Machine learning-based hybrid approach for schizophrenia diagnosis using non-stationary EEG signals

A serious mental illness, schizophrenia (SZ) affects 1% of people worldwide and is characterised by delusions, hallucinations and disorganised thought patterns. Diagnosis mostly is based on subjective interviews by a psychiatrist in which there is a high chance of human errors and bias. In this work, we have conducted a comprehensive analysis of electroencephalogram (EEG) data using empirical mode decomposition (EMD) algorithm which can analyse non-stationary and nonlinear signals and separates them into components at different resolutions called intrinsic mode functions (IMFs). In this work, our primary goal is to introduce a hybrid approach for IMF selection that combines four distinct parameters namely correlation, energy, statistical significance and power spectral density (PSD) distance. From the selected IMFs nine statistical features are computed and performance is evaluated using various classifiers. Among all the classifiers k-nearest neighbour (KNN) showed the best accuracy of 90.29% using the second IMF. These results suggest that EEG signals can effectively distinguish between healthy control and SZ patients and have a potential to help psychiatrists for diagnosis of SZ.