Pregistered movie-fMRI analyses reveal altered visual feature encoding in autism in pSTS

Using preregistered movie-fMRI encoding models, this study reveals that autistic children and adolescents exhibit reduced high-level visual feature representation and a relative shift toward low-level processing in social brain regions like the pSTS, supporting weak central coherence theories over early sensory enhancement hypotheses.

The Gist

The Big Picture: Watching a Movie to Understand the Autistic Brain

Imagine you are trying to understand how two different types of radio receivers work. One is a standard receiver (neurotypical brains), and the other is a specialized receiver (autistic brains). Both are tuned to listen to the same complex radio show (a movie).

For years, scientists had a theory that the "specialized" receiver was just super-powered. They thought autistic people heard every tiny crackle and saw every single pixel with super-sharp clarity (a theory called "Enhanced Perceptual Functioning").

This study says: "Actually, that's not quite right."

Instead of being a super-powered receiver, the autistic brain seems to be a receiver that is tuned differently. It's not that the signal is louder; it's that the brain is focusing on the details of the signal (like the static or the background noise) and missing the big story (the plot or the faces).

---

The Experiment: The "Stacked" Decoder

The researchers didn't just ask kids to sit still and look at a dot. They let them watch natural movies (like Despicable Me and The Present) while inside an MRI machine.

To understand what the brain was doing, they used a tool called an Encoding Model. Think of this like a recipe decoder:

1. The Ingredients: They broke the movie down into "ingredients."
   • Low-level ingredients: Brightness, motion, loudness, simple sounds.
   • High-level ingredients: Faces, bodies, speech, music, social cues.
2. The Decoder: They asked the computer: "Which ingredients does the brain care about most?"
3. The Stacking: They used a "stacked" model. Imagine you have two chefs. One only cooks with basic spices (low-level), and the other only cooks with complex sauces (high-level). The "stacked" model asks: "If we combine both chefs, how much of the final dish does each chef actually contribute?"

The Findings: What They Discovered

1. No "Super-Vision" in the Basics

The Myth: Autistic people see and hear basic things (like a flashing light or a loud beep) much better than everyone else.
The Reality: When the researchers looked at the "primary sensory" parts of the brain (the first stop for sight and sound), there was no difference. The autistic brain wasn't "super-charged" at the entry level. The volume knob wasn't turned up higher.

2. The "Social Hub" Got Stuck on Details

The Discovery: The big difference showed up in a specific area called the pSTS (posterior Superior Temporal Sulcus). You can think of this as the brain's "Social Integration Station." It's where the brain usually takes all the little details and combines them to understand a face, a gesture, or a conversation.

• In Neurotypical Brains: This station acts like a Director. It says, "Ignore the background noise; focus on the actor's face and what they are saying." It prioritizes the high-level story.
• In Autistic Brains: This station acts more like a Sound Engineer focused on the raw audio. It says, "Let's focus on the texture of the voice and the movement of the lips, but maybe we're missing the emotional context."

The autistic brain in this study showed a shift in priority: it was spending more energy on the "low-level" details (motion, brightness) and less on the "high-level" meaning (faces, social cues).

3. The "Volume" of Symptoms

The researchers found a direct link between this "tuning" and how severe a person's social symptoms were.

• Analogy: Imagine a radio dial. The further you turn the dial toward "Static/Details" (low-level), the harder it is to hear the "Music/Story" (high-level).
• The Result: The more a person's brain was tuned to the "details" in the pSTS, the higher their scores were on the Social Responsiveness Scale (SRS). It's like a volume knob: the more you focus on the pixels, the quieter the social story becomes.

4. No "Audio vs. Visual" War

There was a theory that autistic people might rely more on hearing than seeing (or vice versa), like a "Reverse Colavita Effect."
The Reality: The study found that the balance between hearing and seeing was mostly the same for both groups. The autistic brain wasn't suddenly becoming "deaf" to visuals or "blind" to audio. The difference was what they were paying attention to within those senses (details vs. meaning), not which sense they preferred.

