A Multihead Attention Deep Learning Algorithm to Detect Amblyopia Using Fixation Eye Movements

To develop an attention-based deep learning (DL) model based on eye movements acquired during a simple visual fixation task to detect amblyopic subjects across different types and severity from controls. An observational study. We recruited 40 controls and 95 amblyopic subjects (anisometropic = 32;...

Full description

Saved in:
Bibliographic Details
Published in:Ophthalmology science (Online) Vol. 5; no. 5; p. 100775
Main Authors: Upadhyaya, Dipak P., Cakir, Gokce, Ramat, Stefano, Albert, Jeffrey, Shaikh, Aasef, Sahoo, Satya S., Ghasia, Fatema
Format: Journal Article
Language:English
Published: Netherlands Elsevier Inc 01.09.2025
Elsevier
Subjects:
Amblyopia
Artificial intelligence
Eye tracking
Fixation eye movements
Ophthalmology
Original
Strabismus
Strabismus
D
FMN
DL
MV
Fixation eye movements
FEM
AEV
BV
AUPRC
FEV
Amblyopia
Artificial intelligence
Eye tracking
fixation eye movement
amblyopic eye viewing
binocular viewing
fusion maldevelopment nystagmus
deep learning
area under the precision–recall curve
diopters
fellow eye viewing
monocular viewin
ISSN:2666-9145, 2666-9145
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:To develop an attention-based deep learning (DL) model based on eye movements acquired during a simple visual fixation task to detect amblyopic subjects across different types and severity from controls. An observational study. We recruited 40 controls and 95 amblyopic subjects (anisometropic = 32; strabismic = 29; and mixed = 34) at the Cleveland Clinic from 2020 to 2024. Binocular horizontal and vertical eye positions were recorded using infrared video-oculography during binocular and monocular viewing. Amblyopic subjects were classified as those without nystagmus (n = 42) and those with nystagmus with fusion maldevelopment nystagmus (FMN) or nystagmus that did not meet the criteria of FMN or infantile nystagmus syndrome (n = 53). A multihead attention-based transformer encoder model was trained and cross-validated on deblinked and denoised eye position data acquired during fixation. Detection of amblyopia across types (anisometropia, strabismus, or mixed) and severity (treated, mild, moderate, or severe) and subjects with and without nystagmus was evaluated with area under the receiver-operator characteristic curves, area under the precision–recall curve (AUPRC), and accuracy. Area under the receiver-operator characteristic curves for classification of subjects per type were 0.70 ± 0.16 for anisometropia (AUPRC: 0.72 ± 0.08), 0.78 ± 0.15 for strabismus (AUPRC: 0.81 ± 0.16), and 0.80 ± 0.13 for mixed (AUPRC: 0.82 ± 0.15). Area under the receiver-operator characteristic curves for classification of amblyopia subjects per severity were 0.77 ± 0.12 for treated/mild (AUPRC: 0.76 ± 0.18), and 0.78 ± 0.09 for moderate/severe (AUPRC: 0.79 ± 0.16). Th area under the receiver-operator characteristic curve for classification of subjects with nystagmus was 0.83 ± 0.11 (AUPRC: 0.81 ± 0.18), and the area under the receiver-operator characteristic curve for those without nystagmus was 0.75 ± 0.15 (AUPRC: 0.76 ± 0.09). The multihead transformer DL model classified amblyopia subjects regardless of the type, severity, and presence of nystagmus. The model's ability to identify amblyopia using eye movements alone demonstrates the feasibility of using eye-tracking data in clinical settings to perform objective classifications and complement traditional amblyopia evaluations. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:2666-9145
2666-9145
DOI:10.1016/j.xops.2025.100775