Slow update of internal representations impedes synchronization in autism

Abstract

Autism is a neurodevelopmental disorder characterized by impaired social skills, motor and perceptual atypicalities. These difficulties were explained within the Bayesian framework as either reflecting oversensitivity to prediction errors or - just the opposite - slow updating of such errors. To test these opposing theories, we administer paced finger-tapping, a synchronization task that requires use of recent sensory information for fast error-correction. We use computational modelling to disentangle the contributions of error-correction from that of noise in keeping temporal intervals, and in executing motor responses. To assess the specificity of tapping characteristics to autism, we compare performance to both neurotypical individuals and individuals with dyslexia. Only the autism group shows poor sensorimotor synchronization. Trial-by-trial modelling reveals typical noise levels in interval representations and motor responses. However, rate of error correction is reduced in autism, impeding synchronization ability. These results provide evidence for slow updating of internal representations in autism.

Fig. 1. Isochronous finger tapping: mean asynchrony is similar in the three groups, but variability around this mean is substantially larger in Autism (ASD) compared to neurotypical (CON, control) and dyslexia groups (DYS).

a A schematic illustration of the temporal structure of paced tapping: metronome stimuli (presented every 500 ms, black squares), and finger-tap responses (blue circles) as a function of time; $e_k$ - error (asynchrony, typically negative) in tap k; $r_k$ - inter-tap interval; $d_k$ - delay interval from the previous metronome stimulus (beat k-1) to the following finger tap (tap k). Note that $r_k=d_k-e_{k-1}$. b, c Basic tapping parameters: b Mean asynchrony is negative for all three groups (p < 0.001) and similar in the three populations, though more broadly distributed in the ASD group. c Standard deviation is larger in the ASD group compared to the two other groups. Each dot represents the performance of one participant (average of two blocks); the y-axis represents the score in ms, and x-axis and color represent group membership (with a small jitter for readability): blue circles—neurotypical, red triangles—dyslexia, and green squares—ASD. The median of each group is denoted as a line of the same color; error bars around this median denote an interquartile range. Kruskal–Wallis H-statistic and corresponding p values are plotted in the bottom-left corner; p values of comparisons between groups are plotted next to the line connecting the groups’ medians. N = 109 subjects (N_CON = 47, N_DYS = 32, N_ASD = 30). Source data are provided as a Source Data file. Though there are a few outlier results in both mean asynchrony and standard deviation of participants with ASD, these are not the same individuals—scores on these two measures were not correlated in the ASD and dyslexia groups (Spearman correlations: ${\rho }_{ASD}=-0.2$ (p = 0.3), ${\rho }_{DYS}=-0.24$ (p = 0.18)). A significant correlation was found only in the neurotypical group (${\rho }_{CON}=-0.37$, p = 0.01, uncorrected). Statistical tests are two-sided unless stated otherwise.

Fig. 2. Correlation between consecutive asynchronies (errors) is highest in the ASD group revealing reduced online error correction.

a–c Scatter plots showing correlations between consecutive asynchronies: a neurotypical (CON, control), b dyslexia (DYS), and c ASD. Individual asynchronies were plotted with respect to each participant’s mean asynchrony, yielding a mean of 0 ms. Consecutive asynchronies are positively correlated in all groups. This positive correlation is largest in the ASD group, reflecting reduced online error correction. Luminance scale is equal in (a–c): white, the maximum number of asynchronies in a bin, is 165 in all graphs. d Single participant correlations also show the impairment in error correction for the ASD group compared with the neurotypical and dyslexia groups. The median of each group is denoted as a line of the same color; error bars around this median denote an interquartile range. Kruskal–Wallis H-statistic and the corresponding p value are plotted in the bottom-left corner; p values of comparisons between groups are plotted next to the line connecting the groups’ medians. N = 109 subjects (N_CON = 47, N_DYS = 32, N_ASD = 30). Source data are provided as a Source Data file.

Fig. 3. Trial-by-trial computational modeling of isochronous tapping: Parameters estimated for each participant show that individuals with autism have reduced error correction and intact timekeeper and motor noise.

a Schematic illustration of the computational model used to dissociate error correction mechanisms from poor timekeeping or motor noise^,,. Each tapping interval (blue empty arrow) is assumed to be the summation of three mechanisms: (1) error correction based on the previous asynchrony (marked in red, the magnitude of the correction is determined by the phase correction parameter α) (2) timekeeping of the base tempo $T_k$ (composed of a fixed $t_0$, purple, plus the noise at tap k, $n_k$, green), and (3) motor noise (turquoise). See also notations in Fig. 1a. Fitting was performed using the bGLS (bounded General Least Squares) estimation method. b Error correction of phase difference—the fraction corrected (α) is significantly smaller in the ASD group. c Noise in keeping the metronome period, and d Motor noise do not differ between the groups. b–d Each block was modeled separately, and parameters were averaged over the two assessment blocks. The median of each group is denoted as a line of the same color; error bars around this median denote an interquartile range. Kruskal–Wallis H-statistic and corresponding p value are in the bottom-left corner; p values of comparisons between groups are next to the line connecting the groups’ medians. CON control (neurotypical), DYS dyslexia, ASD autism. N = 108 subjects (N_CON = 47, N_DYS = 32, N_ASD = 29), one ASD participant was excluded due to a large number of missing taps (see Methods). Source data are provided as a Source Data file.

