Summary: A new AI algorithm can predict whether a person is on the autism spectrum by examining their brain scans. The algorithm can also predict the severity of symptoms and could be used as an early detection tool for ASD.
Stanford researchers have developed an algorithm that may help discern if someone has autism by looking at brain scans. The novel algorithm, driven by recent advances in artificial intelligence (AI), also successfully predicts the severity of autism symptoms in individual patients.
With further honing, the algorithm could lead to earlier diagnoses, more targeted therapies, and broadened understanding of autism’s origins in the brain.
The algorithm pores over data gathered through functional magnetic resonance imaging (fMRI) scans. These scans capture patterns of neural activity throughout the brain. By mapping this activity over time in the brain’s many regions, the algorithm generates neural activity “fingerprints.” Although unique for each individual just like real fingerprints, the brain fingerprints nevertheless share similar features, allowing them to be sorted and classified.
As described in a new study published in Biological Psychiatry, the algorithm assessed brain scans from a sample of approximately 1,100 patients. With 82% accuracy, the algorithm selected out a group of patients whom human clinicians had diagnosed with autism.
“Although autism is one of the most common neurodevelopmental disorders, there is so much about it that we still don’t understand,” says lead author Kaustubh Supekar, a Stanford clinical assistant professor of psychiatry and behavioral sciences and Stanford HAI affiliate faculty. “In this study, we’ve shown that our AI-driven brain ‘fingerprinting’ model could potentially be a powerful new tool in advancing diagnosis and treatment.”
Unlike many other diseases, autism lacks objective biomarkers—telltale measurements that reveal a medical condition’s presence and sometimes severity—meaning there is no simple test for the disorder. Instead, diagnosis is based on observing patients’ behaviors, which are naturally highly variable and thus make diagnosis a challenge. (Common signs of autism include difficulty navigating everyday social interaction, deficits in communicating and learning, and repetitive speech and motions.)
“We need to create objective biomarkers for autism,” says Supekar, “and brain fingerprints get us one step closer.”
Combining Big Data and XAI
Scientists have long searched for biomarkers via fMRI scans. Yet studies to date with small populations have reported conflicting results, stemming from natural variability in patients’ brains and confounded further by differences in fMRI machines and testing methods.
Like many scientific fields, autism research has embraced the big data approach, Supekar says, where previously unobtainable insights emerge from analyzing large, statistically powerful samples. Supekar’s new study is a case in point, pooling brain scans from medical centers worldwide into a mammoth, demographically and geographically diverse dataset.
The next step was to effectively parse and deal with the data complexity and variability. Supekar and colleagues thought a good place to start would be image recognition algorithms, developed by technology companies. These algorithms have grown increasingly sophisticated at handling significant degrees of variability in the images they assess.
For example, Supekar says, imagine an algorithm designed to identify cats and dogs in online images. That algorithm must contend with the animals being photographed from different angles and distances, as well as nimbly account for the ranges of colors and features among breeds.
“For image recognition AI to be successful, it doesn’t matter if my 5-year-old took the picture or someone with an award in photography—the algorithm has to work in both cases,” says Supekar. “The same kind of heterogeneity you get in pictures of cats and dogs, you get in brain scans, too.”
In deriving their image-recognition algorithms, Supekar and colleagues sought to make the artificial intelligence explainable, or understandable to human researchers. Researchers in recent years have focused on crafting explainable AI, or XAI, in contrast to conventional AI systems that might produce quality results but not in readily apparent ways.
“A challenge has been that AI algorithms can be a ‘black box,’ where we can’t explain where the accuracy of the algorithm comes from,” says Supekar.
Taking the cat-versus-dog example model again, researchers would want to know if the algorithm is picking over the animals’ facial features or neck sizes, say. For the brain fingerprinting algorithm, Supekar and colleagues fashioned a simple mathematical model that assesses brain regional interactions and interconnectivity. In this manner, the XAI algorithm alit upon three brain regions exhibiting significant differences in interconnectivity in a groupable portion of the dataset.
Lending credibility to the XAI algorithm’s findings, those three brain regions have been previously implicated in autism pathology. The regions are the posterior cingulate cortex and precuneus, which form part of the default mode network (DMN), notably active during periods of wakeful rest; the dorsolateral and ventrolateral prefrontal cortex, involved in cognitive control; and the superior temporal sulcus, involved in processing the sounds of human voices. In particular, disruptions to the DMN served as strong predictors of autism symptom severity in the studied population.
Earlier the better
While the XAI algorithm performed admirably at this early stage of development, Supekar and colleagues will need to improve its accuracy further still to raise brain fingerprinting to the level of a definitive biomarker. The researchers intend to explore the algorithm’s efficacy in sibling studies, where one sibling has autism and the other does not, to hone the ability to detect fine-tuned, yet critical differences between potentially very similar brains.
Supekar envisions brain fingerprinting being used to assess the brains of very young children, perhaps as early as six months or a year old, who are at high risk of developing autism. Earlier diagnosis is critical in achieving better outcomes, with therapies proving more effective when introduced while patients are still toddler-aged versus later in childhood
“We hope that the approach demonstrated in our study could diagnose autism during the window of opportunity when interventions are maximally most effective,” says Supekar.
About this AI and ASD research news
Author: Adam Hadhazy
Contact: Adam Hadhazy – Stanford
Image: The image is in the public domain
Original Research: Open access.
“Interpretable Artificial Intelligence–Derived Brain Fingerprints of Autism and Social Communication Symptom Severity” by Kaustubh Supekar et al. Biological Psychiatry
Interpretable Artificial Intelligence–Derived Brain Fingerprints of Autism and Social Communication Symptom Severity
Autism spectrum disorder (ASD) is among the most pervasive neurodevelopmental disorders, yet the neurobiology of ASD is still poorly understood because inconsistent findings from underpowered individual studies preclude the identification of robust and interpretable neurobiological markers and predictors of clinical symptoms.
We leverage multiple brain imaging cohorts and exciting recent advances in explainable artificial intelligence to develop a novel spatiotemporal deep neural network (stDNN) model, which identifies robust and interpretable dynamic brain markers that distinguish ASD from neurotypical control subjects and predict clinical symptom severity.
stDNN achieved consistently high classification accuracies in cross-validation analysis of data from the multisite ABIDE (Autism Brain Imaging Data Exchange) cohort (n = 834). Crucially, stDNN also accurately classified data from independent Stanford (n = 202) and GENDAAR (Gender Exploration of Neurogenetics and Development to Advanced Autism Research) (n = 90) cohorts without additional training. stDNN could not distinguish attention-deficit/hyperactivity disorder from neurotypical control subjects, highlighting the model’s specificity.
Explainable artificial intelligence revealed that brain features associated with the posterior cingulate cortex and precuneus, dorsolateral and ventrolateral prefrontal cortex, and superior temporal sulcus, which anchor the default mode network, cognitive control, and human voice processing systems, respectively, most clearly distinguished ASD from neurotypical control subjects in the three cohorts.
Furthermore, features associated with the posterior cingulate cortex and precuneus nodes of the default mode network emerged as robust predictors of the severity of core social and communication deficits but not restricted/repetitive behaviors in ASD.
Our findings, replicated across independent cohorts, reveal robust individualized functional brain fingerprints of ASD psychopathology, which could lead to more objective and precise phenotypic characterization and targeted treatments.