Support vector machine classification of patients with depression based on resting-state electroencephalography

Depression is one of the most common mental disorders. Although depression is typically diagnosed by identifying specific symptoms and through history, no recognized standard for depression diagnosis exists. This assures the development of objective diagnostic tools for depression.

We investigated the differences in the resting-state electroencephalograms (EEGs) of patients with depression and healthy controls (HCs) to distinguish patients from HCs by using a support vector machine (SVM) classifier with the following two feature selection approaches: t test and receiver operating characteristic analysis.

We used the EEG data from the Patient Repository of EEG Data + Computational Tools; this study included 21 patients with depressive disorder (MDD) and 21 HCs. The relative frequency power, alpha interhemispheric asymmetry, left-right coherence, strength, clustering coefficient (CC), shortest path length, sample entropy (SampEn), multiscale entropy (MSE), and detrended fluctuation analysis (DFA) data were extracted to determine candidate EEG features associated with depression.

With the t-test selection, the SVM classifier demonstrated the highest performance with the accuracy, sensitivity, and specificity of 96.66%, 95.93%, and 97.550% for the eye-open condition and 91.33%, 90.59%, and 91.81% for the eye-closed condition, respectively. For comparisons of features in the 2 selection approaches, the most influential features were relative frequency power and left-right coherence.

Using this information to distinguish patients with MDD from HC subjects with the SVM classifier resulted in a mean accuracy over 90%. Although this result may not be robust enough for clinical applications, further exploration is necessary given the simplicity, objectivity, and efficiency of the classifier.