that autistic people perform worse on objective tasks of interoceptive ability such as heartbeat counting (Garfinkel et al., 2016; Mul et al., 2018; Nicholson et al., 2019; Palser et al., 2018), others have systematically failed to replicate this group difference in similar samples (Failla et al., 2020; Nicholson et al., 2018, 2019; Pickard et al., 2020; Schauder et al., 2015). Furthermore, even when differences in heartbeat counting are evident, significant group differences have not been found on tasks assessing heartbeat discrimination (Garfinkel et al., 2016; Mul et al., 2018; Palser et al., 2018), suggesting that the group differences seen in the literature may be attributable to task-specific factors that do not reflect true interoceptive difficulties. Findings using self-report questionnaires are similarly inconsistent, with published studies indicating that self-reported attention to interoceptive cues in autism is diminished (Fiene & Brownlow, 2015; Mul et al., 2018; Pickard et al., 2020), elevated (Garfinkel et al., 2016), or no different (Palser et al., 2018) when compared to neurotypical controls. Given the discrepant findings regarding group differences in interoception in autism, quantitative synthesis of the literature is needed to reach more robust conclusions about the presence and/or directionality of these effects. Thus, the primary goal of the present study was to perform a systematic review and meta-analysis of case-control studies comparing interoceptive variables in autism, generating summary effects that characterize the magnitude and direction of interoceptive differences between autistic and neurotypical individuals.

Notably, interoceptive ability is not a unitary construct, and researchers have proposed a number of frameworks to better characterize subdimensions of interoception (e.g., Garfinkel et al., 2015; Murphy et al., 2019). One way to conceptualize interoception may be through the so-called “2×2” framework (Murphy et al., 2019), which considers both how interoception is measured (i.e., using an “objective” performance measure vs. a “subjective” self-report) and what aspect of interoception is measured (i.e., accuracy of interoceptive perceptions vs. attention toward interoceptive signals). Within this framework, measures of interoception can fall in to one of four cells in a 2×2 matrix (i.e., objective measure of accuracy, objective measure of attention, subjective measure of accuracy, subjective measure of attention). Although interoceptive domains have also been categorized within the more popular “tripartite” model of “interoceptive accuracy,” “interoceptive sensibility,” and “interoceptive awareness” (Garfinkel et al., 2015), this framework has recently been criticized for conflating multiple dimensions of self-reported interoception into a single construct of “interoceptive sensibility” (Bogaerts et al., 2022; Desmedt et al., 2021; Gabriele et al., 2022; Murphy et al., 2020; Trevisan et al., 2021). Thus, in order to provide a slightly more granular picture of interoceptive differences in autism, the current study adopted the 2×2 framework as a way of categorizing the interoceptive outcomes being compared between groups.

Methods

Identification and Selection of Studies

The procedures of this study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (Page, McKenzie, et al., 2021; Page, Moher, et al., 2021). The protocol for this meta-analysis was registered in PROSPERO (number CRD42020214237), although it is notable that our current analyses deviate significantly from those previously registered (e.g., the current study utilizes the 2×2 interoception framework rather than the tripartite framework to categorize outcomes; we also did not conduct a meta-analysis of correlations between interoception and other variables due to a lack of sufficient data). We searched PubMed, Web of Science, ProQuest, and Embase for peer-reviewed publications on interoception in autism using a combination of keywords and filters specific to each database (example search string for PubMed: Autis* OR ASD OR Asperger OR PDD-NOS) AND (interocept* OR ((heartbeat OR cardiac) NEAR/5 (counting OR tracking OR discrimination OR detection OR “evoked potentials”)) OR “body perception”); see Supplemental Table S1 for a complete list of search terms and specifications). A similar strategy was employed for a gray-literature search of ProQuest Dissertations and Theses, conference abstracts from the International Society for Autism Research Annual Meeting after 2004 (the earliest they were digitally available), and the reference list of a popular textbook on interoception (Tsakiris & de Preester, 2019). Eligible studies included peer-reviewed journal articles, conference posters and presentations, dissertations, and theses published in English before October 21, 2020 (the day of all database searches). Duplicates were removed in Microsoft Excel, double checked using visual inspection, and nonduplicate records were then transferred into the Abstrakr program (Rathbone et al., 2015). Potentially eligible studies were flagged in Abstrackr for full-text review if they were empirical research studies that reported on at least one measure of interoception in a group of participants diagnosed with autism. Two raters screened each abstract in Abstrackr, and agreement was quantified using the percentage of agreement and Krippendorff’s alpha coefficient calculated across all raters (Krippendorff, 2011). All disagreements were handled via consensus meeting between authors ES and ZJW.

