Abstracts

Rationale: Affecting 1% of the world population, epilepsy ranks fourth in severity by the Global Burden of Disease project. Current epilepsy treatment relies on visual electroencephalography (EEG) inspection for seizure detection and localization, which is not only time- and labor-intensive but also subjective. Furthermore, one-third of patients who have refractory epilepsy face limited pharmacological options and turn to surgical interventions. Thus, there is an urgent need for the identification of quantitative EEG (qEEG) features that may be undetectable by eye to further our understanding of seizures, improve therapies, and provide consistent, accessible care.

Methods: Seizure analyses can benefit from information theoretic measures like entropy: the information (in bits) contained in a signal. However, calculating joint entropy for multi-dimensional signals like a multi-electrode EEG is combinatorially intractable. In this work, we propose the inverse compression ratio (ICR), an unbiased estimate for the upper bound of joint entropy, as a potential qEEG method to analyze seizure activity. Using our data repository consisting of continuous, 10 kHz intracranial EEGs acquired from clinical neuromonitoring studies of adult and pediatric participants, ICR was computed over the entire EEG recording of up to two weeks duration. Its performance was compared to standard qEEG methods in the field: variance, Shannon entropy, approximate entropy, and sample entropy.

Results: Our results showed a sharp ICR peak at seizure onset, followed by a large dip before returning to baseline. To assess generalizability, ICR was computed over 29 participants (18 adults and 11 children with a total of 274 seizures), where similar trends were observed. Performance was computed using a patient-specific threshold that determined the classification of seizure and non-seizure states. The median sensitivity/specificity of ICR, variance, Shannon entropy, approximate entropy, and sample entropy were 75/99, 53/99, 70/98, 33/97, 36/96, respectively. Moreover, ICR peak amplitudes changed noticeably across electrodes for those located in seizure onset zones.

Conclusions: There were prominent changes in ICR around seizure onset times and zones. The ICR peak at seizure onset can be explained by the influx of information associated with the abrupt transition from a non-seizure to seizure state. The imminent dip can be attributed to the synchronous firing of neurons during seizures; this fall in signal complexity is consistent with previous studies. The high signal quality and sampling rate of the datasets may have helped in generating pronounced ICR trends.

Due to its permutation sensitivity and generalized nature, ICR may be a compelling way to analyze high-dimensional time-series data like a multi-electrode EEG to characterize as well as to detect and localize seizures.

Funding: Please list any funding that was received in support of this abstract.: This work is supported by Stanford University Wu Tsai Neurosciences Institute and Stanford Bio-X Seed Grant Award IIP9-104.