Design of efficient bi-orthogonal wavelets for EEG-based detection of Schizophrenia

The state-of-the-art bi-orthogonal wavelet filters need infinite precision implementation due to more number of irrational filter coefficients. This paper presents a novel low-complexity bi-orthogonal wavelet filter-bank (LCBWFB) with canonical-signed-digit filters to reduce the computations. The proposed wavelet filter design uses a generalized matrix formulation technique with sharp roll-off to generate rational coefficients. These filters facilitate near-orthogonality, regularity, and perfect reconstruction. Different lengths of highpass and lowpass filters are generated by varying the half-band polynomial factors. The various combinations of filter banks including 9/7, 10/6, and 11/9 are designed using proposed method. This method provides the freedom to select the parameters according to the size of the filter bank. Comparative analysis with earlier reported bi-orthogonal wavelets showed lower computations and higher regularity for the LCBWFB. These rational coefficients are then used in automatic schizophrenia detection to decompose EEG signals. The Fisher score is used to select the most discriminating channels, and each channel is decomposed using LCBWFB into six subbands. A set of 22 features, comprising statistical, entropy, and complexity, are calculated for each subband. A least square support vector machine is tuned using the Grey Wolf optimizer and is trained using the five most significant features selected using the Wilcoxon Signed-rank test. The 10-fold accuracy of 96.84%, sensitivity of 95.95%, and specificity of 96.97%. These values using leave-one-subject-out are 93.92%, 92.30%, and 93.77%, respectively, obtained for an open-source dataset with only 25 out of 1408 features comparable to existing Schizophrenia detection methods.