DB-FFR: A Modified Feature Selection Algorithm to Improve Discrimination Rate Between Bipolar Mood Disorder (BMD) and Schizophrenic Patients

The results of differentiating schizophrenia (SZ) from bipolar mood disorder (BMD) patients by classifying their electroencephalogram (EEG) features remain unconvincing. To overcome this deficiency, an efficient feature selection algorithm, called Davies-Bouldin fast feature reduction (DB-FFR), is proposed here to select the most discriminative features and therefore enhancing the classification rate. Here, 27 patients with BMD and 26 patients with SZ were voluntary-enrolled and their EEG signals were recorded from 22 channels, in both eyes-closed/eyes-open conditions. Then, autoregressive model parameters, fractal dimensions (Katz, Higuchi and Lyapunov exponent) and kurtosis features were extracted from their EEGs. The proposed DB-FFR along with Plus-L Minus-R, Tabu search and fast correlation-based filter was applied to the extracted features, and then the selected features by each method were separately applied to a modified version of nearest neighbor classifier. Experimental results provided 87.51% classification accuracy using DB-FFR, which was statistically (P < 0.05) superior to that of other counterparts. The robustness of DB-FFR was investigated in the presence of additive white noise with different signal-to-noise ratio. Empirical results imply our method could select better discriminative subset of features compared to its rivals.