Using machine learning classifiers to identify glaucomatous change earlier in standard visual fields.

摘要: PURPOSE To compare the ability of several machine learning classifiers to predict development abnormal fields at follow-up in ocular hypertensive (OHT) eyes that had normal visual baseline examination. METHODS The 114 patients with OHT four or more field tests standard automated perimetry over three years and for whom stereophotographs were available assessed. mean (+/-SD) number was 7.89 +/- 3.04. covered 5.92 2.34 (range, 2.81-11.77). Fields classified as based on Statpac-like methods (Humphrey Instruments, Dublin, CA) by classifiers. two types support vector (SVM), a mixture Gaussian (MoG) classifier, constrained MoG, generalized (MGG). Specificity set 96% all classifiers, using data from 94 evaluated longitudinally. cutoffs required confirmation abnormality. RESULTS Thirty-two percent (36/114) converted during methods. All 36 identified least one classifier. In nearly cases, predicted confirmed abnormality, average, 3.92 0.55 earlier than traditional CONCLUSIONS Machine can learn complex patterns trends adapt create decision surface without constraints imposed statistical This adaptation allowed identify abnormality converts much