Prediction of drug response in major depressive disorder using ensemble of transfer learning with convolutional neural network based on EEG

Abstract

Major Depressive Disorder (MDD) is one of the leading causes of disability worldwide. Prediction of response to Selective Serotonin Reuptake Inhibitors (SSRIs) antidepressants in patients with MDD is necessary for preventing side effects of mistreatment. In this study, a deep Transfer Learning (TL) strategy based on powerful pre-trained convolutional neural networks (CNNs) in the big data datasets is developed for classification of Responders and Non-Responders (R/NR) to SSRI antidepressants, using 19-channel Electro-encephalography (EEG) signal acquired from 30 MDD patients in the resting state. Multiple time–frequency images are obtained from each EEG channel using Continuous Wavelet Transform (CWT) for feeding into pre-trained CNN models that are VGG16, Xception, DenseNet121, MobileNetV2 and InceptionResNetV2. Our plan is to adapt and fine-tune the weights of networks to the target task with the small-sized dataset. Finally, to improve the recognition performance, an ensemble method based on majority voting of outputs of five mentioned deep TL architectures has been developed. Results indicate that the best performance among basic models achieved by DenseNet121 with accuracy, sensitivity and specificity of 95.74%, 95.56% and 95.64%, respectively. An Ensemble of these basic models created to surpass the accuracy obtained by each individual basic model. Our experiments show that ensemble model can gain accuracy, sensitivity and specificity of 96.55%, 96.01% and 96.95%, respectively. Therefore, proposed ensemble of TL strategy of pre-trained CNN models based on WT images obtained from EEG signal can be used for antidepressants treatment outcome prediction with a high accuracy.