5. Age Matters More Than Diagnosis

One of the most interesting findings was about growing up.

• Analogy: Think of the brain like a city. When you are a child, the roads are a bit messy, and traffic goes everywhere. As you get older, the city builds better highways.
• The Result: As children got older, their brains naturally got better at separating "visual roads" from "audio roads." This developmental change was actually stronger than the difference between autistic and non-autistic groups. The brain is constantly rewiring itself as it matures, and this study captured that journey.

The "ADHD" Twist

The researchers also noticed something interesting about ADHD.

• The "detail-focused" shift in the autistic brain was mostly driven by autistic kids without ADHD.
• Autistic kids with ADHD looked more like the non-autistic group in how they processed these movies.
• Takeaway: This suggests that autism and ADHD might affect the brain in different ways, and mixing them together in studies can sometimes hide the true picture.

The Conclusion: A New Way to Listen

This study changes the story. It suggests that autism isn't about having "superpowers" in basic senses. Instead, it's about how the brain weighs information.

In the social parts of the brain, the autistic brain seems to be over-weighting the details (the pixels, the motion) and under-weighting the big picture (the face, the emotion). It's not that the signal is broken; it's that the brain is prioritizing the "ingredients" over the "recipe."

By using movies and these advanced "decoders," scientists can now see exactly where and how this weighting happens, opening the door to better understanding and support for neurodevelopmental differences.

Technical Summary

1. Problem Statement & Research Context

Sensory-perceptual differences are a hallmark of autism spectrum disorder (ASD), affecting up to 90% of individuals. However, the neurobiological mechanisms underlying these differences remain unclear. Two dominant theories offer conflicting predictions:

• Enhanced Perceptual Functioning (EPF): Predicts heightened encoding of low-level sensory features in primary sensory cortices.
• Weak Central Coherence (WCC): Predicts a preference for local (low-level) over global (high-level) processing, potentially leading to reduced high-level integration in association areas.

Previous neuroimaging studies have largely relied on static stimuli or simple connectivity metrics, leaving a gap in understanding how information is represented during rich, dynamic, naturalistic perception. This study aims to resolve these theoretical conflicts by testing preregistered hypotheses regarding how autistic children and adolescents encode low-level vs. high-level auditory and visual features during naturalistic movie viewing.

2. Methodology

Dataset & Participants:

• Source: Healthy Brain Network (HBN) dataset (Releases 1–10).
• Participants: Children and adolescents (ages ~5–21).
  • ASD Group: $N=108$ (after quality control), including those with and without ADHD comorbidity.
  • Control Group: $N=63$ (non-ASD, no ADHD).
• Stimuli: Naturalistic movie clips (Despicable Me and The Present) viewed during fMRI scanning.
• Behavioral Measures: Social Responsiveness Scale (SRS-2) for autism severity and a derived "Sensory Subset Score" (SSS) for sensory symptoms.

Feature Engineering:
The authors extracted low- and high-level features from the audio and visual streams:

• Low-Level Visual: Mean brightness and mean motion energy.
• High-Level Visual: Presence of faces and bodies (hand-coded).
• Low-Level Audio: Perceptual loudness and cochleagram principal components.
• High-Level Audio: Categorical labels (e.g., Speech, Music, Animal Sounds) derived from AudioSet/YAMNet.

