Anomaly Detection of Moderate Traumatic Brain Injury Using Auto-Regularized Multi-Instance One-Class SVM

Detection and quantification of functional deficits due to moderate traumatic brain injury (mTBI) is crucial for clinical decision-making and timely commencement of functional therapy. In this work, we explore magnetoencephalography (MEG) based functional connectivity features i.e. magnitude squared coherence (MSC) and phase lag index (PLI) to quantify synchronized brain activity patterns as a means to detect functional deficits. We propose a multi-instance one-class support vector machine (SVM) model generated from a healthy control population. Any dispersion from the decision boundary of the model would be identified as an anomaly instance of mTBI case (Glasgow Coma Scale, GCS score between 9 and 13). The decision boundary was optimized by considering the closest anomaly (GCS =13) from the negative class as a support vector. Validated against magnetic resonance imaging (MRI) data, the proposed model at high beta band yielded an accuracy of 94.19% and a sensitivity of 90.00%, when tested with our mTBI dataset. The results support the suggestion of multi-instance one-class SVM for the detection of mTBI.