Study on Feature Selection Methods for Depression Detection Using Three-Electrode EEG Data.

The early diagnosis of depression is important to the treatment of this condition, whereas a timely diagnosis can reduce the incidence of mortality caused in patients with depression. In the present study, we collected the EEG signals of Fp2, Fpz and Fp1, compared with 128 channels EEG, a simpler test (3 channels EEG) can make diagnosis more accessible and widespread, researchers can perform more tests on more patients given the same amount of time and money. The difference between the depressed and the non-depressed patients was explored by the linear and non-linear characteristics of these EEG signals. A total of 152 patients with depression and 113 healthy subjects participated in the study. In the current report, the linear features were as follows: peak, variance, inclination, kurtosis and Hjorth parameter. The nonlinear features included C0 complexity, correlation dimension, Shannon entropy, Kolmogorov entropy and power spectrum entropy. With regard to the aforementioned characteristics, the present report utilized four feature selection algorithms, namely WrapperSubsetEval, CorrelationAttributeEval, GainRatioAttributeEval, and PrincipalComponents and five classification algorithms that included Support Vector Machine, K-Nearest Neighbor, Decision Tree, Logistics Regression and Random Forest. The experimental results indicated that the WrapperSubsetEval of the wrapper class exhibited higher performance compared with the other three feature selection algorithms on each classifier, whereas the highest classification accuracy was 76.4. It is suggested that this analysis may be a complementary tool to aid psychiatrists in the diagnosis of depressed patients.