Moderate traumatic brain injury identification for MEG data using PU (Positive and Unseen) learning

Traumatic brain injury (TBI) is a source of disability and morbidity worldwide. Mild cognitive impairment (MCI) and mild TBI cause functional connectivity interruption for a very limited time frame; however, the patient diagnosed with moderate to severe forms of TBI requires quick, hassle free and precise identification of functional deficits in order to provide timely care. Magnetoencephalography (MEG) is the neuroimaging modality that provides the required information, and is useful for non-contact recording of functional connectivity assessment of TBI subjects. Default mode network (DMN) has been studied and described using functional magnetic resonance imaging (fMRI). This paper proposes a method to develop a default model of biomagnetic activations, as sensed over cortical region using MEG scans. The model is used to classify and assess TBI subjects. The classification is performed by devising default coherence limits between all pairs of MEG sensors for positive (control) group, and the assessment of severity is carried out by using PU learning method (single class model), where P (positive) data is from control population is utilized to compute significant functional connectivity deficits. © 2017 IEEE.