AI is diagnosing autism spectrum disorder. Here’s how—and what—the experts say
Key Takeaways
A study published in the Journal of the American Medical Association found that artificial intelligence (AI) could accurately distinguish between people with autism spectrum disorder (ASD) and typical development by analyzing retina photographs.
The study evaluated 1,890 eyes belonging to 958 participants, half with ASD and half with TD. The authors state that people with ASD have “structural retinal changes that potentially reflect brain alterations, including visual pathway abnormalities through embryonic and anatomic connections.”
Experts think that this sort of technology could aid clinicians by making faster and more accurate diagnoses, but that it shouldn’t replace the human clinician’s empathy, intuition, and contextual understanding of ASD.
One in 36 people have autism spectrum disorder (ASD), which can cause social communication challenges as well as repetitive interests or behaviors, ranging in severity. This number may be higher, however: Research shows that 25% of children under the age of eight—predominantly from minority backgrounds—go undiagnosed, thereby preventing early support.[][]
ASD can be challenging to diagnose accurately and early on, according to a review published by Cureus. These challenges are rooted in ASD’s unique presentation in each patient, making it hard to develop a universally comprehensive diagnostic test.[]
To help address this barrier, a recent study published in the Journal of the American Medical Association ( JAMA) explores how deep learning algorithms, a type of artificial intelligence (AI), could be used to identify ASD through retinal photography analysis that is designed to differentiate between individuals with ASD versus “typical development” (TD) and those with severe ASD versus mild to moderate ASD.[]
The use of retinal analysis stems from the fact that “[i]ndividuals with ASD have structural retinal changes that potentially reflect brain alterations, including visual pathway abnormalities through embryonic and anatomic connections,” the authors state.
At Severance Hospital, Yonsei University College of Medicine in Seoul, researchers analyzed 1,890 eyes of 958 participants. The study included 479 participants with ASD and 479 participants with TD. Most were boys (81.8%), whose mean age was 7.8 years. The retinal photographs of patients with ASD were collected between April and October 2022, and photographs of age- and sex-matched TD participants were collected between December 2007 and February 2023.[]
To analyze the eyes, “Deep ensembles of 5 models were built with 10-fold cross-validation using the pre-trained ResNeXt-50 (32×4d) network,” the authors write. “Score-weighted visual explanations for convolutional neural networks, with a progressive erasing technique, were used for model visualization and quantitative validation.” The data was then analyzed between December 2022 and October 2023.
The researchers state that the AI model could accurately identify the eyes of patients with ASD, with a mean area under the receiver operating characteristic curve (AUROC) of 1.00. A curve of 0 would have indicated inaccuracy.
"Our models had promising performance in differentiating between ASD and TD using retinal photographs, implying that retinal alterations in ASD may have potential value as biomarkers. Interestingly, these models retained a mean AUROC of 1.00 using only 10% of the image containing the optic disc, indicating that this area is crucial for distinguishing ASD from TD," the authors state.
What does this technology mean for clinicians?
Ryan Sultan, MD, a board-certified psychiatrist who works with children, says that AI’s potential for enhancing the screening and diagnosis of conditions like autism is undeniable.
“Currently, our field employs various scales and tests, alongside clinical observation and history, to diagnose autism,” Dr. Sultan says. “Integrating AI into these methods, as we're doing at Integrative Psych and our Lab for Mental Health Informatics at Columbia University in New York, shows promise in saving time for psychiatrists and therapists—an important achievement given the shortage of mental health providers and the growing mental health pandemic.”
It’s all about balance; Dr. Sultan emphasizes: “Early results indicate improved patient outcomes with AI assistance. [But]...even if AI reaches high levels of sensitivity and specificity in diagnosis, it should not overshadow the value of clinical experience. We must establish guardrails to ensure that AI remains a supplementary tool, enhancing rather than replacing the nuanced understanding and judgment that experienced clinicians bring to patient care,” Dr. Sultan says.
“This approach allows us to leverage the efficiency and analytical capabilities of AI while preserving the irreplaceable human elements of empathy, intuition, and contextual understanding in mental health care,” he continues.
Sharief Taraman, MD, who is board-certified in neurology and holds special qualifications in child neurology, echoes Dr. Sultan’s statements, noting that AI may be able to improve on the work clinicians are doing, “to detect features of autism that we are aware of more effectively, efficiently, and equitably.” He also notes that other AI diagnostic tools, including Cognoa's Canvas Dx and EarliPoint, are used to identify autism.
Dr. Taraman says that the study may also give clinicians new insights into the biology of autism. “Given the extreme specialist shortages [in] identify[ing] autism and evaluat[ing] developmental delays, AI has a significant promise to help us, and adoption is already happening,” Dr. Taraman says.
However, he adds, the study only looked at a single biomarker to identify autism. “As a clinician who evaluates children with autism, it is extremely important to look at the whole child,” he says.