Autism spectrum disorder has long been studied through the lens of group averages. For decades, researchers have compared autistic volunteers with neurotypical volunteers on a handful of tasks and reported where the two groups differ on average, treating the resulting gap as the signature of the condition. That approach has produced a large and useful literature, but it carries an assumption that is rarely tested directly: that the average describes something real about most people in the group, rather than a statistical summary that conceals wide person to person spread.
Naturalistic neuroimaging has made that assumption harder to defend. Instead of showing participants isolated faces or static shapes, modern experiments play films and record whole brain activity with functional magnetic resonance imaging, or fMRI. Films carry sound, motion, narrative and social meaning at the same time, and they hold attention in a way that reduced laboratory stimuli often do not. The cost is complexity, because the stimulus cannot be described by a single variable, and analysis depends on computational models that learn which features of the audio and video drive which brain responses.
Those models have become good enough to predict activity in held out data, which makes them a test bed for individual differences. If a model trained on many people predicts a new person well, that person's brain is behaving like the group. If the model predicts the group average well but individual patterns scatter, the average is hiding heterogeneity. The distinction matters for a condition as varied as autism, where clinicians have argued for years that two people with the same diagnosis can present very differently.
A new study led from the University of Iowa takes that test seriously. The work, published in The Journal of Neuroscience and presented at the Society for Neuroscience annual meeting in Washington, D.C., on Sunday, 15 November 2026, compared autistic and neurotypical adults watching short films and asked not only whether group averages looked alike, but whether individuals inside each group looked alike. The answer, according to the authors, is that the averages converge while the individuals do not.
Key Facts
The paper is titled Heterogeneous Encoding Patterns of Dynamic Audiovisual Social Stimuli in Autism Spectrum Disorder and carries the DOI 10.1523/JNEUROSCI.0055-26.2026. Its authors are Avery L. Van De Water, Lisa Byrge, Daniel P. Kennedy, Dorit Kliemann and James A. Traer, with Kliemann and Traer of the University of Iowa leading the collaboration. The Journal of Neuroscience reported on November 15 that the article appeared in its early access pages dated 21 October 2026.
The dataset covered 138 adults: 48 autistic participants, 13 of them female, and 90 neurotypical participants, 25 of them female. None of the scans were collected for this paper. The team analyzed data previously gathered at Indiana University and Caltech while participants watched six videos lasting between 5 and 20 minutes, spanning animation, action, romantic comedy and Hitchcockian thriller. EurekAlert! reported on November 15 that the release describing the work came from the Society for Neuroscience.
The analysis relied on encoding models. The researchers trained a computational model to relate measured brain activity to high dimensional audio features, visual features and abstract social features such as valence, the pleasantness or unpleasantness of a moment. Once trained, the model predicted activity significantly above chance across brain networks and across videos. Auditory and language networks were the most accurately predicted, which is unsurprising given how strongly dialogue and sound drive those systems.
The headline comparison was between groups, and it came out nearly flat. Group averaged encoding accuracy was strikingly consistent between autistic and neurotypical adults, and within participant prediction quality was equally variable in the two groups. The divergence appeared one level down. Autistic participants showed lower within group encoding pattern similarity than neurotypical participants, and their patterns were more similar to the other group's average than to their own group's average. The effect was not confined to one system: variability rose in nearly all brain systems assessed and was most pronounced in the social brain system. News-Medical reported on November 16 that analyses restricted to socially relevant features produced the largest across group differences in encoding pattern similarity.
Analysis
The most interesting number here is not a difference but a similarity. Two groups separated by a clinical diagnosis produced nearly identical average encoding accuracy, a result the researchers themselves describe as surprising. Traer said in the release that the team was surprised by the similarity of group averages and that the field needs to understand how and why individual brain responses differ. A null result at the group level, paired with a positive result at the individual level, is a specific kind of finding: it says the conventional comparison was not wrong, only blunt.
What this really means is that the standard case control design in autism neuroimaging may be systematically understating the biology it aims to describe. If autistic adults vary more among themselves than neurotypical adults do, an average difference between groups can stay small even when the individual differences underneath are large, because the autistic distribution is wider and its mean is a poor summary of its members. The authors argue that idiosyncratic individual patterns must be incorporated into translational neuroimaging approaches to autism, rather than treated as noise around a group signal.
The social brain result deserves particular attention. The largest across group divergence came from models built only on socially relevant features, which means the heterogeneity is concentrated where the diagnosis is defined. That is a coherent story, since social information is exactly the kind of input where personal history, learned strategies and co-occurring conditions would be expected to shape neural response. It also cautions against reading a group level social brain difference as a mechanism, because the same mean can be produced by very different individual arrangements. The Journal of Neuroscience reported on November 15 that cross-validated models were assessed on prediction accuracy, variance and pattern similarity, which is how the team could compare people against each other rather than only against a mean.
The bigger picture here is methodological, and it extends past autism. Naturalistic paradigms plus encoding models give researchers a way to measure how alike two people are, not just how far each sits from a mean. That is a different measurement with different statistical consequences, and it is likely to be applied well beyond this one diagnosis, including to conditions where heterogeneity is suspected but hard to demonstrate.
Why It Matters
Current approaches to understanding and treating autism spectrum disorder rely heavily on group average comparisons, a point the researchers make explicitly. If individual variation is as large as this dataset suggests, a treatment validated against a group mean may work well for one subset and poorly for others, with the average hiding both outcomes. Kliemann said identifying subgroups of autistic people may help build a more precise understanding of the condition and develop individually tailored support. That is a research agenda rather than a clinical recommendation, but it reframes what a useful study would look like.
The result also bears on how imaging findings are communicated. A report that autistic and neurotypical brains respond similarly to films is accurate at the group level and misleading at the individual level, and both statements come from the same 138 participants. News-Medical reported on November 16 that its coverage emphasized preserving individual level variation rather than treating it as noise, a framing the authors endorse. For a community that has often objected to being reduced to a single profile, that distinction is not academic.
Next Up
The obvious next step is subgroup identification. Kliemann's suggestion that distinct ASD subgroups may exist points toward clustering participants by encoding pattern rather than by diagnosis, then asking whether those clusters track behavior, sensory profile or support needs. Traer's question, how and why individual brain responses differ, requires longitudinal designs and better control of what participants bring to a film: attention, memory, prior familiarity with a genre, and the strategies people use to interpret social scenes.
Replication in independent samples will also matter, since the current analysis reuses data from Indiana University and Caltech and the autistic sample, at 48 adults, is large for neuroimaging but small for subgroup work. The paper was presented at Neuroscience 2026, the Society for Neuroscience annual meeting in Washington, D.C., on 15 November 2026. Whether the field adopts individual level metrics as standard practice is now an empirical question rather than a rhetorical one.
Comments (0)
Log in or sign up to leave a comment.
No comments yet. Be the first to share your thoughts.