A Spectral Clustering Approach for the Classification of Waveform Anomalies in High-Dimensional Brain Signals

Transient electrophysiological anomalies in the human brain have been associated with neurological disorders such as epilepsy, may signal impending adverse events (e.g, seizurse), or may reflect the effects of a stressor, such as insufficient sleep. These, typically brief, high-frequency and heterogeneous signal anomalies remain poorly understood, particularly at long time scales, and their morphology and variability have not been systematically characterized. In continuous neural recordings, their inherent sparsity, short duration and low amplitude makes their detection and classification difficult. In turn, this limits their evaluation as potential biomarkers of abnormal neurodynamic processes (e.g., ictogenesis) and predictors of impending adverse events. A novel algorithm is presented that leverages the inherent sparsity of high-frequency abnormalities in neural signals recorded at the scalp and uses spectral clustering to classify them in very high-dimensional signals spanning several days. It is shown that estimated clusters vary dynamically with time and their distribution changes substantially both as a function of time and space.