Unsupervised seizure detection using modulation spectra measures: A preliminary study

Epileptic seizures affect millions worldwide, impairing their quality of life in incapacitating ways. Many epilepsy patients undergo electroencephalography (EEG) continuously for days to weeks in a hospital to have seizures that might help doctors identify the anatomical focus of those seizures. But screening days to weeks of iEEG for seizures that can happen anytime or not at all is a time-consuming clinical burden for clinicians and staff. Thus, a computerized method to perform this duty objectively in lieu of subjective manual labor is highly beneficial toward robust timely clinical care of patients. I present such a method using two unsupervised machine learning techniques applied to cross-frequency coupling measures, comparing the classification performance of the methods. The methods perform similarly in accuracy, sensitivity, specificity, and selectivity (positive predictive value). Overall a proof-of-concept for a new approach is made. With more development, either approach could be used in practical clinical settings.