Skip to main content

Table 5 Summary of studies using machine learning classifier for different KC or subclinical KC eyes from normal eyes

From: Machine learning helps improve diagnostic ability of subclinical keratoconus using Scheimpflug and OCT imaging modalities

Authors Year Instruments ML classifier Subjects Results
Current Study 2019 UHR-OCT, Scheimpflug camera Neural network 38 eyes with KC, 33 eyes with subclinical KC, 50 normal eyes 93% precision for subclinical KC eyes, 99% precision for KC eyes
Smolek et al. [9] 1997 Corneal topography Neural network 6 KC suspect eyes, 33 eyes with KC 100% accuracy, sensitivity and specificity for all KC suspect and KC eyes
Accardo et al. [24] 2002 Corneal topography Neural network 120 eyes with early KC eyes, 120 normal eyes 94.1% sensitivity, 97.6% specificity for early KC eyes
Arbelaez et al. [11] 2012 Scheimpflug camera and Placido corneal topography SVM 877 eyes with KC, 426 eyes with subclinical KC, 1259 healthy control eyes 98.2% accuracy (95.0% sensitivity and 99.3% specificity) for KC eyes and 97.3% accuracy (92.0% sensitivity and 97.7% specificity) for subclinical KC eyes
Smadja et al. [10] 2013 Scheimpflug camera Decision tree 148 eyes with KC, 177 eyes with forme fruste KC, 372 healthy control eyes 100% sensitivity and 99.5% specificity for KC eyes, 93.6% sensitivity and 97.2% specificity for forme fruste KC eyes
Kovacs et al. [25] 2016 Scheimpflug camera Neural network 60 eyes with KC, 15 eyes with preclinical KC, 60 healthy control eyes 0.99 AUC, 100% sensitivity and 98% specificity for KC eyes, 0.96 AUC, 92% sensitivity and 85% specificity for preclinical KC eyes
Saad et al. [26] 2016 Placido based corneal topography and corneal wavefront measurements Neural network 62 eyes with forme fruste KC, 114 normal eyes 0.97 AUC, 63% sensitivity and 82% for forme fruste KC, 100% sensitivity and 82% specificity for KC eyes
Hidalgo et al. [27] 2016 Scheimpflug camera SVM 454 eyes with KC, 67 eyes with forme fruste KC, 194 normal eyes 98.9% accuracy, 99.1% sensitivity and 98.5% specificity for KC eyes, 93.1% accuracy, 79.1% sensitivity and 97.7% specificity for forme fruste KC eyes
Ambrosio et al. [21] 2017 Scheimpflug camera and biomechanical camera SVM, random forest 111 eyes with KC, 227 normal eyes 1.0 AUC for KC eyes
Lopes et al. [12] 2018 Scheimpflug camera Random forest 71 eyes with ectasia susceptibility, 182 eyes with KC, 2980 normal eyes 85.2% sensitivity and 0.966 specificity, 0.968 AUC for suspected KC eyes.
Issarti et al. [28] 2019 Scheimpflug camera Neural network 77 eyes with suspect KC, 312 normal eyes 96.56% accuracy, 97.78% sensitivity and 95.56% specificity for suspect KC eyes
  1. KC= keratoconus; UHR-OCT= ultra-high-resolution optical coherence tomography; ML= machine learning