2023-12-23 00:05:49
WASHINGTON, December 22, 2023 (APMnews) – The analysis of retinal photographs of children using artificial intelligence (AI) appears to constitute a tool for objective screening of autistic disorders and possibly for assessing their severity, according to a Korean study published in JAMA Network Open.
There are already various tools for identifying children potentially suffering from autistic disorders, with rather good performance, but the growing demand for diagnosis and evaluation faces limited resources in trained professionals. New screening methods are needed, say Jae Han Kim of Yonsei University School of Medicine in Seoul and colleagues.
They were interested in retinal photographs which are increasingly used to indirectly assess structural abnormalities of the brain. Studies have shown alterations visible on retinography in children with autism spectrum disorders (ASD) compared to those with typical development, and machine learning models from these types of images have suggested that they may used to screen for ASD but in a small study.
Korean researchers wanted to confirm the interest of AI applied to retinograms as a tool for screening ASD but also for assessing their severity in a large cohort.
This is a single-center study carried out among 958 children and adolescents (7.8 years on average, 81.8% boys), 479 with ASD whose retinograms were collected prospectively and 479 with developmental typical matched on age and sex whose data were collected retrospectively, i.e. a total of 1,890 retinograms.
The researchers applied a so-called “deep set” approach to image analysis, i.e. an aggregation of several convolutional neural networks in order to quantify uncertainty in machine learning models and obtain better performance. .
It appears that to differentiate children with ASD and those without according to the DSM-5 criteria, the model developed presents an average area under the curve (AUROC) of 100%, with a sensitivity and specificity of 100%.
To evaluate the performance of the model in assessing the severity of disorders on the ADOS-2 score, 305 retinograms were used and it was possible to distinguish severe cases from mild to moderate cases with an AUROC of 74%, a sensitivity of 58 % and a specificity of 74% as well as a precision of 66%.
On the other hand, the model applied to 556 retinograms did not make it possible to evaluate the severity of the disorders according to the SRS-2 score (AUROC of 0.44).
The researchers then applied the model to 962 retinograms of patients with fundus pathologies to estimate its uncertainty by calculating entropy from 0 to 1, with a larger value indicating greater uncertainty.
The model had an entropy of 0.8 for spotting ASD cases in the first group of images and only 0.01 in the second.
These results suggest that AI analysis of the optic disc on retinograms, the part of the retina from which the optic nerve emerges, appears to be a promising objective tool for screening ASD and potentially assessing its severity, conclude Researchers.
Larger studies must now be carried out to consider a generalization of such a tool which would help resolve the problems of access to specialists to obtain a diagnosis, they add.
(JAMA Network Open, online publication of December 15)
ld/ab/APMnews
1703290972
#Emergency #News #APM #French #Society #Emergency #Medicine