Reduced Context Updating but Intact Visual Priors in Autism

R. Randeniya; I. Vilares; J. B. Mattingley; M. I. Garrido

Introduction

Sensory processing alterations, which can affect one or more sensory modalities (), are reported in around 90% of autistic adults (). Autistic individuals can show improved performance in visual search tasks (; ) and visual and auditory target detection (), compared to neurotypicals. However, sensory processing dysfunction (e.g. hypo- or hyper-sensitivities) can inhibit participation in activities such as learning and social interaction, which in turn impose lifelong challenges (). Due to the heterogeneous nature of perceptual function in autism, current diagnostic measures, such as the Autism Diagnostic Observation Schedule (ADOS), have significant limitations in characterizing the nature of perceptual experiences in people with a diagnosis of Autism Spectrum Disorder (ASD), and can only be applied for specific degrees of severity and age along the autism spectrum (). Thus, better characterization of the mechanisms that give rise to autistic perception should help in understanding perceptual subtypes and could pave the way for personalized interventions.

“Bayesian brain” accounts for autism have offered explanations for perceptual differences in autism (; ; ; ; ). Simply stated, a Bayesian approach to sensory learning posits that during learning individuals form models, encoded as priors, by detecting patterns in the environment (Figure 1A). New incoming information (i.e., likelihood) is then matched against these priors. The internal model (prior) about the sensory environment is updated with new sensory information until model updating is no longer necessary in a stable environment (). This framework has proven useful in explaining typical sensory learning and decision-making under uncertainty. According to Bayes theorem, posteriors (or perceptual decisions) will combine both prior and likelihood information but give more weight to whichever source has a higher precision (i.e., the lower variance). In autism, however, it is theorised that this process may be altered due to an imbalance of precision ascribed to sensory observations relative to prior beliefs (); (Note we use precision here in its statistical sense, which is the inverse of variance). From a Bayesian perspective, this likelihood over-reliance can be formalised mathematically as a shift in the posterior toward the sensory observation (likelihood) and away from the prior (belief). This shift can be attributed to different underlying causes. The “hypo-priors model” ((); Figure 1B) suggests that the shift is caused by noisier or less precise (high variance) priors, equivalent to a weak model (or belief) of the environment. The “precise likelihood model” ((); Figure 1C) argues instead that priors are intact, but there is an increase in the precision associated with sampling new information, such that sensory representations are more narrowly tuned. The conundrum is that given the relative difference in the precision of priors and likelihoods under the hypo-priors and precise likelihood models, the two accounts effectively give rise to the same posterior means.

Figure 1

Bayesian models of atypical perception in Autism. A) Bayesian models postulate that in typical learning – new sensory observations (likelihood; red) are integrated with the learned model of the world (Prior; blue), leading to a perception/decision (posterior; orange). B) Hypo-prior model for Autism (): A weak model of the world (high variance in the prior) increases the reliance on new sensory observations (likelihood) that are relatively more precise than the prior (but as precise in autistic as in typical individuals) C) Precise Likelihood Model (): Overly precise new sensory observations (likelihood has less variance than in typical individuals, and more precise than the prior) can increase the reliance on new sensory observations.

Amongst the criticisms of these Bayesian theories, it has been argued that while they can explain hypersensitivities and sensory overload, they do not account for hyposensitivity or other perceptual disruptions in autism such as weak global coherence (; ). Some studies investigating aberrant precision models in autism have found that children and adults on the autism spectrum are able to learn priors (; ), while other studies have shown that this process is altered (). Evidence for the precise likelihood model (Figure 1C) has been supported by a study showing that increasing autistic traits correlated with increasing precision in likelihood in a neurotypical sample (), but not in prior representations. In contrast, another study which employed a signal detection approach provided evidence in support of the (weak) hypo-priors model (), again in a sample of neurotypical individuals. Further a study investigating central tendency in autistic children demonstrated poorer performance in autistic children than matched controls () indicating evidence for a hypo-priors model.