Fig. 4. Individuals with autism adapt to changes in tempo only partially, even when changes are very salient.

a, b 90 ms step-size, c, d 70 ms step-size, and e, f 50 ms step-size. In each panel, the x-axis represents the metronome-beat number around the moment of tempo change (beat 0), and the y-axis measures the delay interval in each beat aligned to the pre-change metronome (mean group values, ±SEM; values were calculated by first averaging responses within each participant and then across the group; error bars denote SEM across participants). The dashed lines represent the metronome beat. Changes are quickly corrected, particularly for the larger steps (panels a–d). Reduced updates are seen for the smaller 50 ms step changes (panels e, f), where neurotypicals (CON control) take three–four steps to correct, and individuals with dyslexia (DYS) take longer, perhaps since these steps are less salient. The difficulties of individuals with autism (ASD) are seen in all step changes (including the smallest step-size, panels e, f), and their error is not fully corrected even within seven taps. Each participant tapped through eight-ten accelerations and eight-ten decelerations in each condition. Sample sizes: a, b 90 ms step-size: N_CON = 46, N_DYS = 31, N_ASD = 29 for acceleration and N_ASD = 26 for deceleration. c, d 70 ms step-size: N_CON = 47, N_DYS = 31, N_ASD = 29 for acceleration and N_ASD = 27 for deceleration. e, f 50 ms step-size: N_CON = 47, N_DYS = 32, N_ASD = 29 for acceleration and N_ASD = 25 for deceleration. See Methods for the exclusion criteria. Source data are provided as a Source Data file.

Fig. 5. Distributions of delay intervals 5–12 taps after the tempo switch show that asynchronies in the autism group remain uncorrected.

a–c, e–g and i–k show group delay intervals probability density functions separately for the longer tempo (light color) and shorter tempo (dark color), for each tempo-change condition (90 ms—top, 70 ms—middle, 50 ms—bottom) and for each population (CON (control, neurotypical)—blue, DYS (dyslexia)—red, ASD (autism)—green). The mean of each distribution is denoted by a black vertical line. Values of d′ and the difference between the means (diff. means) are in the top right corner. d, h, l show for each group the receiver-operator-curves (ROC), and area under the curve (AUC) for classification between delay intervals under the two tempos. In the 70 and 90 ms step-sizes, the measures of dyslexia and neurotypical groups nearly overlap, while the values for the ASD group are smaller, reflecting reduced updating to changes in external tempo. For the small 50 ms step-size, the dyslexia group values are midway between the neurotypical and autism values, though this difference was not significant (see main text). Source data are provided as a Source Data file.

Fig. 6. Trial-by-trial computational modeling of tapping with changing tempos: Parameters estimated for each participant show that individuals with ASD have reduced period correction when the period changes abruptly.

a Illustration of the dynamics added to the isochronous computational model (Fig. 3a), which enables tracking of changing tempos. The internal period estimate ($t_k$, purple) is adjusted in each trial based on the recent asynchrony (in red, the magnitude of the correction is determined by the period correction parameter β). To produce the tapping interval, noise is added to this estimate ($n_k$, green) as in the isochronous model. b–e: Each dot represents the combined value from all conditions of tempo-switches per participant (50–90 ms; after z-scoring). CON control (neurotypical), DYS dyslexia, ASD autism. N = 109 subjects (N_CON = 47, N_DYS = 32, N_ASD = 30). b Period correction (β) is smaller in the autism group compared with the two other groups, c while phase correction (α) is similar in this Experiment. d Timekeeper noise and e Motor noise estimates do not differ between the groups. The median of each group is denoted as a line of the same color; error bars around this median denote an interquartile range. Kruskal–Wallis H-statistic and corresponding p value are in the bottom-left corner; p values of comparisons between groups are next to the line connecting the groups’ medians. Source data are provided as a Source Data file.

Fig. 7. Rate of online error correction in stationary and in changing environments reflect a single updating mechanism.

The estimated phase correction from Experiment 1 and the estimated period correction from Experiment 2 are highly correlated in all groups a neurotypical (CON control), b dyslexia (DYS), c autism (ASD), suggesting that both are manifestations of a common underlying mechanism of error correction, which determines the speed of integrating new sensory data to guide behavior. The significance of Spearman correlation was calculated using a two-sided test, p values are uncorrected. Overlayed regression lines, predicting phase correction (Experiment 1) from period correction (Experiment 2) with an intercept term. d The combined update rate is significantly smaller in the ASD group but does not differ between the neurotypical and dyslexia groups. The median of each group is denoted as a line of the same color; error bars around this median denote an interquartile range. Kruskal–Wallis H-statistic and the corresponding p value are plotted in the bottom-left corner; p values of comparisons between groups are plotted next to the line connecting the groups’ medians. N = 108 subjects (N_CON = 47, N_DYS = 32, N_ASD = 29), one ASD participant was excluded from the computational modeling of Experiment 1 due to a large number of missing taps (see Methods). Source data are provided as a Source Data file.

Fig. 8. Update rate is correlated with communication and mindreading skills in both autism and neurotypical groups.

a Distribution of communication\mindreading skills was measured using a factor of the AQ50^, in the neurotypical (CON control) and autism (ASD) groups. b–d Update rate is correlated with communication\mindreading skills in the neurotypical (b) and autism (c) groups, and when combining both groups together (d). Each dot represents a single participant (neurotypical N_CON = 37, ASD N_ASD = 19). Lines represent regression lines predicting update rate from communication and mindreading skills with an intercept term. Blue and green lines (b, c) are based on the neurotypical and autism groups respectively, the black line (d) is based on data from both groups combined. Source data are provided as a Source Data file (b–d).