Studies were included in the formal meta-analysis if they met the following criteria based on full-text review: (a) at least 10 participants diagnosed with autism, (b) a control group of neurotypical participants with which the autistic group’s scores could be compared, (c) at least one interoception measure (not including measures unique to a single study) that could be unambiguously classified within the 2×2 framework, and (d) reported statistics necessary for calculation of Hedges’ g for outcomes of interest (i.e., means and standard deviations, t- or F-statistics, or effect size measures that could be converted to g). Eligible measures of objective interoceptive accuracy included the heartbeat counting task (HCT; Schandry, 1981) and forced-choice heartbeat discrimination task (HDT; Whitehead et al., 1977), whereas eligible measures of subjective interoceptive accuracy included self-reported confidence in one’s HCT or HDT performance (confidence ratings; Hickman et al., 2020). Eligible measures of subjective interoceptive attention (sIAt) included the “awareness” subscale of the Porges Body Perception Questionnaire (BPQ; Porges, 1993) and derivative measures (Cabrera et al., 2018; Palser et al., 2018), as well as the “noticing” subscale of the Multidimensional Assessment of Interoceptive Awareness (MAIA; Mehling et al., 2012, 2018), and eligible measures of objective interoceptive attention included experience-sampling methods (Murphy et al., 2019). Corresponding authors of studies that met criteria for inclusion but that did not report sufficient statistics to derive all effect sizes were contacted via email to request additional data.

Data Extraction

For each eligible study, we extracted group comparison statistics for all interoception outcomes of interest. Study quality was also graded by authors ES and ZJW using a version of the Newcastle-Ottawa Scale proposed by Rødgaard et al. (2019), with scores ranging from 0 (lowest quality) to 9 (highest quality). The specific items used to grade study quality are presented in Supplemental Table S2. In addition, we extracted numerous putative moderator variables, including publication year, autism group mean age, age group (pediatric vs. adult), proportion of females in the autism group, autism group full-scale IQ, degree of age-matching (in Cohen’s d units), degree of IQ-matching (in Cohen’s d units), and degree of alexithymia-matching (in Cohen’s d units).

Statistical Analyses

All statistical analyses were performed in R version 4.1.0 (R Core Team, 2021). Descriptive statistics (i.e., means and standard deviations/standard errors), t values, or F values were used to calculate Hedges’ g effect sizes (Hedges, 1981) using the R package compute.es (Del Re, 2013). Regardless of how effects were reported in their original studies, the sign of g was standardized such that a negative effect size indicated smaller values of a variable in the autism group (e.g., worse performance on a heartbeat counting task, less self-reported attention to internal cues).

Based on the included studies, the following interoceptive outcomes (each present in at least three eligible studies) were meta-analyzed: (a) HCT performance, (b) HDT performance, (c) confidence ratings of HCT performance (CR-HCT), and (d) sIAt. Notably, a recent meta-analysis found relatively small correlations between HCT and HDT scores across studies (Hickman et al., 2020); therefore, we opted to analyze scores from these tasks as separate outcomes despite them often being treated as interchangeable. Taking advantage of the fact that many studies collected several interoceptive outcomes on the same sample, we fit a multiple-outcome three-level meta-analysis model (Cheung, 2019; Moeyaert et al., 2017; Van den Noortgate et al., 2015), treating effect size (level 3) as a random effect nested within study (level 2). The meta-analytic model additionally contained four different outcome-specific random intercept terms, and all between-outcome correlations were freely estimated. Models were fit to the outcome data reported for each study, a process similar to full-information maximum likelihood estimation of frequentist models. Summary data were presented using posterior density forest plots, which depict the probable distribution of effect sizes in each study conditional on prior beliefs and all observed data (Vuorre, 2016).

Model fitting was performed in a Bayesian framework using the R package brms (Bürkner, 2017, 2018). Similar to previous Bayesian meta-analyses (Z. J. Williams, Abdelmessih, et al., 2021; Z. J. Williams, Suzman, et al., 2021), we employed weakly informative priors, including a Normal(0, 1) prior on all fixed effects, a Half-Cauchy(0.3) prior on the standard deviations of the random intercept terms (D. R. Williams et al., 2018), and a Lewandowski-Kurowicka-Joe (LKJ; Lewandowski et al., 2009) prior (η = 2) on the correlation matrix of random effects. Model parameters were estimated via Markov chain Monte Carlo (MCMC) using the No U-turn Sampler (Hoffman & Gelman, 2014). Posterior distributions of g and other parameters were generated using 28,000 post-warmup MCMC draws from seven separate Markov chains. In order to quantify the heterogeneity of studies for a given outcome, we calculated the unstandardized outcome-specific variance parameter (τ²), as well as the standardized I² statistic (Higgins & Thompson, 2002). Additionally, we calculated model-based 95% prediction intervals (Graham & Moran, 2012; IntHout et al., 2016), which estimate the range of observed effects that would be most likely to occur if additional similar studies were conducted comparing these outcomes between autistic and neurotypical individuals.

