Abstracts

Rationale: Depression refers to a wide range of mental health problems characterized by the absence of a positive affect, low mood and a range of associated emotional, cognitive, physical and behavioral symptoms. Diagnosing depression is based mainly on information provided by the patient. Hence, the diagnostic method in use today tends to be subjective and sometimes can be considered as inadequate and often mistaken. As a result, it is essential to develop an objective diagnosis device based on Electroencephalogram (EEG) recording, inexpensive and non-invasive test using machine learning techniques. Methods: In this study, the EEG data was obtained from Beilinson Hospital, Neurology Department, Rabin Medcical Center, Israel. Looking at the details, the complete dataset consists of 17 depression and 17 normal EEG records, and each record contains 18 channels. The sampling frequency was 500 Hz and all signals were low-pass filtered with 70 Hz cutoff frequency. During the research, two consecutive and randomized minutes of recording were analyzed and each channel (60,000 samples) was decomposed in to 239 windows using Hamming window with 50% overlap (window size = 500 samples). The first step in EEG analysis was to decompose the signals using Discrete Wavelet Transform (DWT). The Mother wavelet was Deubechies 7 and 7 levels were used because after seven levels of decomposition, the signals were decomposed into five EEG sub-bands that approximate to Delta (0-4Hz), Theta (4-8Hz), Alpha (8-15Hz), Beta (15-30Hz) and Gamma (30-100Hz). Then, two groups of features were extracted: frequency-domain features and time-domain features. The frequency-domain features refer to the power band of each EEG frequency sub-band (α, β, γ, δ, θ). Whereas the time-domain features were calculated for the reconstructed EEG sub-bands and include: maximum of each reconstructed sub-band, minimum of each reconstructed sub-band, mean of each reconstructed sub-band and standard deviation of each reconstructed sub-band. In the next stage, the analyzed data was split in to training data and testing data 70% and 30% respectively. Then, training data was divided to training set (70%) which was used as an input to the classifiers in order to train the model and Validation set (30%). While the Validation set had a central role in assessing classifiers' optimal parameters, the testing set was later used in order to evaluate the performance of the model. Principal Component Analysis (PCA) was applied for dimensionality reduction and three classifiers, such as Support Vector Machine (SVM), Random Forest (RF) and K-Nearest Neighbor (KNN) were examined for classification. The last step of the analysis was to estimate the classifiers' performance by examination of the following measurements: Accuracy, Sensitivity, Specificity and AUC (Area Under Curve). Results: The findings of this study illustrate that the highest classification accuracy using PCA+SVM is marked among time-domain features (91.21%), while the highest classification accuracy using PCA+RF and PCA+KNN is marked among all features (79.3% and 77% respectively).  Conclusions: SVM classifier tends to be implemented most efficiently with time-domain features, while RF and KNN represent the highest accuracy with all feature. Another interesting finding is that SVM was found to be the best classifier among the mentioned above. The present project illustrates that, the application of time-domain features extraction together with SVM classifier can serve as a promising alternative for intelligent diagnosis system in the future.  Funding: No funding