Identifying the location of spinal cord injury by support vector machines using time-frequency features of somatosensory evoked potentials

Somatosensory evoked potentials (SEP) have been found to contain a series of time-frequency components that conveys information about the location of neurological deficits within the spinal cord. This study aims to develop a classification system for identifying the location of neurological deficit in cervical spinal cord based on the time-frequency patterns of SEPs. Waveforms of SEPs after compressive injuries at various locations (C4, C5, and C6) of rats' spinal cord were decomposed into a series of time-frequency components (TFCs) by a high resolution time-frequency analysis method, matching pursuit (MP). A classification system was build according to the distributional distinction of these TFCs among different levels using support vector machine (SVM). This distinction manifests itself in different categories of SEP TFCs. High-energy TFCs of normal state SEP have significantly higher power and frequency compared with those of injury state SEP. The level of C5 is characterized by a unique distribution pattern of middle-energy TFCs. And the difference between C4 and C6 level is evidenced by the distribution pattern of low-energy TFCs. The proposed classification system was proved to be able to distinguish the four functional status (normal, injury at C4, C5, and C6) with an accuracy of 80.17%.