Summary of the AI studies for the diagnosis of glaucoma using the structural evaluations
| Author | Journal | Population (training/testing) | Assessment | Type of AI | Data type | AUC | Sensitivity | Specificity | Comments |
| Bowd et al 45 | Invest Ophthalmol Vis Sci | 189 N and 108 G patients (10-fold cross-validation) | Diagnosis (S) | MLC | CSLO parameters | 0.945 | 83% | 90% | – |
| Townsend et al 46 | Br J Ophthalmol | 60 N and 140 G patients (eightfold cross-validation) | Diagnosis (S) | MLC | CSLO parameters | 0.904 | 85.7% | 80% | – |
| Zangwill et al 47 | Invest Ophthalmol Vis Sci | 135 N and 95 G patients (10-fold cross-validation) | Diagnosis (S) | MLC | CSLO parameters | 0.964 | 85% | 90% | – |
| Uchida et al 48 | Invest Ophthalmol Vis Sci | 43 N and 53 G patients (threefold cross-validation) | Diagnosis (S) | MLC | CSLO parameters | 0.940 | 92% | 91% | – |
| Adler et al 49 | Methods Inf Med | 98 N and 98 G patients (threefold cross-validation) | Diagnosis (S) | MLC | CSLO parameters | – | 82.79% | 86.54% | Error=15.4% |
| Bowd et al * 50 | Invest Ophthalmol Vis Sci | 226 GS patients | Diagnosis (S) | MLC | CSLO parameters | – | – | – | HRs (95% CI): 1.57 (1.03 to 2.59) for SVM forward; and 1.70 (1.18 to 2.51) for SVM back |
| Weinreb et al 51 | Arch Ophthalmol | 84 N and 83 G patients (12-fold cross-validation) | Diagnosis (S) | DA | SLP parameters | 0.890 | 74% | 92% | – |
| Bowd et al 52 | Invest Ophthalmol Vis Sci | 72 N and 92 G patients (10-fold cross-validation) | Diagnosis (S) | MLC | SLP parameters | 0.940 | 77% | 90% | – |
| Li et al 60 | Ophthalmology | Training: 23 433 N, 6122 G Testing: 6033 N, 1537 G | Diagnosis (S) | DL | FP | 0.9825 | 95.60% | 92% | – |
| Ting et al 61 | JAMA | Training: 494 661 images (125 189 G) Testing: 71 896 G | Diagnosis (S) | DL | FP | 0.942 | 96.40% | 87.2% | – |
| Medeiros et al*59 | Ophthalmology | Training: 476 N, 674 GS and 699 G patients (GS) Testing: 128 N,164 GS and 171 G patients | Diagnosis (S) | DL | FP | 0.944 (95% CI 0.902 to 0.966) | 90% | 80% | Accuracy: 83.70% |
| Huang and Chen64 | Invest Ophthalmol Vis Sci | 100 N and 89 G patients (10-fold cross-validation) | Diagnosis (S) | MLC | OCT parameters | 0.821–0.991 | 72%–100% | 80% | – |
| Burgansky-Eliash et al 65 | Invest Ophthalmol Vis Sci | 42 N and 47 G patients (sixfold cross-validation) | Diagnosis (S) | MLC | OCT parameters | 0.981 | 92.50% | 95% | – |
| An et al 66 | J Glaucoma | 105 G patients (10-fold cross validation) | Diagnosis (S) | MLC | OCT parameters | – | – | – | Accuracy: 87.8% |
| Kim et al 67 | PLoS One | 202 N and 292 G (training and testing split: 80:20%) | Diagnosis (S) | MLC | OCT parameters and VFs | 0.979 | 98.3 % | 97.5% | Accuracy: 98% |
| Barella et al 68 | J Ophthalmol | 46 N and 57 G patients (10-fold cross-validation) | Diagnosis (S) | MLC | OCT parameters | 0.877 | 64.9% | 80% | – |
| Christopher et al 69 | Invest Ophthalmol Vis Sci | 28 N and 93 G patients (training and testing: LOO) | Diagnosis (S) | UL | OCT parameters | 0.95 | – | – | – |
| Muhammad et al 73 | J Glaucoma | 45 N and 57 G patients | Diagnosis (S) | DL | OCT images | – | – | – | Accuracy: 93.1% |
| Girard et al 74 | ARVO General Meeting | 2566 N and 135 G (training and testing split: 70:30%) | Diagnosis (S) | DL | OCT Images | 0.90 | – | – | – |
| Maetschke et al 75 | PloS One | Training: 216 N and 672 G scans Testing: 17 N and 93 G scans | Diagnosis (S) | DL | OCT images | 0.94 | – | – | – |
| Zhang et al 77 | ARVO General Meeting | 55 902 N and 39 096 G images (training, validation and testing split: 70:15:15%) | Diagnosis (S) | DL | OCT images | 0.90 | – | – | – |
| Xu78 | IEEE EMBC | 2048 G images (training and testing: LOO) | Diagnosis (S) | DL | AS-OCT images | 0.921 | – | 85% | |
| Fu et al 79 | Am J Ophthalmol | 2113 G (fivefold cross-validation) | Diagnosis (S) | DL | AS-OCT images | 0.960 | 90% | 92% | – |
| Fu et al 80 | IEEE Trans Cybern | Data set 1: 2113 G patients Data set 2: 202 G patients | Diagnosis (S) | DL | AS-OCT images | Data set 1: 0.962 Data set 2: 0.952 | Data set 1: 93% Data set 2: 87.4% | Data set 1: 90% Data set 2: 95% |
Balanced accuracy
Data set 1: 91.8% Data set 2: 91.24% F-measure Data set 1: 67.7% Data set 2: 81.8% |
| Niwas et al 81 | Comput Methods Programs Biomed/ | 74 G patients (training and testing: LOO and 10-fold cross-validation) | Diagnosis (S) | MLC | AS-OCT images | 0.95 (LOO method) | 88.90% (LOO method) | 93.33% (LOO method) |
Accuracy
LOO method: 89.20% 10-fold cross-validation: 85.12% |
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; AS-OCT, anterior segment optical coherence tomography; AUC, area under the curve; CSLO, confocal scanning laser ophthalmoscopy; DA, discriminant analysis; DL, deep learning; FP, fundus photograph; G, glaucoma; GS, glaucoma suspect; LOO, leave one out validation; MLC, machine learning classifier; N, normal; OCT, optical coherence tomography; S, evaluation using structural information; SLP, scanning laser polarimetry; SVM, support vector machine; UL, unsupervised learning; VF, visual field.