Meta-analytic summary effects were operationalized as the median and the 95% highest-density credible interval (CrI) of a parameter’s posterior distribution. To quantify the evidence for an effect, we calculated the probability of direction (P_d; Makowski et al., 2019), defined as the proportion of the posterior distribution that has the same sign as the point estimate. P_d values can be thought of as the Bayesian analog of a frequentist one-tailed p-value; thus, P_d values of 0.975 (similar to a two-tailed p-value of 0.05) were chosen as the threshold for a “significantly nonzero” effect. We also used Bayesian inference to evaluate the interval null hypothesis that the overall summary effect was too small to be practically significant, falling within a region of practical equivalence (ROPE; Kruschke, 2018). The ROPE, containing all effect sizes that were judged as equivalent to zero, was defined as g = [−0.1, 0.1] based on the suggestion of Kruschke (2015). Evidence for or against the true parameter falling within this region was quantified using the ROPE Bayes factor (BF_ROPE; Makowski et al., 2019), with BF_ROPE values interpreted using standard guidelines (Wagenmakers et al., 2011). BF_ROPE values >3 provide substantial support for the alternative hypothesis (i.e., |g| > 0.1), whereas values <1/3 provide substantial support for the null hypothesis (i.e., |g| < 0.1). Although Bayes factors quantify evidence on a continuous scale, values between 1/3 and 3 are typically considered inconclusive or providing only “anecdotal” evidence (Wagenmakers et al., 2011).

For outcomes reported in at least five independent samples (i.e., HCT scores and sIAt), we additionally conducted univariate sensitivity analyses based on robust Bayesian meta-analysis with publication selection model averaging (RoBMA-PSMA; Bartoš et al., 2021; Maier et al., 2022), as implemented in the RoBMA R package (Bartoš & Maier, 2021). The RoBMA-PSMA meta-analytic estimates were constructed from an averaging of 20 separate meta-analytic models varying in terms of (a) the presence/absence of a nonzero summary effect, (b) the presence/absence of effect heterogeneity, and (c) the presence/absence of publication bias. Publication bias analyses were based on both selection models (Vevea & Hedges, 1995) and the precision-effect test/precision-effect estimate with standard errors method (PET-PEESE; Stanley & Doucouliagos, 2014). Only two-tailed selection models were utilized in the current investigations. RoBMA-PMSA models utilized the same Normal(0, 1) and Half-Cauchy(0.3) priors as the multiple-outcome model, as well as Cauchy(1) priors on the PET coefficients, Cauchy(5) priors on the PEESE coefficients, and Dirichlet priors on omega coefficients (see Bartoš et al., 2021 for additional technical details). Each model produces three inclusion Bayes factors (Hinne et al., 2020; Maier et al., 2022), individually reflecting evidence for or against (a) a nonzero summary effect (BF_EFF), (b) the presence of effect heterogeneity (BF_HET), and (c) the presence of publication bias (BF_PB). These Bayes factors are interpretable on the same scale as BF_ROPE. In cases where there was substantial evidence in favor of a nonzero effect (i.e., BF_EFF > 3), we additionally examined the model-averaged posterior distribution of effect sizes (conditional on a nonzero effect), which adjusts for publication bias using both the selection model and PET-PEESE procedures. These adjusted effect estimates were again tested for practical significance using BF_ROPE.

For outcomes reported in at least ten independent samples (i.e., HCT scores), we conducted meta-regression analyses in order to determine whether heterogeneity of effects could be meaningfully explained by putative moderator variables (Baker et al., 2009). Each meta-regression model was compared with its respective baseline (intercept-only) model using a model-comparison Bayes factor (BF₁₀). The R²_HET coefficient (Z. J. Williams, Suzman, et al., 2021), defined as the proportional reduction in τ² after adding the moderator variable, was also reported for each regression model. Missing data in meta-regression analyses were handled with 20-fold multiple imputation using the random forest procedure implemented in the MissForest R package (Stekhoven, 2013; Stekhoven & Buehlmann, 2012).

Results

The initial literature search identified 3977 results. After removing duplicates (n = 1809), the research team (including all listed authors) screened the remaining 2168 abstracts to identify studies eligible for full-text review. Average pairwise agreement between authors was good (97.4% agreement, overall Krippendorff’s α = 0.611), and all records that two or more raters agreed should be included were subjected to full-text review (n = 46). Authors ES and ZJW independently reviewed the full texts of these articles, with perfect (100%) agreement on inclusion/exclusion decisions. This process resulted in 10 records (describing 15 separate studies) meeting study inclusion criteria, reporting comparisons between a total of 467 autistic and 478 neurotypical participants (Table 1). Notably, relevant outcome data were obtained from authors for two of the included studies (Failla et al., 2020; Yao et al., 2019). Forward searching of all included articles (date of last search: December 11, 2021) revealed no additional eligible records, although some studies published in that time period (e.g., Palser et al., 2021) reported on samples that overlapped with those already included in analyses (e.g., Palser, 2018). Average quality scores for included studies were relatively high (7.13 out of 9 total; range 5.5–8.5). A PRISMA flow diagram is presented in Figure 1, and the quality scale item scores and effect sizes for each study included in each meta-analysis are reported in Supplemental Table S3.

Table 1.

