A recent study published in JAMA Network Open reported on the efficacy of a newly developed deep-learning-based artificial intelligence (AI) system to screen children for autism spectrum disorder (ASD) using retinal photographs.
Let’s talk about ASD first.
ASD is a neurodevelopmental disorder characterized by symptoms related to social communication impairment and restricted and repetitive behaviors or interests.
From a visual standpoint, individuals with ASD have structural retinal changes that potentially reflect brain alterations, such as visual pathway abnormalities through embryonic and anatomic connections.
Give me some background on this research.
A previous study demonstrated that machine learning models developed to screen for ASD using retinal photographs had a sensitivity and specificity of 95.7% and 91.3%, respectively.
As such, a research team from South Korea sought to validate and build on these findings using deep learning AI models on a larger number of participants.
Now this study.
In this single-center, diagnostic study, investigators prospectively collected retinal photographs of individuals with ASD and retrospectively collected those of age- and sex-matched individuals with typical developments (TD).
In addition to evaluating whether the deep learning models could differentiate between individuals with ASD vs. TD, investigators also assessed if they could discern between individuals with severe or mild to moderate ASD.
What were the main outcomes?
- Participant-level area under the receiver operating characteristic curve (AUROC)
- Sensitivity
- Specificity
To note, AUROC is a metric for measuring a model’s ability to discriminate between cases (i.e., participants with ASD) and non-cases (TD participants).
The closer to 1.00 the AUROC is, the better the model is at discerning between cases and non-cases.
Talk about the cohorts.
A total of 1,890 eyes of 958 participants were included in the study, and the ASD and TD cohorts were comprised of 479 participants (945 eyes).
Both ASD and TD groups had a mean (standard deviation [SD]) age of 7.8 (3.2) years and were mostly males (392 [81.8%]).
Findings?
For ASD screening, the deep learning models had a mean AUROC, sensitivity, and specificity of 1.00 (95% CI 1.00-1.00) on the test set—meaning the system found all participants with ASD, with no false positives.
They maintained a mean AUROC of 1.00 using only 10% of the image containing the optic disc, indicating that this area was critical for discerning ASD from TD.
Tell me more.
In terms of symptoms severity screening, the models produced a:
- Mean AUROC of 0.74 (95% CI 0.67-0.80)
- Sensitivity of 0.58 (95% CI 0.49-0.66)
- Specificity of 0.74 (95% CI 0.67-0.82)
These findings indicated that the system was less effective at detecting the severity of ASD in participants.
Expert opinion?
Per the study authors, “Our sequential age-based modeling suggested that retinal photographs may serve as an objective screening tool starting at least at age 4 years. Moreover, the newborn retina continues to develop and mature up to age 4 years."
They added that, “Taken together, our models are potentially viable for screening children from this age onward, which is earlier than the average age of 60.48 months at ASD diagnosis.”
Take home.
The study findings suggest that deep learning algorithms using retinal photographs could be a potential objective screening tool for ASD and possibly for symptom severity.
While future multicenter studies are required to establish generalizability, using retinal photographs may expedite the ASD screening process—providing improved accessibility to specialized pediatric psychiatry assessments.