Abstract
Purpose :
Amblyopia damages both visual sensory and ocular motor functions. One manifestation of the damage is abnormal fixation eye movements (FM)(Figure 1). Advances in eye-tracking have enabled an accurate and objective assessment of fixation in pediatric patients. Our study aims to integrate eye tracking technology with high-accuracy AI algorithms for detection of amblyopia.
Methods :
High-resolution video-oculography was used to record FMs from 46 controls and 160 amblyopia subjects under binocular viewing (BV) and under right vs left eye viewing (MV, one eye blocked) conditions. The fixation and vergence stability were quantified using Bivariate Contour Ellipse Area (BCEA). The fast and slow FMs were analyzed using Engbert algorithms. The data was split 70:30 for training and testing ML and AI algorithms. We used fixation and vergence instability, fast and slow FM metrics to train ML models namely Random Forest, Decision Tree, Gradient Boost, AdaBoost classifiers to detect amblyopia. We also used two deep learning algorithms, namely Long Short-Term Memory (LSTM) and transformer in which the inputs were calibrated eye positions.
Results :
The ML models detected the presence of amblyopia with an accuracy ranging 81-88% and ROC ranging 0.85-0.96. The LSTM and Transformer models predicted the severity of amblyopia with a ROC-AUC ranging from 0.71-0.81 for mild/treated and 0.82-0.89 for moderate/severe amblyopia subjects.
Conclusions :
We have developed ML/AI tools using FM biomarkers that can reliably detect the presence and severity of amblyopia. Future work would comprise of developing AI algorithms that combine unique selected FM features with sequences of actual eye position data to optimize model performances.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.