List and Characteristics of Studies Included in Meta-analysis

Sample	n AUT	n NT	Age Group	MAge (AUT)	MAge (NT)	Prop. Female (AUT)	MFSIQ (AUT)	MFSIQ (NT)	d AGE	d FSIQ	d ALEX	HCT	HDT	CR-HCT	sIAt	Study Quality (0–9)
Hatt (2013)	34	35	Child	11.8	11.6	0.265	107.3	114.5	0.062	−0.476	1.623	+	−	−	−	7
Shah et al. (2016)	20	20	Adult	32.7	34.1	0.150	108.6	111.0	−0.107	−0.178	0.169	+	−	−	−	7
Garfinkel et al. (2016)	20	20	Adult	28.1	27.8	0.100	n.r.	n.r.	0.040	n.r.	n.r.	+	+	−	+	6
Nicholson et al. (2018)	46	48	Adult	40.2	41.2	n.r.	108.2	109.1	0.090	0.060	1.190	+	−	−	−	6.5
Palser (2018) [Study 3]/Palser et al. (2018)	30	30	Child	12.5	11.9	0.167	100.8	104.1	0.214	−0.250	n.r.	+	+	+	+	8.5
Palser (2018) [Study 5]	22	19	Child	13.5	12.2	0.227	99.8	103.0	0.391	−0.183	0.151	+	−	−	+	8
Palser (2018) [Study 6]/Palser et al. (2021)	44	55	Child	13.3	13.2	0.205	98.7	104.1	0.029	−0.334	0.320	+	+	−	+	7.5
Mul (2019) [Study 1]	20	24	Adult	25.4	25.1	0.350	112.2	107.4	0.035	0.341	0.995	+	−	−	+	7
Mul (2019) [Study 3]/Mul et al. (2018)	26	26	Adult	25.9	25.4	0.269	113.8	110.9	0.066	0.224	0.955	+	+	−	−	8.5
Nicholson et al. (2019) [Children]	21	21	Child	13.0	12.7	0.238	111.8	112.5	0.190	0.050	n.r.	+	−	−	−	7
Nicholson et al. (2019) [Adults]	21	21	Adult	37.2	41.2	0.381	106.0	105.5	0.300	0.040	1.580	+	−	−	−	6
Yao et al. (2019)	35	37	Child	12.2	12.7	0.257	99.8	110.7	−0.201	−0.576	n.r.	−	−	−	+	5.5
Failla et al. (2020) [Children]/Schauder et al. (2015)	49	52	Child	12.4	11.9	0.122	104.2	113.5	0.171	−0.550	n.r.	+	−	+	+	8
Failla et al. (2020) [Adults]	18	8	Adult	30.4	27.1	0.333	104.2	114.5	0.336	−0.801	n.r.	+	−	+	−	7
Pickard et al. (2020)	61	62	Child	13.5	13.5	0.311	98.2	100.8	−0.036	−0.202	0.0083	+	−	+	+	7.5

Note. Cells with “+” indicate that a particular outcome was reported in a study, whereas cells containing “−” indicate that a particular outcome was not reported. n.r. = not reported; AUT = autism; NT = neurotypical; M_Age = mean age (in years); M_FSIQ = mean full-scale IQ; d_AGE = standardized mean difference for age (AUT – NT); d_FSIQ = standardized mean difference for full-scale IQ (AUT – NT); d_ALEX = standardized mean difference for alexithymia scores (AUT – NT); HCT = heartbeat counting task; HDT = heartbeat discrimination task; CR-HCT = confidence ratings from HCT; sIAt = subjective interoceptive attention.

Figure 1.

PRISMA flow diagram detailing the identification, screening, and selection of studies for inclusion in analyses.

The multi-outcome meta-analytic model included 30 effect sizes: 14 based on HCT scores (Mdn = −0.213, range [−1.088, 0.096]), 4 based on HBD scores (Mdn = −0.105, range [−0.344, 0.045]), 4 based on CR-HCT (Mdn = 0.459, range [0.278, 0.699]), and 8 based on sIAt scores (Mdn = −0.148, range [−0.580, 1.993]). Full results of this model can be found in Table 2 (see also Figure 2). On average, autistic individuals demonstrated significantly diminished objective interoceptive accuracy, as measured using HCT performance, compared to neurotypical controls (g = −0.333, CrI_95% [−0.535, −0.138], BF_ROPE = 9.80). However, this same pattern was not demonstrated with HDT performance, an outcome for which the ROPE Bayes factor demonstrated practical equivalence between diagnostic groups (g = −0.086, CrI_95% [−0.410, 0.234], BF_ROPE = 0.112). Notably, subjective interoceptive accuracy, as measured using CR-HCT, also significantly differed between groups, with autistic individuals rating themselves as more confident in their performance on average (g = 0.430, CrI_95% [0.123, 0.750], BF_ROPE = 5.50). Lastly, scores on sIAt measures demonstrated no significant diagnostic group differences when averaged across studies, with the ROPE Bayes factor additionally indicating practical equivalence between groups (g = −0.003, CrI_95% [−0.523, 0.516], BF_ROPE = 0.171). Heterogeneity was low to moderate for HCT, HDT, and CR-HCT outcomes (Table 2), whereas a large amount of heterogeneity was present in sIAt effects (I² = 87.7%, CrI_95% [69.5, 98.4]). Furthermore, model-based correlations between interoceptive outcomes across studies were generally quite low (rs between −0.244 and 0.050; Supplemental Table S4).

