High internal noise and poor external noise filtering characterize perception in autism spectrum disorder

Abstract

An emerging hypothesis postulates that internal noise is a key factor influencing perceptual abilities in autism spectrum disorder (ASD). Given fundamental and inescapable effects of noise on nearly all aspects of neural processing, this could be a critical abnormality with broad implications for perception, behavior, and cognition. However, this proposal has been challenged by both theoretical and empirical studies. A crucial question is whether and how internal noise limits perception in ASD, independently from other sources of perceptual inefficiency, such as the ability to filter out external noise. Here, we separately estimated internal noise and external noise filtering in ASD. In children and adolescents with and without ASD, we computationally modeled individuals' visual orientation discrimination in the presence of varying levels of external noise. The results revealed increased internal noise and worse external noise filtering in individuals with ASD. For both factors, we also observed high inter-individual variability in ASD, with only the internal noise estimates significantly correlating with severity of ASD symptoms. We provide evidence for reduced perceptual efficiency in ASD that is due to both increased internal noise and worse external noise filtering, while highlighting internal noise as a possible contributing factor to variability in ASD symptoms.

Figure 1

Predictions from PTM and experimental procedure. (a) Typical threshold-vs-noise (TvN) functions and predictions from the PTM. TvN functions have a characteristic shape where a flat segment is followed by a rising segment. At low external noise, internal noise dominates such that added external noise has little effect on perceptual performance. This produces the flat portion of the TvN curve. Once the external noise level begins to exceed that of internal noise, perceptual thresholds will be affected by external noise. Consequently, thresholds will increase because stronger signal is needed to overcome high levels of external noise. The PTM predicts distinct changes in TvN depending on the source of noise that is atypical in ASD (blue curves), relative to a TD baseline (red curves). Elevated internal additive noise will result in increased thresholds at the flat portion of the curve (top left). Worse external noise filtering, on the other hand, causes increased thresholds at the rising portion of the curve (top right). A combination of the two should yield increased thresholds across all levels of external noise levels (bottom left). A similar pattern is expected if internal multiplicative noise is elevated in ASD. However, the two can be distinguished by comparing the curves across at least two different difficulty levels. The ratio between thresholds at two difficulty levels will be different across groups in the case of elevated internal multiplicative noise in ASD (bottom right). (b) Stimuli, task, and timeline. Each trial began with a dynamic fixation point (see Methods). After a blank screen, the stimulus sequence started. Stimuli were oriented gratings temporally sandwiched by Gaussian pixel noise. This sequence yields merged perception of grating and noise (see insert). Participant’s task was to judge whether the grating was tilted left or right from vertical.

Figure 2

Results from the conventional PTM analysis. The two panels show thresholds (filled dots) at each external noise level for two difficulty levels. Overall, individuals with ASD (blue) performed worse than those with TD (red) across external noise levels. Error bars are ±SEM. Curves were obtained by fitting the conventional PTM (Equation (3) in Methods), which revealed a 70% increase in internal additive noise and 13% worse external noise filtering in ASD (p < 0.001; compared to the model that assumed no group differences).

Figure 3

Results from hierarchical Bayesian analysis. (a) TvN functions estimated from the hierarchical Bayesian analysis show worse performance in ASD (blue) compared to TD (red), consistent with the conventional PTM analysis. Shaded regions are the 68% credible intervals obtained from the posterior distributions. (b) The average individual parameter estimates from the hierarchical Bayesian model. The estimated population means show 48% elevated internal additive noise, and 25% greater effect of external noise (i.e., worse external noise filtering) in ASD (+marginal significance using a 90% criterion on difference in population posterior samples; *significance using a 95% criterion). No significant group difference was found for internal multiplicative noise. Note that the absolute values of different types of noise cannot be meaningfully compared. Error bars are ±SEM. (c) Population posterior distributions estimated from the hierarchical Bayesian model for internal additive noise (top) and external noise filtering (bottom). Horizontal lines indicate 95% credible intervals. In both cases, the population posterior distributions for ASD (blue) are broader than TD (red), indicating greater inter-individual variability in ASD. In the case of internal additive noise, the 95% credible interval for ASD fully includes and extends beyond the credible interval for TD.

Figure 4

Relationship between the model parameters and ADOS calibrated severity score within ASD. Correlation analyses indicated that symptom severity was related to internal additive noise (left panel; r₁₉ = 0.59, p = 0.005) and not to external noise filtering (right panel; r₁₉ = 0.28, p = 0.22; see main text for results from multiple regression analysis which include both noise estimates in the model).