AI-model diagnoses childhood autism from eye pics with 100% accuracy
The identification of autism spectrum disorder (ASD) faces challenges due to resource constraints, especially the shortage of trained professionals for assessments. People with ASD exhibit structural changes in the retina, which may mirror underlying brain alterations, including anomalies in the visual pathway stemming from embryonic and anatomical connections.
Aiming to make such tests more accessible and reliable, researchers in Korea have now developed a solution utilizing deep learning algorithms to screen for ASD objectively and assess symptom severity based on retinal photographs.
Deep ensemble models were created by researchers with an expanded participant pool. Additionally, their potential applicability in a pediatric population was assessed through sequential age-based modeling.
The details regarding the team’s study were published in the journal Jama Network Open.
Detailed process
ASD presents with two primary symptom categories, which are social communication impairment and restricted and repetitive behaviors or interests. As of 2020, the US Centers for Disease Control and Prevention projected that the prevalence of ASD stood at 1 in 36 individuals. This figure is on the rise, potentially attributed to heightened awareness within the general public, medical practitioners, and research circles, according to researchers.
The study involved retinal photographs of 1890 eyes of 958 participants under 19. The participants were selected from the Department of Child and Adolescent Psychiatry, Severance Hospital, Yonsei University College of Medicine in Korea, between April and October 2022. Also, the team gathered retinal photographs from control participants with typical development (TD), matching their age and sex. This retrospective collection occurred at the hospital’s Department of Ophthalmology from December 2007 to February 2023.
Using the data, the team created a convolutional neural network, a deep learning algorithm, to train models for ASD screening and assessing symptom severity. This training involved 85 percent of the retinal images and corresponding scores from symptom severity tests. Subsequently, the remaining 15 percent of images were set aside specifically for testing purposes.
The severity of ASD symptoms was gauged through calibrated severity scores from the Autism Diagnostic Observation Schedule – Second Edition (ADOS-2) and the Social Responsiveness Scale – Second Edition (SRS-2).
Encouraging results
When evaluating the test set of images for ASD screening, the AI model demonstrated the ability to accurately identify children with an ASD diagnosis, yielding a mean area under the receiver operating characteristic (AUROC) curve of 1.00. AUROC values range from 0 to 1, with 0 indicating a model with predictions entirely incorrect and 1 indicating a model with predictions 100 percent correct.
In the current study, the AI’s predictions achieved 100 percent accuracy, and even after removing 95 percent of the least important areas of the image (excluding the optic disc), there was no significant decline in the mean AUROC, according to the team.
“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,” said the study.
The researchers noted that their findings indicate potential additional information about symptom severity from retinal photographs. It was observed that feasible classification was attainable only for ADOS-2 scores and not SRS-2 scores. This distinction may arise from the fact that ADOS-2 assessments are conducted by trained professionals with sufficient time for evaluation, offering a more accurate reflection of severity status, whereas SRS-2 assessments are typically completed by caregivers in a shorter time frame, potentially resulting in a less precise assessment of severity, according to the team.
While further research is needed to establish generalizability, researchers note that their study marks a significant advancement in creating objective screening tools for ASD. These tools could alleviate pressing concerns, such as the limited accessibility of specialized child psychiatry assessments due to resource constraints.
Source: Interesting Engineering
A Dentist Reveals What Happens to Your Teeth as You Get Older
AI-model diagnoses childhood autism from eye pics with 100% accuracy