Table 2.

Three-level Meta-analytic Summary Effects and Heterogeneity Metrics

Outcome	k	n AUT	n NT	g [95% CrI]	P d	BF ROPE	τ2 [95% CrI]	I2 [95% CrI]	95% PI
Heartbeat Counting Task (HCT)	14	432	441	−0.333 [−0.535, −0.138]	0.999	9.80	0.052 [0.000, 0.195]	44.0% [0.0, 74.7]	[−0.938, 0.237]
Heartbeat Discrimination Task (HDT)	4	120	131	−0.086 [−0.410, 0.234]	0.718	0.112	0.004 [0.000, 0.182]	17.3% [0.0, 73.6]	[−0.596, 0.468]
HCT Confidence Ratings (CR-HCT)	4	158	152	0.430 [0.123, 0.750]	0.994	5.50	0.016 [0.000, 0.210]	20.6% [0.0, 77.5]	[−0.119, 0.994]
Subjective Interoceptive Attention (sIAt)	8	281	299	−0.003 [−0.523, 0.516]	0.505	0.171	0.407 [0.033, 1.283]	87.7% [69.5, 98.4]	[−1.525, 1.572]

Note. k = number of studies; AUT = autism; NT = neurotypical; CrI = highest-density credible interval; P_d = probability of direction; BF_ROPE = Bayes factor for the region of practical equivalence [−0.1, 0.1]; PI = prediction interval (posterior predictive interval).

Figure 2.

Posterior density forest plots comparing (A) heartbeat counting task effects, (B) heartbeat discrimination task effects, (C) heartbeat counting confidence rating effects, and (D) subjective interoceptive attention effects. The standardized mean difference (SMD, i.e., Hedges’ g) and 95% highest-density credible interval (CrI) for each study represent the posterior distribution of that study’s mean effect size, conditional on prior beliefs and the observed data. Negative values of g indicate lower interoceptive accuracy/attention in the autism group compared with neurotypical control subjects. The gray shaded areas indicate the region of practical equivalence for each comparison, g = [−0.1, 0.1]. Raw effect sizes from each study are provided in Supplemental Table S2.

Sensitivity analyses utilizing univariate RoBMA-PSMA models were then examined for HCT and sIAt outcomes, both generally agreeing with the conclusions of the three-level model (Table 3). For the HCT model, there was substantial evidence for the presence of a nonzero effect (BF_EFF = 28.0), although Bayes factors examining effect heterogeneity (BF_HET = 1.15) and publication bias (BF_PB = 0.580) were both inconclusive. Adjustment of the meta-analytic HCT effect for publication bias slightly attenuated the difference between groups, but the presence of a practically meaningful group difference was not altered (g = −0.299, CrI_95% [−0.544, −0.092], BF_ROPE = 6.14). For the sIAt model, there was substantial evidence against the presence of a nonzero marginal effect (BF_EFF = 0.238), combined with extremely strong evidence of heterogeneity (BF_HET = 2.40 × 10³) and slight evidence against the presence of publication bias (BF_PB = 0.367). As the presence of a nonzero sIAt group difference was not supported by the RoBMA-PSMA model, the conditional effect distribution (which conditions on the presence of a nonzero summary effect) was not interpreted for this outcome.

Table 3.

Robust Bayesian Model-averaged Meta-analysis Model Coefficients

Outcome	BF EFF	BF HET	BF PB	Conditional g [95% CrI]	τ2 [95% CrI]	ω[0, 0.05]	ω[0.05, 0.1] [95% CrI]	ω[0.1, 1.0] [95% CrI]	PET [95% CrI]	PEESE [95% CrI]
Heartbeat Counting Task (HCT)	28.0	1.15	0.580	−0.299 [−0.544, −0.092]	0.052 [0.000, 0.195]	1.0 [fixed]	0.956 [0.181, 1.0]	0.914 [0.162, 1.0]	−0.053 [−0.697, 0.0]	−0.087 [−1.351, 0.0]
Subjective Interoceptive Attention (sIAt)	0.238	2404	0.367	−0.003 [−0.523, 0.516]	0.407 [0.033, 1.283]	1.0 [fixed]	1.0 [0.408, 1.0]	1.0 [0.345, 1.0]	0.094 [0.0, 0.984]	0.242 [0.0, 3.162]

Note. For all credible intervals (CrIs) other than that of the conditional effect (described using a highest-density interval), equal-tailed intervals (based on the 2.5^th and 97.5^th quantiles) were used to summarize the model-averaged posterior due to the multimodal nature of the mixture distributions. BF_EFF = Bayes factor for presence of a nonzero effect; BF_HET = Bayes factor for presence of nonzero heterogeneity; BF_PB = Bayes factor for presence of publication bias; Conditional g = summary effect based on conditional posterior (effect assumed to be nonzero); PET = precision effect test; PEESE = precision-effect estimate with standard errors.

