Feature optimization method for machine learning-based diagnosis of schizophrenia using magnetoencephalography

• Published in: Journal of neuroscience methods Vol. 338; p. 108688
• Main Authors: Kim, Jieun, Kim, Min-Young, Kwon, Hyukchan, Kim, Ji-Woong, Im, Woo-Young, Lee, Sang Min, Kim, Kiwoong, Kim, Seung Jun
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 15.05.2020
• Subjects: Diagnosis; Feature optimization; Machine learning; Magnetoencephalography; Schizophrenia; Schizophrenia; Diagnosis; Feature optimization; Magnetoencephalography; Machine learning
• ISSN: 0165-0270; 1872-678X; 1872-678X
• DOI: 10.1016/j.jneumeth.2020.108688

•Feature reduction method is proposed when there are redundant or correlated features.•For feature reduction, we use FDR value in conjunction with feature correlation.•An ANT task is used to distinguish patient with schizophrenia.•The 149 and 500 visual P300 m are used as features in sensor and source level.•Our method can be applied to EEG, MEG and fMRI data. When many features and a small number of clinical data exist, previous studies have used a few top-ranked features from the Fisher's discriminant ratio (FDR) for feature selection. However, there are many similarities between selected features. New method: To reduce the redundant features, we applied a technique employing FDR in conjunction with feature correlation. We performed an attention network test on schizophrenic patients and normal subjects with a 152-channel magnetoencephalograph. P300m amplitudes of event-related fields (ERFs) were used as features at the sensor level and P300m amplitudes of ERFs for 500 nodes on the cortex surface were used as features at the source level. Features were ranked using FDR criterion and cross-correlation measure, and then the highest ranked 10 features were selected and an exhaustive search was used to find combination having the maximum accuracy. At the sensor level, we found a single channel of the occipital region that distinguished the two groups with an accuracy of 89.7 %. At source level, we obtained an accuracy of 96.2 % using two features, the left superior frontal region and the left inferior temporal region. At source level, we obtained a higher accuracy than traditional method using only FDR criterion (accuracy = 88.5 %). We used only the P300 m amplitude (not latency) on a single channel and two brain regions at a fairly high rate.