Analytical Approach: Stacked Encoding Models
Instead of standard univariate activation analysis, the study employed stacked encoding models:

1. Base Models: Separate ridge regression models were trained for low-level and high-level features (and audio vs. visual) to predict brain activity at every grayordinate.
2. Stacking: A secondary "stacking" regression combined the predictions of the base models.
3. Metrics:
   • $R^2$ & Unique $R^2$ ($R^2_u$): Explained variance and variance unique to a feature class.
   • Stacked Weights ($W$): The relative contribution of each feature space to the joint prediction.
   • Preference Indices: Calculated as $W_{High} - W_{Low}$ (for level preference) and $W_{Visual} - W_{Audio}$ (for modality preference).
4. Statistical Framework: Linear mixed-effects models (GLMM) were used to test group differences (ASD vs. non-ASD) and correlations with SRS/SSS, controlling for age, sex, SES, and site. Analyses were run across three framewise displacement (FD) motion thresholds (40%, 60%, 80%) to ensure robustness.

3. Key Contributions

• Methodological: Demonstrates the utility of stacked encoding models applied to naturalistic stimuli to disentangle low-level vs. high-level feature representations in the human brain, moving beyond simple activation maps.
• Theoretical: Provides empirical evidence favoring Weak Central Coherence (WCC) over Enhanced Perceptual Functioning (EPF) regarding visual processing in autism, specifically locating the deficit in higher-order social integration hubs rather than primary sensory cortices.
• Clinical: Identifies a specific neural signature in the posterior Superior Temporal Sulcus (pSTS) that tracks with social symptom severity (SRS scores).

4. Key Results

A. Low-Level vs. High-Level Encoding (Hypothesis 1)

• Primary Sensory Cortices: Contrary to EPF predictions, there was no evidence of increased low-level encoding in primary visual or auditory cortices in the ASD group.
• Posterior STS (pSTS): The ASD group showed a significant reduction in high-level visual representations and a relative shift toward low-level visual feature encoding in the pSTS (specifically STSvp and STSdp).
• Clinical Correlation: In the pSTS, the preference for low-level over high-level visual features ($W_{Low} - W_{High}$) was significantly correlated with SRS scores (higher autism severity = stronger low-level bias).
• ADHD Stratification: Post-hoc analysis revealed that these visual encoding differences were primarily driven by the ASD-only subgroup. The ASD+ADHD subgroup did not differ significantly from controls in these visual metrics.

B. Audio-Visual Modality Preference (Hypothesis 2)

• Conserved Dominance: Contrary to behavioral reports of "reverse Colavita effects" (auditory dominance), the study found no significant group differences in audio-visual modality preference ($W_{Visual} - W_{Audio}$) or unimodal dominance in sensory cortices.
• Conclusion: The balance between auditory and visual feature weighting appears broadly conserved in ASD during passive naturalistic viewing.

C. Developmental Effects

• Age is a Major Driver: Age had a stronger effect on encoding metrics than diagnosis.
• Maturation: Encoding performance increased with age in modality-congruent regions (e.g., visual features in visual cortex). Modality preference became more congruent with regional specialization as children aged.
• Interaction: Age-by-diagnosis interactions were sparse and localized, suggesting that the observed ASD differences are not merely a developmental delay but a distinct reweighting of features in specific social brain regions.

5. Significance & Conclusion

This study refines the understanding of sensory processing in autism by shifting the focus from "enhanced primary sensory input" to altered feature weighting in higher-order integration networks.

• Mechanism: The findings suggest that the core deficit in autism may not be an inability to process basic sensory data, but rather a failure to appropriately weight high-level, socially relevant visual information (such as faces and biological motion) in the pSTS. This leads to a "local-over-global" bias specifically within social-perceptual networks.
• Implications for Theory: The results strongly support Weak Central Coherence and predictive coding accounts (altered priors/precision weighting) while challenging the notion of pervasive early sensory enhancement.
• Future Directions: The study highlights the critical need to account for ADHD comorbidity and developmental trajectories in sensory research. It also underscores the value of naturalistic stimuli combined with encoding models to capture the complexity of real-world perception, which static stimuli often miss.

In summary, the paper provides a rigorous, preregistered demonstration that autism involves a specific reorganization of how the brain prioritizes visual information in social brain regions, favoring low-level details over high-level social integration, a mechanism that scales with symptom severity.