The meta-regression models of HCT effects did not find sufficient evidence in favor of any tested moderator, with significant evidence demonstrated against moderation by publication year, quality score, mean age of autism sample, binary age group (pediatric vs. adult), mean IQ of the autism sample, or degree of matching on alexithymia scores (Table 4). There was inconclusive evidence to support or refute claims of moderation by proportion of females in the autism group (BF₁₀ = 0.912, R²_HET = −0.026) or degree of IQ-matching (BF₁₀ = 1.59, R²_HET = 0.383), although the latter variable explained nearly 40% of between-study heterogeneity in HCT scores. Post-hoc examination of model coefficients found that studies with larger differences in full-scale IQ between autistic and neurotypical groups tended to demonstrate larger differences in HCT scores (β = −0.548, CrI_95% [−1.157, 0.067], P_d = 0.960, BF_ROPE = 1.585), though this effect did not quite reach the threshold for statistical or practical significance.

Table 4.

Bayes Factor Values and Heterogeneity Explained for Tested Moderator Variables

Moderator	BF 10	R 2 HET
Publication Year	0.109	−0.232
Quality Score Total	0.112	−0.177
Mean Age in Autism Group	0.010	−0.268
Age Group (Children vs. Adults)	0.201	−0.204
Proportion Female in Autism Group	0.912	−0.026
Mean IQ in Autism Group	0.175	−0.060
Age Matching (dAGE)	0.705	−0.083
IQ Matching (dFSIQ)	1.594	0.383
Alexithymia Matching (dALEX)	0.912	0.193

Note. Bayes factors that provide substantial evidence against moderation are presented in italics.

Due to the high level of heterogeneity present in sIAt effects across studies, we conducted an additional post-hoc sensitivity analysis to examine the effect of excluding a single outlier study with a very large effect size (g = 1.993; Garfinkel et al., 2016) and studentized deleted residual value (Viechtbauer & Cheung, 2010) of 5.06 (CrI_95% [3.54, 5.87]). Notably, when excluding this outlier from a univariate meta-analysis of sIAt effects, the model-estimated heterogeneity dropped from τ² = 0.408 (CrI_95% [0.037, 1.339]) to τ²_(-i) = 0.032 (CrI_95% [0, 0.184]), a reduction of 92.2%. Moreover, the sIAt summary effect without the Garfinkel study was estimated to be g = −0.222 (CrI_95% [−0.460, 0.032], P_d = 0.961). Although this outlier-excluded estimate no longer met the criteria for equivalence between groups (BF_ROPE = 0.533), the Bayes factor value was inconclusive regarding the presence or absence of a practically meaningful average group difference in sIAt scores between autistic and neurotypical individuals.

Discussion

The current study was the first meta-analysis to quantitatively synthesize extant studies comparing various indices of interoception between autistic and neurotypical individuals. Grouping measures according to the 2×2 interoceptive classification put forth by Murphy and colleagues (2019), we found substantial evidence that autistic individuals as a group perform more poorly than control participants on one test of objective interoceptive accuracy (HCT) despite the two groups demonstrating practically equivalent performance on a similar performance-based interoceptive task (HDT). Furthermore, studies measuring subjective interoceptive accuracy (i.e., self-reported confidence in HCT performance) demonstrated that autistic participants typically provide significantly higher average confidence ratings than controls despite their worse HCT performance on average. In contrast to the significant differences seen in subjective interoceptive accuracy, autistic and neurotypical individuals tended to score similarly on measures of subjective interoceptive attention such as the BPQ, though between-study heterogeneity was quite large for this particular outcome. In general, these findings suggest that autistic individuals do differ from neurotypical controls on multiple aspects of interoception, although the magnitudes of these differences are small to moderate and sometimes inconsistent across tasks. We, therefore, take this to mean that interoceptive difficulties, as broadly conceptualized, are unlikely to represent a core pathophysiologic feature of autism, as some individuals have previously proposed (Quattrocki & Friston, 2014). Nevertheless, individual differences in interoceptive accuracy and attention may still be important in the autistic population, potentially contributing to the development of alexithymia (Kinnaird et al., 2019), mood and anxiety disorders (Hollocks et al., 2019), somatic symptoms (Grant et al., 2022; Z. J. Williams & Gotham, 2022), and other associated features affecting only a subset of autistic individuals.