In this study, we took a Bayesian approach to better understand how prior and likelihood information is utilized during visual sensory learning and perceptual decision-making in autism. Specifically, we investigated: 1) whether autism spectrum (AS) individuals rely more on sensory than on prior information compared with neurotypicals (NT), and 2) whether precision in prior and likelihood distributions differs between AS and NT groups. Further, given the utility of undertaking a hybrid of categorical and dimensional approaches to understanding autism (; ), we also investigated how autistic traits and sensory sensitivities are related to the relative weighting of prior and new information during a perceptual decision. Here, we empirically assess these theoretical models by employing Bayesian modelling of behavioural data captured in a task that manipulates uncertainty in priors and likelihoods (; ; ). These studies have consistently demonstrated that people integrate prior and likelihood information in a qualitatively Bayesian-like fashion. While people often do not behave in a (quantitatively) perfectly Bayesian optimal way, the ways in which they deviate from this optimality can give us insights into the subjective information they have available (). Thus, because we are using Bayes as a framework to understand perceptual processes in autism, we do not test, nor do we compare the above-mentioned Bayesian models with non-Bayesian accounts of perceptual processes in autism. Our goal was to shed light on computational models of perception in autism.

Methods

Participants

We recruited a total of 80 adult participants (48 Neurotypicals and 32 participants who self-identified as having received a diagnosis of autism spectrum disorder). Recruitment was undertaken via Asperger’s Services Queensland, Autism Queensland, Mind and Hearts, The University of Queensland (UQ) online SONA system, and online advertisements. All participants (or their guardians) completed an online screening form. Participants were included in the study if they were between the ages of 18 – 35 years and had no history of neurological abnormalities. Autistic participants were recruited initially if they reported having received a diagnosis of autism spectrum disorder from a clinician. Neurotypical participants had an additional inclusion criterion of no history of neurological abnormalities or psychological disorders and reported no current use of any medication acting on the nervous system. For group analysis, the autism spectrum (AS) group had 25 participants with a confirmed diagnosis of ASD, and the neurotypical (NT) group had age and gender matched 25 participants. For the dimensional analysis all 80 participants were included, this consisted of 48 neurotypical, 25 autistic and 7 ‘other’ participants who could not be confirmed as being on the autism spectrum using the ADOS. Thus, no participant’s data were excluded from the study. All participants provided written informed consent to participate in the study and were compensated at a rate of AUD20 per hour for their participation in the study. This study was approved by the Human Research Ethics Committee of The University of Queensland (Approval No.: 2019000119).

Procedure

32 participants who reported having received a diagnosis of ASD from a clinician undertook a diagnostic interview with a clinical psychologist using the Autism Diagnostic Observation Schedule (ADOS) for Adults (; ), to confirm diagnosis and characterize symptom severity. The assessment lasted approximately 1-hour and was conducted on a separate day from the experimental sessions. Of the 32 participants, six individuals who scored below 3 on the ADOS were excluded from the autism spectrum group analysis. Additionally, one participant with a self-r of ASD was not available to complete the ADOS assessment. Thus, a total of 25 participants were confirmed to be on the autism spectrum and were included in the autism spectrum (AS) group (See Figure 1).

Psychometric measures

All participants completed self-report questionnaires including the 50-item Autism Quotient (AQ) questionnaire () and 93-item Sensory Processing Quotient (SPQ) Questionnaire () which were used to measure autistic traits and sensory sensitivities respectively. Participants also completed the Beck Anxiety Inventory () and Beck Depression Inventory ().

2-Prior Coin Task ()

Participants engaged in a modified version of the visual decision-making task developed by Vilares and colleagues (), performed either in a 3T Magnetic Resonance Imaging scanner (AS = 28 and NT = 47) or outside of the scanner at a computer (AS = 4 and NT = 1). Participants were shown an image of a pond on a screen (Figure 2) and were told that someone was throwing a coin to the middle of the pond (i.e., the middle of the screen). Participants were told that they would see trials from two different coin-throwers and that one was better at throwing to the centre than the other. Unbeknownst to the participant, thrower A was more precise than thrower B (with order counterbalanced across participants), throwing the coin closer to the middle more often (narrow prior) than thrower B (wide prior). Before each block, participants were shown which thrower (Thrower ‘A’ or Thrower ‘B’) was throwing next. On each trial, participants were shown five blue dots representing the splashes that the coin made when falling into the pond. For each trial, participants were instructed to use a keyboard/button box to move a blue bar (“net”) horizontally across the screen to where they thought the coin had fallen on that trial. Next, participants moved a bar horizontally to rate how confident they were about their decision on a scale ranging from 0 (Guessing) – 100 (Confident). All these events were self-paced. The participant was then shown the true position of the coin, as a yellow dot, for 1500 ms.

