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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Psychiatry

Background:

• Early depression diagnosis is vital for effective treatment and reducing mortality.
• Traditional 128-channel EEG is complex and less accessible for widespread screening.
• A simpler, more accessible diagnostic method is needed for depression.

Purpose of the Study:

• To evaluate the efficacy of a simplified 3-channel EEG for depression diagnosis.
• To explore linear and non-linear EEG signal characteristics for differentiating depressed patients.
• To compare the performance of various feature selection and classification algorithms.

Main Methods:

• Collected 3-channel EEG data (Fp1, Fpz, Fp2) from 152 depressed patients and 113 healthy controls.
• Analyzed linear (e.g., variance, kurtosis) and non-linear (e.g., entropy, complexity) EEG features.
• Employed four feature selection algorithms (WrapperSubsetEval, CorrelationAttributeEval, GainRatioAttributeEval, PrincipalComponents) and five classifiers (SVM, KNN, Decision Tree, Logistics Regression, Random Forest).

Main Results:

• The WrapperSubsetEval feature selection method demonstrated superior performance across all classifiers.
• The highest classification accuracy achieved was 76.4% in distinguishing depressed from non-depressed individuals.
• Specific linear and non-linear EEG features effectively differentiated between the two groups.

Conclusions:

• A simplified 3-channel EEG analysis is a viable and accessible tool for aiding depression diagnosis.
• This method can complement clinical assessments, potentially improving diagnostic accessibility and efficiency.
• Further research can refine this approach for broader clinical application in mental health screening.