Assessing the external validity of machine learning-based detection of glaucoma

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Machine learning (ML) models demonstrate high accuracy in detecting glaucoma from optical coherence tomography (OCT) scans.
• Previous studies have not evaluated the performance of these ML models across different ethnic populations.

Purpose of the Study:

• To externally validate ML models for glaucoma detection using OCT data.
• To assess the generalizability of ML models across Asian and Caucasian ethnicities.

Main Methods:

• Prospective, cross-sectional study involving 514 Asians and 138 Caucasians.
• Developed two ML classifiers using retinal nerve fibre layer (RNFL) thickness: one with original OCT data and another with data corrected by a compensation model.
• Validated models on independent Asian and Caucasian datasets.

Main Results:

• ML models significantly outperformed measured data in Asian and Caucasian datasets.
• The ML model using compensated RNFL data showed consistent performance across ethnicities (Asian AUC=0.96, Caucasian AUC=0.93).
• The ML model using original OCT data exhibited poor reproducibility across ethnic groups (Asian AUC=0.92, Caucasian AUC=0.83).

Conclusions:

• ML models, particularly those using compensated RNFL data, show promise for glaucoma detection.
• Model performance can vary significantly across different ethnicities.
• Caution is advised when applying ML models to diverse patient cohorts due to potential ethnic variations.