Figure 2

Coin Task Set up. A) Task design adapted from Vilares et al. (): the four conditions of the task – with two types of prior (P_N = narrow prior; P_W= wide prior) and two types of likelihood (L_N= narrow likelihood; L_W= wide likelihood) uncertainty. B) The time course of a single trial: participants are asked to estimate the position where a coin fell given where the 5 splashes appeared. C) The trials were organized into 24 short blocks of 12 trials each with participants being told at the beginning of each block which thrower (“A” or “B”) will be throwing the coin (blue bar). Four types of trials are shown – Narrow Prior – Narrow Likelihood (P_NL_N, black); Narrow Prior – Wide Likelihood (P_NL_W, grey); Wide Prior – Narrow Likelihood (P_WL_N, red); and Wide Prior – Wide Likelihood (P_WL_W, orange).

The prior variance was manipulated across blocks as follows. The coin position was drawn from a Gaussian distribution in every trial, centred on the centre of the screen with a standard deviation that was either narrow (σ_PN = 2.5% of the screen width; Thrower A) or wide (σ_PW = 8.5% of the screen width; Thrower B) across blocks. The true prior variance was constant within a block, whereas the variance of the likelihood changed pseudo-randomly within a block. The variance of the splashes (i.e., the five blue dots) was the true likelihood variance. The spread of these dots could be narrow or wide, corresponding to narrow or wide likelihood variance, respectively. The position of the five dots on the x-axis of the screen was drawn from a Gaussian distribution with a mean corresponding to the true coin position and standard deviation either narrow (σ_LN = 6%) or wide (σ_LW = 15%). Where the variance of the five dots on the x-axis was narrow, the standard deviation of the dots on the y-axis was wide (σ = 15%), and correspondingly narrow (σ = 6%) on y-axis if wide on the x-axis. This was to ensure that the total area of the spread of the dots was uniform across both narrow and wide conditions and was the same area on the visual regions. However, since participants moved the bar only on the x-axis the true likelihood variance on each trial was calculated as the standard deviation on the x axis.

Thus, the experiment conformed to a 2 × 2 design with Prior (wide and narrow) by Likelihood (wide and narrow), consisting of 4 types of trials/conditions: Narrow Prior – Narrow Likelihood (P_NL_N, black); Narrow Prior – Wide Likelihood (P_NL_w, grey); Wide Prior – Narrow Likelihood (P_wL_N, red); and Wide Prior – Wide Likelihood (P_wL_w, orange).

The practice task: consisted of only 2 blocks (one per thrower/prior type, with 40 trials each thrower), taking between 10–15 minutes to complete.

The main task: consisted of 12 blocks per Prior, with each block having 12 trials from a single thrower (Thrower A or B) with 2 types of likelihoods (narrow/wide). Before the thrower changed participants were instructed as to which thrower, A or B, would be throwing next (5 seconds); See Figure 2C. In total, the main task consisted of 288 trials (72 trials per condition). The task duration was self-paced and took between 35 and 60 minutes to complete.

