Artificial intelligence for predicting orthodontic patient cooperation: Voice records versus frontal photographs

Abstract

The purpose of this study was to compare the predictive ability of different convolutional neural network (CNN) models and machine learning algorithms trained with frontal photographs and voice recordings. Two hundred and thirty-seven orthodontic patients (147 women, 90 men, mean age 14.94 ± 2.4 years) were included in the study. According to the orthodontic patient cooperation scale, patients were classified into two groups at the 12th month of treatment: Cooperative and non-cooperative. Afterward, frontal photographs and text-to-speech voice records of the participants were collected. CNN models and machine learning algorithms were employed to categorize the data into cooperative and non-cooperative groups. Nine different CNN models were employed to analyze images, while one CNN model and 13 machine learning models were utilized to analyze audio data. The accuracy, precision, recall, and F1-score values of these models were assessed. Xception (66%) and DenseNet121 (66%) were the two most effective CNN models in evaluating photographs. The model with the lowest success rate was ResNet101V2 (48.0%). The success rates of the other five models were similar. In the assessment of audio data, the most successful models were YAMNet, linear discriminant analysis, K-nearest neighbors, support vector machine, extra tree classifier, and stacking classifier (%58.7). The algorithm with the lowest success rate was the decision tree classifier (41.3%). Some of the CNN models trained with photographs were successful in predicting cooperation, but voice data were not as useful as photographs in predicting cooperation.