MACHINE LEARNING CLASSIFICATION OF FIRST-ONSET DRUG-NAIVE MDD USING STRUCTURAL MRI

초록

Objectives: We tried to train a high-accuracy predictive classification that could discriminate scans from adolescent patients with MDD from scans from healthy control subjects. We wanted to know whether it is possible to reliably train a predictive classification with a high accuracy, even in the first-onset drug-naïve adolescent with MDD just using structural MRI without using any other clinical data from the patient. We also estimated the probability of the subject belonging to the predicted class to determine the confidence of the prediction. Methods: Medication-naïve adolescent patients with the first episode of MDD were prospectively recruited from the child and adolescent mental health clinic. For group comparison, healthy volunteers matched for age, gender, and education years were also prospectively recruited. A total of 6 classification models were used to build the adolescent MDD classification model: tree-based bagging, random forest, multilayered feedforward neural network, adaptive boosting, gradient boosting machine, and support vector machine. Results: The study participants consisted of 27 patients and 27 control subjects. Among the classification models, support vector machine (SVM) yielded the highest F1-score (0.943) followed by adaptive boosting (0.830) and random forest (0.808). The two most significant variables are the standard deviation of intensity of the right ventral diencephalon and the thickness of the superior segment of the circular sulcus of the insula. Conclusions: The structural brain changes can be found even in the adolescents with the first onset of MDD and can be used to build an accurate classification model for machine learning, although the duration of illness is relatively short and the influence of MDD on the brain structure would be minimalized. A participant is diagnosed as an adolescent with MDD when either the variation of intensity of the right ventral diencephalon region or the thickness of superior segment of the circular sulcus of the insula increases.