The most frequently studied interoceptive outcome in this body of literature was the HCT, a performance-based measure of objective interoceptive accuracy in the cardiac domain. Despite inconsistent reports of significant between-group differences in HCT scores in individual studies, we found a small yet robust summary effect of diagnostic group that was minimally affected by adjusting for publication bias. Notably, if we assume a true effect size of approximately g = −0.333, all studies included in this meta-analysis would have been woefully underpowered to detect such an effect—in fact, the largest sample examined (n = 123; Pickard et al., 2020) had just 45% power to detect an effect of this magnitude using a two-sample t-test. For the median sample size of 52, power to detect g = −0.333 drops to just under 22%. Given the low statistical power achieved by most primary studies for detecting effects of this size, it is perhaps unsurprising that nearly two thirds of samples examining HCT effects did not report statistically significant differences (Lakens & Etz, 2017). For future studies aiming to characterize differences in HCT performance between autistic and neurotypical individuals, it will be important to adequately power investigations to detect these relatively subtle differences between diagnostic groups.

Furthermore, although we ran meta-regression models in an attempt to explain the heterogeneity found in HCT effect sizes, none of the tested study-level characteristics was found to significantly moderate these effects across studies. This lack of significant moderators was possibly due to issues of statistical power, as simulations have found that many moderator effects are not easily detectable with fewer than approximately 20 studies (López-López et al., 2014). Notably, the significant evidence against moderation by age group appears to counter the claim by Nicholson et al. (2019) that interoceptive difficulties are present in autistic children but not in autistic adults. Additionally, while there was some evidence to suggest that samples poorly matched on full-scale IQ may show larger discrepancies in HCT scores, this finding did not reach our a priori significance threshold. Nevertheless, as the degree of IQ matching did explain an estimated 38% of HCT effect heterogeneity, we believe that researchers should be aware of IQ differences as a potential confounding factor. Thus, out of an abundance of caution, we recommend that future studies comparing HCT performance between autistic and neurotypical participants covary for full-scale IQ as a way of combating this potential source of bias.

One concerning finding in the present study was the discrepancy between HCT and HDT scores, which was present in all studies that collected both measures (Garfinkel et al., 2016; Mul et al., 2018; Palser et al., 2018). Unlike HCT scores, average HDT scores did not significantly differ between autistic and neurotypical samples, and our Bayesian analysis found substantial evidence in favor of equivalence between groups on this measure. A recent meta-analysis has demonstrated that scores on these two measures of cardiac interoceptive accuracy do not correlate highly in cross-sectional studies (Hickman et al., 2020); however, as both HCT and HDT scores respond in tandem to manipulations of cardiac interoception such as tilt table maneuvers (Schulz et al., 2021) and an interoceptive training intervention (Quadt et al., 2021), the shared variance that does exist likely represents the underlying interoceptive ability to detect cardiac signals. Given the discrepant HCT and HDT findings in the autism literature, it is therefore difficult to determine whether group differences in HCT performance are attributable to true underlying differences in cardiac interoception or one of the many construct-irrelevant factors that contribute to variance in HCT scores (Brener & Ring, 2016; Corneille et al., 2020; Desmedt et al., 2018; Hickman et al., 2020; Murphy, Millgate, et al., 2018; Zamariola et al., 2018). Notably, the study-level variable with the largest moderating effect on HCT effect size was the degree of IQ (mis)matching between AUT and NT groups, consistent with prior studies showing confounding of HCT scores by general intelligence (Murphy, Millgate, et al., 2018). Thus, it is likely that at least some of the group differences in HCT between autistic and neurotypical individuals are primarily driven by non-interoceptive factors. The HDT has also been criticized in terms of its potential insensitivity to individual differences, as a relatively small proportion of individuals in the general population tend to perform at above-chance levels on this measure (Brener & Ring, 2016), and a sizable number of trials may be needed to provide an index of interoceptive ability with sufficient reliable variance (Kleckner et al., 2015). Thus, discrepancies between HCT and HDT findings in autism may be a result of methodological issues with one or both of these classical interoceptive tasks. To better estimate objective interoceptive accuracy in autistic individuals, future studies on this topic should attempt to utilize newer and more psychometrically robust cardiac interoceptive tasks that overcome many of the problems inherent in both the HCT and HDT (e.g., Fittipaldi et al., 2020; Khalsa et al., 2009; Legrand et al., 2022; Plans et al., 2021; Pohl et al., 2021).

Another notable finding of our meta-analysis is the clear mismatch between autistic individuals’ performance as a group on the HCT and their self-rated performance on that same task. Despite on average performing slightly worse than their neurotypical counterparts, autistic participants provided higher average confidence ratings with a small to moderate effect size. Although differences in so-called subjective interoceptive accuracy may not reflect a true underlying difference in the perception of internal cues, it is notable that autistic individuals on average appear to have somewhat limited insight into their performance on the HCT. In the one study that calculated relative discrepancies between subjective and objective HCT performance, autistic individuals did indeed demonstrate reduced “interoceptive insight” compared to neurotypical controls (Pickard et al., 2020). However, without covarying for group differences in confidence ratings and HCT scores, it remains unclear whether this group difference is truly driven by just one of the two interoceptive domains rather than their interaction. Although metacognitive awareness of interoceptive performance (i.e., “interoceptive awareness” in the tripartite framework; Garfinkel et al., 2015) may potentially be a clinically relevant outcome in this area of research, investigators utilizing this construct must be aware of the statistical complexities inherent in working with outcomes that are themselves the difference between two component variables (Humberg et al., 2019; see also Z. J. Williams, 2022).

