An objective and easy-to-use glaucoma severity classification system based on artificial intelligence

Abstract

Purpose : To develop a simple and easy-to-use glaucoma staging system based on visual fields (VFs) and to evaluate the system using an independent validation dataset.

Methods : We developed an unsupervised k-means algorithm to identify clusters with similar VFs (Fig. 1). We annotated the clusters based on their respective mean deviation (MD). To establish an objective criterion for glaucoma staging, we computed optimal MD thresholds that discriminated clusters with the highest accuracy based on Bayes minimum error principle. We validated the entire pipeline based on an independent validation dataset and evaluated the accuracy of the staging system based on the identified MD thresholds (Fig. 1).

Results : k-means model discovered four clusters with 6784, 4034, 1541, and 872 VFs and average MDs of 0.0 dB (±1.4: SD), -4.8 dB (±1.9), -12.2 dB (±2.9), and -23.0 dB (±3.8), respectively. The supervised Bayes minimum error classifier identified optimal MD thresholds of -2.2 dB, - 8.0 dB, and -17.3 dB for discriminating normal eyes and eyes at the early, moderate, and advanced stages of glaucoma (Fig. 2). Approximately 2.2% of normal eyes and eyes in the early stage of glaucoma had VF loss in at least one of the four central VT test locations. The accuracy of the glaucoma staging system based on the identified MD thresholds with respect to the initial k-means clusters was about 94%.

Conclusions : We used unsupervised and supervised machine learning models to develop an objective glaucoma staging system without expert intervention. We discovered that MDs of approximately -2 dB, - 8 dB, and -17 dB provide the optimal thresholds for staging glaucoma to four severity levels. The proposed staging system is unbiased, easy-to-use, and consistent, thus may be used in glaucoma research and clinical practice to improve glaucoma care.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Pipeline to develop the glaucoma staging system. Unsupervised model discovered four clusters that were labeled based on their respective average mean deviation (MD). Bayes minimum error classifier identified optimal MD thresholds to stage glaucoma to four severity levels.

 

Figure 2. Mean deviation (MD) of VFs in the discovered clusters. Left: Distribution of MD across four discovered clusters. Right: Optimal MD thresholds of the glaucoma staging system were identified based on the Bayes minimum error-rate principle.