Fig. 5

The micro average ROC of N, PPG, and G groups classified by multimodal models. Multimodal models achieved high AUROCs with the validation data. The RF (99.7%), SVM (99.4%), and DNN (99.1%) models showed similar results with tenfold cross-validation data. The best three models in the test set were the SVM (95.1%), DNN (95.0%), and RF (94.1%). There was a decrease in AUROCs in the test set, but they were still within an acceptable range. N, normal; G, glaucoma; RF, random forest; Ada, adaptive boosting; SVM, support vector machine; LogReg, logistic regression; NB, Naïve Bayes; KNN, k-nearest neighbor; CART: classification and regression decision tree; C4.5, C4.5 decision tree; DNN, dense neural network