Likelihood Only Task

After completing the 2-Prior Coin Task, participants completed a Likelihood Only Task outside of the scanner at a computer. Four participants in the AS group and one participant in the NT did not complete this task, as they did not wish to continue due to fatigue. The aim of this task was to estimate the participant’s perceived likelihood distribution, i.e., how the participants represent the centre of the dots on their own, without prior knowledge. Participants saw trials as in Figure 2B. However, participants did not report confidence on this task and the true coin position was always the centre of the splashes. Participants were simply instructed to move the net to where they thought the centre of the splashes was (i.e., the middle of the 5 blue dots). The true centre of the dots was shown in yellow as feedback to each participant on every trial. This task consisted of only one block with 144 trials from 2 types of likelihood (narrow/wide) and took between 15–20 minutes to complete. This task was conducted after the 2-Prior Task, as we expected that the participants would be biased towards the centre of the splashes if they performed the Likelihood Only Task first and thus would have a difficulty in learning the prior in the main 2-Prior Task that followed. In hindsight, however, we acknowledge that this caused the participants to carry over their priors (from the 2-Prior Task) to the Likelihood only tasks, hence defeating our purpose of having a no-prior task. Future studies may wish to consider administering the Likelihood-only task first or counterbalance the two tasks.

Behavioural Analysis

From Bayes rule, we can obtain what would be the optimal estimate for the position of the coin on each trial of the coin-catching task (For detailed workings see (; )):

(1)

X est = σ L 2 σ L 2 + σ P 2 μ P + σ P 2 σ L 2 + σ P 2 μ L

M1 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{{\rm{{\rm X}}}_{est}} = \frac{{\sigma _L^2}}{{\sigma _L^2 + \sigma _P^2}}\;{\mu _P} + \frac{{\sigma _P^2}}{{\sigma _L^2 + \sigma _P^2}}\;{\mu _L}\] \end{document}

where X_est is the estimated position of the coin (i.e., participant responses on each trial). (σ²_P, µ_P) and (σ²_L, µ_L) represent the variance and mean of the prior (i.e., participant’s subjective model of where each thrower would throw the coin) and likelihood/sensory observation distributions (i.e., participant’s subjective measure of the five blue dots).

Estimating likelihood vs. prior reliance

As described in Vilares et al. (), for each condition we can fit a linear regression line to predict the participants’ estimated position of the coin for each trial (X_est) as a function of the likelihood mean (µ_L here are the centre of the splashes). The slope of the regression line (the term $σ P 2 σ L 2 + σ P 2$ M6 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\textstyle{{\sigma _P^2} \over {\sigma _L^2 + \;\sigma _P^2}}}\] \end{document} ) in equation (1) is the sensory weight (sw/likelihood reliance), which indicates how much the participant relies on the likelihood/sensory information (see Figure 3A). The closer the slope is to one, the more the participant relied on likelihood information (i.e., centre of five blue dots). If we assume that participants only rely on current and/or prior information (e.g., if we exclude random behaviour), then a slope closer to zero corresponds to relying more on the prior (i.e., centre of the screen), with anything in between indicating integration of the likelihood and the prior.

Figure 3

Likelihood vs. Prior Reliance in the 2-Prior Task. A) Sensory weight (likelihood reliance) for a participant is calculated by obtaining the slope of the regression (orange) between the true centre of the likelihood and participant’s estimates of the coin position for each condition. Slopes closer to 1 (red) indicate that participants are more reliant on likelihood information (mean of the splashes) while slopes closer to zero (green) indicate that the participant didn’t rely much on likelihood information, suggesting that they may have relied on prior information instead. Figure shows an example participant. Each dot corresponds to the response on a given trial. B) Sensory weights per group (Left, NT- orange; right, AS-green) averaged across the four conditions. No group differences were found in likelihood reliance in the Main (2-Prior) Coin Task. C) Average sensory weights per group, separated by condition. The AS group shows reduced context adjustment compared to NT. D) Estimated subjective prior variances, divided per group and condition. No evidence was found for group differences in subjective prior variance. Conditions: P_NL_N = narrow prior and likelihood; P_NL_W = narrow prior, wide likelihood; P_WL_N = wide prior, narrow likelihood; P_WL_W = wide prior and likelihood.

