Classification of single-trial MEG during sentence processing for automated schizophrenia screening

This paper presents a novel computer-aided system for assisting schizophrenia (SZ) diagnosis. Power Spectral Density Ratios (PSDRs) covering 7 brain regions and 5 frequency sub-bands are extracted as features, from single-trial magnetoencephalography (MEG) recorded while subjects read sentence stimuli silently. A two-stage feature selection algorithm combining F-score and Adaptive Boosting (Adaboost) model is proposed to rank the features. The top ranked features are used to build a boosted non-linear classifier using linear decision stumps as the base classifiers. A majority voting scheme is employed to combine single trial classification results from each test subject to make final classification decisions. Following a leave-one-out cross validation procedure, the proposed system achieves 82.61% classification accuracy (92.31% specificity and 70% sensitivity) on 13 healthy controls and 10 SZ patients. The most discriminating PSDR features are selected from the right temporal, right parietal and right frontal regions and are related to alpha (8-13Hz) and beta (13-30Hz) frequency ranges. This information may help in gaining knowledge about the abnormal neural oscillations associated with sentence-level language disorder in SZ.