Electroencephalogram (EEG)-based computer-aided technique to diagnose major depressive disorder (MDD)

Abstract Recently, Electroencephalogram (EEG)-based computer-aided (CAD) techniques have shown their promise as decision-making tools to diagnose major depressive disorder (MDD) or simply depression. Although the research results have motivated the use of CAD techniques to help assist psychiatrists in clinics yet their clinical translation has been less clear and remains a research topic. In this paper, a proposed machine learning (ML) scheme was tested and validated with resting-state EEG data involving 33 MDD patients and 30 healthy controls. The EEG-derived measures such as power of different EEG frequency bands and EEG alpha interhemispheric asymmetry were investigated as input features to the proposed ML scheme to discriminate the MDD patients and healthy controls, and to prove their feasibility for diagnosing depression. The acquired EEG data were subjected to noise removal and feature extraction. As a result, a data matrix was constructed by the columns-wise concatenation of the extracted features. Furthermore, the z-score standardization was performed to standardize each column of the data matrix according to its mean and variance. The data matrix may have redundant and irrelevant features; therefore, to determine the most significant features, a weight was assigned to each feature based on its ability to separate the target classes according to the criterion, i.e., receiver operating characteristics (roc). Hence, only the most significant features were used for testing and training the classifier models: Logistic regression (LR), Support vector machine (SVM), and Naive Bayesian (NB). Finally, the classifier models were validated with 10-fold cross-validation that has provided the performance metrics such as test accuracy, sensitivity, and specificity. As a result of the investigations, most significant features such as EEG signal power and EEG alpha interhemispheric asymmetry from the brain areas such as frontal, temporal, parietal and occipital were found significant. In addition, the proposed ML framework proved automatic identification of aberrant EEG patterns specific to disease conditions and provide high classification results i.e., LR classifier (accuracy = 97.6%, sensitivity = 96.66%, specificity = 98.5%), NB classification (accuracy = 96.8%, sensitivity = 96.6%, specificity = 97.02%), and SVM (accuracy = 98.4%, sensitivity = 96.66%, specificity = 100%). In conclusion, the proposed ML scheme along with the EEG signal power and EEG alpha interhemispheric asymmetry are proved suitable as clinical diagnostic tools for MDD.