J Korean Ophthalmol Soc.  2024 Oct;65(10):675-682. 10.3341/jkos.2024.65.10.675.

Strabismus Screening Using Eyetracker Combined with Machine Learning

• Son SM
• Kim JH
• Moon S

• KMID: 2559995
• DOI: http://doi.org/10.3341/jkos.2024.65.10.675

Abstract

Purpose
To assess the effectiveness of an automated screening program that diagnoses horizontal strabismus using machine learning based on ocular deviation data captured by the wearable eyetracker, Tobii pro glasses 2 (TPG2).
Methods
The TPG2 which locates the pupil center to measure ocular movement was used. In normal adults wearing TPG2, horizontal ocular deviation was induced by covering the left eye and applying prisms of varying strengths (2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 25, 30, 35, and 40 PD base-in and out) to the right eye. TPG2 automatically recorded ocular deviation before and after prism induction generating 28 types of ocular deviation sets. From each set, 20 X-axis values before and after ocular deviation were randomly extracted using an oversampling technique creating a total of 61,600 ocular deviation sets. For training, 56,000 sets were used and 5,600 were evaluated for sensitivity, specificity, and area under the curve (AUC).
Results
Eleven normal adults (5 males) participated with a mean age of 34.8 ± 7.37 years. Based on an 8 PD threshold, deviations of 8 PD or less demonstrated a sensitivity of 1.0, a specificity of 0.95, and an AUC of 0.97. When categorized into three groups based on 8 PD and 20 PD thresholds, the results were: sensitivity of 0.90 and specificity of 0.95 for ≤ 8 PD; sensitivity of 0.60 and specificity of 1.00 for 8-20 PD; sensitivity of 1.00 and specificity of 0.88 for > 20 PD.
Conclusions
The machine learning program developed using induced ocular deviations measured with prisms and TPG2 shows promise for use in future strabismus screening tests.

Keyword

Exotropia, Machine learning, Strabismus, Wearable eye tracker