When examining group differences in SIA, operationalized using scores such as the BPQ awareness subscale and MAIA noticing subscale, autistic and neurotypical individuals did not differ in their average scores. Additionally, this finding was qualified by a high degree of heterogeneity, seemingly driven by one extreme outlier (Garfinkel et al., 2016). As only eight sIAt effects were included in the meta-analysis, our analysis was markedly underpowered to detect moderation (Baker et al., 2009; López-López et al., 2014), and thus meta-regression was not performed on this outcome. Nevertheless, a sensitivity analysis that excluded the outlying study was performed, finding insufficient evidence to make conclusions about the presence of either significant group differences or equivalence in sIAt scores. While the reasons for the outlying effect reported by Garfinkel et al. (2016) were not systematically explored, recent work in the general population has revealed that the sIAt measure used in that study (the BPQ awareness subscale) is variably interpreted by different individuals as measuring either interoceptive attention or interoceptive accuracy (Gabriele et al., 2022). Although the specific interpretations of BPQ items were not assessed in any of the studies included in our meta-analysis, it is quite possible that sample-level differences in BPQ item interpretation were responsible for at least a portion of the observed heterogeneity in sIAt effects. To better characterize autism-associated sIAt differences in future studies, we recommend utilizing a more specific measure of sIAt such as the Interoceptive Attention Scale (Gabriele et al., 2022) or a measure such as the Interoceptive Sensitivity and Attention Questionnaire (Bogaerts et al., 2022) that separately assesses interoceptive attention and accuracy using different subscales. Although sIAt scores were found to be equivalent across groups in the current meta-analysis, the lack of robustness of this finding to outlier exclusion and the ambiguity with which BPQ items may be interpreted suggests that this conclusion may not hold in the case of other sIAt measures. Additional studies are, therefore, necessary to clarify the degree to which autism is associated with differences in various facets of self-reported interoception, including but not limited to sIAt.

Strengths and Limitations

This study has a number of strengths, including the use of a modern framework to characterize interoceptive outcomes, a large gray literature search that successfully uncovered a number of unpublished findings, a meta-analytic model that capitalized on the multivariate nature of outcome data in the reviewed studies, and statistically rigorous publication bias tests based on cutting-edge meta-analytic techniques. However, it was not without limitations. First and foremost, all meta-analytic studies are limited by the body of literature available for inclusion, and while the literature on autism and interoception is steadily growing, relatively few effect sizes were available for synthesis. Estimates of group differences in HDT scores and HCT confidence ratings were based on only four samples each, producing imprecise and potentially biased estimates of true group differences on these outcomes. In addition, due to the scope of the literature examining interoceptive variables in autism, we were only able to produce summary estimates for a subset of interoceptive variables. For instance, only one study (Nicholson et al., 2019) reported on group differences in a non-cardiac (in this case, respiratory) measure of objective interoceptive accuracy; thus, we were unable to generate summary effects for this outcome. In addition, despite the 2×2 framework defining the construct of “objective interoceptive attention” (quantified by experience-sampling data; Murphy et al., 2019), measures of this construct have not yet been assessed in autism and therefore could not be meta-analyzed. Self-report questionnaires assessing subjective interoceptive accuracy (rather than attention), including the Interoceptive Sensory Questionnaire (ISQ; Fiene et al., 2018) were also unable to be analyzed due to the lack of multiple studies reporting group differences on the same questionnaires. Given that autistic individuals have been shown to score substantially lower on the ISQ than neurotypical controls (Fiene et al., 2018), subjective accuracy measures other than confidence ratings may represent a unique dimension of interoception that warrants study specifically in the autistic population. Finally, this study did not examine correlations between interoceptive measures and behavioral outcomes of clinical relevance, such as anxiety, alexithymia, other sensory features, or autistic traits. Further research characterizing the relations between these outcome domains may provide valuable information regarding links between interoception and core and associated features of autism, guiding additional studies on this topic.

Conclusion

In conclusion, this systematic review and meta-analysis suggests that autistic individuals demonstrate group-level differences in multiple interoception-related outcomes when compared to neurotypical controls, although several caveats exist. Though the most robust finding of diminished HCT performance indicates potentially reduced objective interoceptive accuracy in autism, the lack of group differences in HDT suggests that this difference may be unrelated to true differences in interoceptive ability. Differences in self-reported dimensions of interoception were also mixed, with autistic individuals reporting higher confidence in HCT performance (despite lower objective performance) but no average differences in self-reported attention to internal cues. Among the outcomes that differed between groups, effect sizes were modest, demonstrating that nearly all studies published to date in this area are greatly underpowered to detect the most likely true effects. Though the limitations of the extant literature may preclude definitive conclusions about the ways in which autism is associated with interoceptive differences, this study was able to provide a meaningful summary of work in this area to date, paving the way for more targeted future investigations into interoceptive processes both within and outside of autism research.