Author | Journal | Population (training/testing) | Assessment | Type of AI | Data type | AUC | Sensitivity | Specificity | Comments |
Goldbaum et al 28 29 | Invest Ophthalmol Vis Sci | 60 N and 60 G patients (10-fold cross-validation) | Diagnosis (F) | ANN | VFs | – | 65% | 71% | ANN/expert agreement: 74% |
Goldbaum et al 30 | Invest Ophthalmol Vis Sci | 189 N and 156 G patients (10-fold cross-validation) | Diagnosis (F) | MLC | VFs | 0.922 | 79% | 90% | – |
Chan et al 31 | IEEE Trans Biomed Eng | 189 N and 156 G patients (25-fold cross-validation) | Diagnosis (F) | MLC | VFs | 0.88–0.92 | 58.3%–78.2% | 90% | – |
Bizios et al 32 | J Glaucoma | 116 N and 100 G patients Training: 53.7:46.3% (N:G) Testing: 46.3:53.7% (N:G) | Diagnosis (F) | ANN | VFs | 0.984 | 93% | 94% | – |
Lietman et al 33 | J Glaucoma | 249 N and 106 G patients (training and testing: half-half) | Diagnosis (F) | ANN | VFs | – | – | – | Outperformed global indices at high specificities (90%–95%) |
Li e t a l 34 | BMC Med Imaging | 4012 PD images (training: 3712; testing: 300) | Diagnosis (F) | DL | VFs (PD images) | – | 93.20% | 82.60% | Accuracy: 87.60% |
Kucur et al 35 | PloS O ne | 1979 N and 2811 G VF images (10-fold cross-validation) | Diagnosis (F) | DL | VFs (Voronoi images) | – | – | – | Average precision: 87.40% |
Brigatti et al 37 | Arch Ophthalmol | 181 G patients (threefold cross-validation) | Prognosis (F) | ANN | VFs | – | 73% | 88% | – |
Sample et al 38 | Invest Ophthalmol Vis Sci | Training: 189 N and 156 G patients (10-fold cross-validation) Tested: 114 OHT | Prognosis (F) | MLC | VFs | – | – | – | Predicted abnormal fields 3.92±0.55 years earlier than traditional methods |
Sample et al*39 | Invest Ophthalmol Vis Sci | 66 OHT, 73 GS and 52 G patients | Prognosis (F) | UL | VFs | – | – | – | ICA quantitatively identified a larger percentage of progression in eyes. |
Yousefi et al 40 | Am J Ophthalmol | Cross-sectional cohort: 1146 N and 677 patients Longitudinal cohort: 139 N patients Test–retest cohort: 71 G patients | Prognosis (F) | UL | VFs | – | 87% | 96% | Machine learning helps to consistently detect the progression of glaucoma much earlier than other conventional methods. |
Wen et al 41 | PloS O ne | Training: 3972 patients Testing: 903 patients | Prognosis (F) | DL | 24-2 HVFs | – | – | – | Prediction of HVFs up to 5.5 years from single HVF |
The abbreviations and acronyms used in the table can be found in the online supplementary appendix.
*Study considered glaucoma suspect/early-stage glaucoma subjects.
AI, artificial intelligence; ANN, artificial neural network; AUC, area under the curve; DL, deep learning; F, evaluation using functional information; G, glaucoma; GS, glaucoma suspect; HVF, Humphrey visual fields; ICA, independent component analysis; MLC, machine learning classifier; N, normal; OHT, ocular hypertension; PD, pattern deviation; UL, unsupervised learning; VF, visual field.