Estimating a participant’s subjective prior mean

From equation (1), the intercept of the regression line (the term $σ L 2 σ L 2 + σ P 2 μ P$ M7 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\textstyle{{\sigma _L^2} \over {\sigma _L^2 + \;\sigma _P^2}}}\;{\mu _P}\] \end{document} ) can be rearranged to calculate the prior mean acquired by the participant:

(2)

μ P = β 0 1 − sw

M2 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\mu _{\rm{P}}} = \frac{{{\rm{\beta }_0}}}{{\left({1 - sw} \right)}}\] \end{document}

Estimating participant’s trial by trial sensory weights

Equation (1) can be rearranged to obtain a trial-by-trial slope (i.e., sensory weight)

(3)

Slope = X est − μ P μ L − μ P

M3 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\rm{Slope}} = \frac{{{X_{est}} - {\mu _{\rm{P}}}}}{{\left({{\mu _{\rm{L}}} - {\mu _{\rm{P}}}} \right)}}{\rm{\;}}\] \end{document}

Estimating participants’ subjective prior variance

Again, from equation (1), the slope or the sensory weight (sw) is equal to $σ P 2 σ L 2 + σ P 2$ M8 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[{\textstyle{{\sigma _P^2} \over {\sigma _L^2 + \;\sigma _P^2}}}\] \end{document} . This can be rearranged to obtain the participant’s subjective prior:

(4)

σ P 2 = σ L 2 ∗ sw 1 − sw

M4 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\sigma _P^2 = \frac{{\sigma _L^2 * sw}}{{\left({1 - sw} \right)}}\] \end{document}

In equation 2, σ²_L can be assumed as the true or objective likelihood variance (given by the variance of the 5 blue dots). or can be estimated from the subjective likelihood variance (σ²_Ls; described below in equation 3).

Estimating a proxy for subjective likelihood variance

The variance of the participant’s estimates of the mean (µ_est) relative to the true mean of the splashes (µ_L) on the Likelihood Only task can be determined as a proxy for the participant’s subjective likelihood variance $σ Ls 2$ M9 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\sigma _{Ls}^2\] \end{document} is as follows:

(5)

σ Ls 2 = Σ μ est − μ L) 2 nTrials

M5 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[\sigma _{Ls}^2 = \frac{{{\rm{\Sigma }}\left({{\mu _{{\rm{est}}}}} \right. - {\mu _{\rm{L}}}{)^2}}}{{nTrials}}\] \end{document}

Statistical Analysis

We aimed to understand if any of the current Bayesian models can explain sensory learning in autism. To test these models, we estimated the subjective prior and likelihood variance measures for each participant as described above.

For group analyses, we conducted 2 × 2 × 2 repeated measures analysis of variance for each measure of interest, with Prior (narrow vs wide) and Likelihood (narrow vs wide) and Group (AS vs NT) as factors. The different outcome variables analysed were: 1) Sensory weight or likelihood reliance; 2) estimation error (i.e., participants’ estimate minus the true position of the coin) for narrow and wide likelihood conditions as the outcome variables; 3) subjective prior variance for each condition, and 4) average confidence. Independent Samples T-tests were conducted for post-hoc pairwise comparisons. Where appropriate, Bayes factors (BF₀₁) are also reported in support of evidence for the null hypothesis. For all group tests estimated marginal mean (EM Mean) and effect size (partial eta²/η_p²) are reported.

For the dimensional analyses, we conducted bootstrapped Spearman rank correlations with AQ and SPQ scores with likelihood reliance and subjective prior variance for the main 2-prior task. Bootstrapped 95% confidence intervals for Spearman correlations reported are based on 1000 samples. For the Likelihood Only Task we excluded outliers based on Tukey’s 1.5 Interquartile Range. Based on this test 7 participants’ data (1NT and 3AS and 3 Other) were excluded.

In order to assess if continuum results were driven by group differences (AS vs NT vs other) we further, conducted multivariate analysis of covariance with AQ as covariate, Group as fixed factor and variable interest as the outcome variable to assess the interaction between group and AQ. Corrections for multiple comparisons are reported based on the Bonferroni correction procedure (p_bonf) alongside the uncorrected p value. Bonferroni corrections are applied considering all statistical comparisons with a trait of interest (e.g., AQ score) within a task (i.e., within Likelihood Only or within 2-Prior Task). Statistical analysis was conducted in SPSS version 26 and R. Figures are presented using ggplot2 () in RStudio.