Stanford Researchers Have Developed a Machine Learning-Based Algorithm To Detect Autism in Brain “Fingerprints”

This research summary article is based on the paper 'Robust, generalizable, and interpretable AI-derived brain fingerprints of autism and social-communication symptom severity' and the Stanford article 'New AI-Driven Algorithm Can Detect Autism in Brain “Fingerprints"' 

Please don't forget to join our ML Subreddit

Patients with autism who are diagnosed early and definitively may benefit from earlier therapies and better results.

Stanford researchers have created an algorithm that can tell if someone has autism by analyzing brain images. Inspired by current breakthroughs in artificial intelligence (AI), the unique system also accurately predicts the degree of autism symptoms in individuals. The algorithm might lead to faster diagnosis, more tailored therapy, and a better understanding of the brain’s roots in autism with further refinement.

The program sifts through data from functional magnetic resonance imaging (fMRI) images. These scans record patterns of neuronal activity in the brain. The system develops neural activity “fingerprints” by mapping this activity over time throughout the brain’s numerous areas. Although unique to each individual, just like actual fingerprints, brain fingerprints contain comparable characteristics that allow them to be sorted and categorized.

According to new research published in Biological Psychiatry, the algorithm evaluated brain scans from around 1,100 people. The algorithm correctly identified a group of patients with autism who had been diagnosed by human doctors.

Autism is one of the most prevalent neurodevelopmental diseases, yet there is still so much we don’t know about it. The AI-driven brain ‘fingerprinting’ model has the potential to be a powerful new tool in enhancing diagnosis and therapy.

Autism, unlike many other disorders, lacks objective biomarkers—telltale readings that identify the existence and severity of a medical condition—which means there is no straightforward test for the disorder. Instead, diagnosis is focused on monitoring patients’ naturally changeable actions, making diagnosis difficult. (Common symptoms of autism include difficulties managing daily social interactions, communication and learning problems, and repetitive speech and gestures.)

We need objective biomarkers for autism, and brain fingerprints get us one step closer. Scientists have traditionally used fMRI scans to look for biomarkers. However, studies with small populations have shown contradictory results, owing to inherent variability in patients’ brains and further complicated by variances in fMRI machines and testing methodologies.

Autism research, like many other scientific domains, has embraced the extensive data method, in which previously unattainable discoveries arise from studying vast, statistically robust samples. The present study combines brain scans from medical institutions into a massive, demographically and geographically varied dataset. The program sifts through data from functional magnetic resonance imaging (fMRI) images. These scans record patterns of neuronal activity in the brain. The system develops neural activity “fingerprints” by mapping this activity over time throughout the brain’s numerous areas.

The following stage was to efficiently analyze and deal with the data’s complexity and variability. Image identification algorithms built by technology businesses, according to colleagues, would be an excellent place to start. These algorithms have become increasingly adept at dealing with high levels of unpredictability in the photos they evaluate.

Consider an algorithm developed to recognize cats and dogs in internet photographs. That algorithm must deal with the animals being shot from various angles and distances and account for the wide variety of colors and traits seen in different breeds. It doesn’t matter if some 5-year-old took the picture or someone with a photography award—the algorithm has to function in both circumstances for image recognition AI to be effective. Brain scans show the same type of variability that you see in photographs of cats and dogs.


Researchers have concentrated on developing explainable AI, or XAI, instead of traditional AI systems, which may yield good outcomes but not in obvious ways. An issue has been that AI algorithms might be a ‘black box,’ where we can’t explain where the algorithm’s accuracy comes from.

Using the cat-versus-dog model as an example, researchers would need to know if the algorithm is selecting over the animals’ face traits or neck sizes. A basic mathematical model was created that evaluates brain regional relationships and interconnectivity for the brain fingerprinting technique. As a result, the XAI algorithm was applied to three brain areas that showed substantial variations in interconnectivity in a groupable component of the dataset. Those three brain areas previously linked with autism pathophysiology lend legitimacy to the XAI algorithm’s findings.

The regions are the posterior cingulate cortex and precuneus, which are part of the default mode network (DMN) and are particularly active during periods of wakeful rest. The dorsolateral and ventrolateral prefrontal cortex is involved in cognitive control. And the superior temporal sulcus processes human voice sounds. Disruptions to the DMN, in particular, were found to be significant predictors of autism symptom severity in the examined group.

While the XAI system worked excellently at this early stage of development, its accuracy needs to be increased before brain fingerprinting can be considered a valid biomarker. The researchers plan to test the algorithm’s effectiveness in sibling studies. One sibling has autism, and the other does not, to perfect the capacity to detect fine-tuned yet essential distinctions across potentially similar brains.

Using brain fingerprinting to analyze the brains of highly young infants, maybe as little as 6 months or a year old, who are at high risk of developing autism. Earlier diagnosis is crucial for improved results, with medications being more successful when begun when patients are still toddlers rather than later in childhood.

The method revealed in this study will allow identifying autism during the window of opportunity when therapies are most successful. You can read further about it here.



✅ [Featured Tool] Check out Taipy Enterprise Edition