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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Traditional feature selection methods like Fisher's discriminant ratio (FDR) often yield highly similar features when dealing with limited clinical data and numerous features.
• This redundancy can impact the effectiveness of diagnostic models in neurological disorders.

Purpose of the Study:

• To develop and validate a novel feature selection technique that reduces redundancy by integrating FDR with feature correlation analysis.
• To improve the accuracy of distinguishing individuals with schizophrenia from healthy controls using magnetoencephalography (MEG) data.

Main Methods:

• An attention network test was conducted on schizophrenic patients and normal subjects using a 152-channel magnetoencephalograph.
• P300m amplitudes from event-related fields (ERFs) were extracted as features at both sensor and source levels (500 cortical nodes).
• Features were ranked using FDR and cross-correlation, with the top 10 features selected for an exhaustive search to maximize classification accuracy.

Main Results:

• At the sensor level, a single occipital channel achieved 89.7% accuracy in differentiating the groups.
• At the source level, an accuracy of 96.2% was achieved using two features from the left superior frontal and left inferior temporal regions.
• The proposed method significantly outperformed the traditional FDR-only approach (88.5% accuracy at source level).

Conclusions:

• The combined FDR and feature correlation method effectively reduces feature redundancy and enhances classification accuracy for schizophrenia detection.
• Source-level analysis using P300m amplitudes from specific brain regions offers a highly accurate approach for identifying schizophrenia.
• This method holds promise for improving diagnostic tools in psychiatric neuroimaging.