Artificial intelligence in glaucoma progression: A systematic review

Abstract

Aims/Purpose: Glaucoma is the leading cause of irreversible blindness worldwide. Effective assessment of glaucoma progression could facilitate timely interventions and improve prognosis. This systematic review aims to evaluate the use of machine learning (ML) in the detection and prediction of longitudinal glaucoma progression.Methods: An online search was conducted on 31.12.2022, and relevant studies were identified. Outcomes included ML maximum sensitivity, specificity and area under the curve (MSSAUC) (mean (range)).Results: Of the 577 records initially identified, 40 studies with over 156 000 eyes were included. Overall, 30 studies included traditional ML methods, and 13 studies included representational methods. 28 studies used supervised ML, whilst 12 studies used unsupervised ML. 27 studies performed classification, and 16 studies included regression analysis. 11 studies evaluated structural glaucoma data, such as retinal nerve fibre layer thickness. Six studies detected glaucoma progression, with pooled MSSAUC of 68.9% (36.9–86.0), 80.5% (49.0–94.0) and 0.80 (0.60–0.91). Five studies predicted glaucoma progression, with pooled MSSAUC of 84.9% (81.0–93.0), 73.5% (59.0–91.7) and 0.82 (0.71–0.90). 23 studies evaluated functional visual field data. 10 studies detected glaucoma progression, with pooled MSSAUC of 74.8% (3.8–95.0), 94.3% (90.0–98.1) and 0.89 (0.82–0.94). Four studies predicted glaucoma progression, with pooled MSSAUC of 52.2% (20.0–72.0), 72.3% (49.0–95.0) and 0.73 (0.44–0.86). Eight studies combined structural and functional data. Two studies detected glaucoma progression and six studies predicted glaucoma progression, with pooled MSSAUC of 66.6% (44.2–83.7), 90.0% (85.0–100%) and 0.79 (0.61–0.88).Conclusions: ML detection and prediction of longitudinal glaucoma progression is highly accurate. However, there is significant heterogeneity in ML methodologies. Standardization is needed for robust evidence of the clinical utility of ML.