AI and eye tracking: Technological companions to assess amblyopia in pediatric patients

(Image Credit: AdobeStock/Maxim Kukurund)

Artificial intelligence (AI) tools have a promising future for reliably detecting the presence and severity of amblyopia in pediatric patients.

During the Fourth Annual Cleveland Eye Bank Foundation Virtual Vision Research Symposium,¹ Fatema Ghasia, MD, discussed the different visual function defects in amblyopia, focusing on ocular motor deficits and how they can be used to create AI algorithms to detect amblyopia. She is an associate professor at Cleveland Clinic Cole Eye Institute and director of the Vision Neuroscience and Ocular Motility Laboratory at Cleveland Clinic in Ohio.

“Amblyopia, a visual impairment caused by abnormal vision experienced during childhood, affects about 3% to 5% of the population and typically develops during the first 7 years of life,” she said.

Amblyopia results from unequal visual input from the 2 eyes of a patient. Causes include anisometropia, in which there is a difference in the refractive errors between the eyes; strabismus, causing ocular misalignment; deprivation from cataract or ptosis; or mixed mechanisms.

Amblyopia had been considered clinically as a monocular problem in an eye with a vision deficit that is not resolved by spectacle correction alone. However, it has been recognized that amblyopia affects visual functions in natural viewing, the most common being depth perception, reading difficulties,² contrast sensitivity,³ visual acuity,³ and fixation instability.⁴

“Amblyopia often has a lifelong impact on children’s academic and physical abilities, with adverse psychosocial effects,” Ghasia commented.

Detecting amblyopia

Ghasia pointed out that children with amblyopia without strabismus or with mild strabismus are not recognized until the age of 5, and the duration of the disrupted binocular visual input affects visual function deficits. The ability to detect these children earlier results in less visual impairment.⁵ However, screening in schools is less than adequate because of the lack of skilled screeners and reliable responses from the children.

Photoscreening can detect amblyopia risk factors, but the technology is limited because low specificity results in over-referrals.

The Blinq pediatric vision scanner (Rebion) detects foveal fixation in both eyes simultaneously and can detect actual amblyopia, which requires specialized equipment that may not be readily available.

Next steps

Ghasia and colleagues sought to answer whether eye movement abnormalities can aid clinicians in detecting amblyopia.

Eye tracking is advantageous because it provides objective data without relying on patient responses. Recent advances have produced more affordable, portable eye-tracking systems that provide enhanced data accuracy, and reliable data can be obtained in young children and those with developmental delays, Ghasia explained.

A downside is that specialized expertise is required to analyze and interpret eye movement data. In light of this, her research team investigated if deep learning models that use fixation eye positions can be developed to detect amblyopia.

The study included 40 controls and 110 patients with amblyopia who ranged in age from 3 to 65 years. For patients with amblyopia, investigators broke down the cohort as follows: 34% with anisometropia, 30% with strabismus, and 36% with mixed amblyopia; 60% had moderate and severe amblyopia and 40% mild disease.

High-precision video-oculography (EyeLink 1000 Plus; SR Research Ltd) recorded binocular horizontal and vertical eye positions over 45 seconds during simple visual fixation tasks while viewing a target using primary gaze during binocular viewing, fellow eye viewing, and amblyopic eye viewing, Ghasia recounted.

The study cohort was split 80:20 to train and test the model. The results showed that the controls had stable fixation, but there was increased instability with monocular viewing, particularly in the occluded eye. In contrast, the patients with amblyopia showed more unstable fixation in the fellow and amblyopic eyes during binocular and monocular viewing.

Amblyopia detection using AI models

Amblyopia type

Ghasia outlined the results obtained with the AI models for detecting amblyopia based on fixation eye positions for controls vs patients. For those with anisometropic amblyopia, investigators observed the highest area under the curve (AUC) of 0.74, and they used fixation data obtained in monocular viewing.

For patients with strabismic amblyopia, the accuracies exceeded the results for anisometric amblyopia, with the highest accuracy being 0.79 for data obtained in binocular viewing.

Using binocular viewing data, investigators observed the highest AUC value of 0.87 for patients with mixed amblyopia.

Amblyopia severity

Ghasia reported that the model performed well, with accuracies ranging from 0.79 to 0.87; the greatest accuracy was seen with binocular viewing data. For moderate to severe amblyopia, accuracies ranged up to 0.92 for the binocular viewing data.

Considering the presence or absence of nystagmus, the investigators wanted to assess how the model performed. In amblyopia without nystagmus, the model performed reasonably well, according to Ghasia. The greatest accuracy of 0.78 was associated with binocular viewing data.

In patients with nystagmus, the accuracy was higher, with the highest being 0.87, as seen in binocular viewing.

Ghasia noted that deep learning models using fixation eye position data can detect the presence of amblyopia.

“The models demonstrate good performance in detecting various types and severity of amblyopia,” Ghasia concluded. “Models effectively detect amblyopic [patients] with and without nystagmus, as fixation is seen in both groups. Our pilot study with the small cohort underscores the need for future research with larger sample sizes.”

Fatema Ghasia, MD

E: ghasiaf@ccf.org

Ghasia is an associate professor at Cleveland Clinic Cole Eye Institute and director of the Vision Neuroscience and Ocular Motility Laboratory at Cleveland Clinic in Ohio. She has no financial interest in this subject matter.

References

1. Ghasia F. Beyond the eye chart: unveiling visual sensory and oculomotor deficits in amblyopia. Presented at: Fourth Annual Cleveland Eye Bank Foundation Virtual Vision Research Symposium; October 15, 2024; Cleveland, Ohio.

2. Bhutada I, Skelly P, Jacobs J, Murray J, Shaikh AG, Ghasia FF. Reading difficulties in amblyopia: consequence of visual sensory and oculomotor dysfunction. J Neurol Sci. 2022;442:120438. doi:10.1016/j.jns.2022.120438

3. Dulaney CS, Murray J, Ghasia F. Contrast sensitivity, optotype acuity and fixation eye movement abnormalities in amblyopia under binocular viewing. J Neurol Sci. 2023;451:120721. doi:10.1016/j.jns.2023.120721

4. Kang SL, Beylergil SB, Otero-Millan J, Shaikh AG, Ghasia FF. Fixational eye movement waveforms in amblyopia: characteristics of fast and slow eye movements. J Eye Mov Res. 2019;12(6):10.16910/jemr.12.6.9. doi:10.16910/jemr.12.6.9

5. Williams C, Northstone K, Harrad RA, Sparrow JM, Harvey I; ALSPAC Study Team. Amblyopia treatment outcomes after screening before or at age 3 years: follow up from randomised trial. BMJ. 2002;324(7353):1549. doi:10.1136/bmj.324.7353.1549