Abstract
Purpose :
To validate a machine learning model from spectral-domain optical coherence tomography (Cirrus, Carl Zeiss Meditec, Dublin, CA) data along with patients’ background information for glaucoma detection and assess its external validity.
Methods :
In this cross-sectional study, 514 Asians (257 glaucoma and 257 controls) were enrolled to construct a machine learning model for glaucoma detection, which was then tested on an independent dataset of 356 Asians (183 glaucoma and 173 controls) and an external dataset of 138 Caucasians (57 glaucoma and 81 controls). The machine learning model was also compared to the existing multivariate-adjusted retinal nerve fiber layer (RNFL) thickness compensation model and Cirrus-generated (measured) RNFL model. The area under the receiver operating characteristics curve (AUC) for glaucoma detection was calculated.
Results :
The machine learning-based model (AUC=0.97) outperformed both the compensation model (AUC=0.93; P<0.001) and the measured RNFL model (AUC=0.93; P<0.001) for glaucoma detection in the Asian dataset. However, in the Caucasian dataset, the compensation model (AUC=0.90 to 0.93) outperformed both machine learning-based (AUC=0.85; P<0.001) and measured RNFL (AUC=0.82; P<0.001) models, and there was no significant difference in the AUC between the machine learning-based model and measured RNFL model (P=0.174) for glaucoma detection.
Conclusions :
While the machine learning model detected glaucoma at a higher accuracy as compared to the compensation model or measured RNFL model in the internal dataset, this finding was not fully replicated in the external validation with the Caucasian dataset. Care must be taken when machine learning models are applied to patient cohorts of different ethnicities.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.