Abstracts

Rationale: Epilepsy is a neurological disorder that can affect up to 1% of the population, especially in developing countries.  The gold standard for diagnosis of epilepsy is an electroencephalogram (EEG), and recordings are shown to be most effective if they last up to three or four days.  Experienced EEG readers are typically used to score EEG recordings for the presence of spikes, seizures, and sharp waves.  However, there is a high workload for EEG readers due to the long EEG records obtained to detect seizures, a high inter- and intra-reader variability, and a lack of EEG readers in many developing countries.  Given these limitations, it is imperative that alternate solutions are developed for the detection of spikes and seizures in EEG records. BioSerenity A.I. has developed algorithms to automatically detect spikes and sharp waves (interictal events) as well as seizures in the EEG.

Methods: The seizure algorithm was developed using a regularized gradient boosting classifier to analyze an EEG signal and automatically detect seizures.  Training and testing used the Temple University seizure database v1.2.0, which had been scored and annotated by trained neurologists. The spike algorithm works by pre-detecting candidate interictal events and then performs feature extraction.  It detects interictal events by using a random forest classifier. It was trained and tuned using the annotated Epilepsiae database, and then validated using an external dataset (Temple University seizure database). 

Results: The BioSerenity A.I. seizure algorithm could detect seizures with similar accuracy to predicate algorithms. The algorithm showed positive percent agreement with expert scorers of 73.6%, slightly better than the Persyst algorithm (68%). Its false detection rate for seizures was 11.28 seconds per hour. The spike algorithm had a false positive alarm per hour rate of 22.66. The spike algorithm was able to distinguish between spikes and sharp waves, a feature that is unique to this algorithm. The spike and the seizure algorithms were able to perform well in the presence of missing channels, gaps in the record, and with 10-10 and 10-20 montages. 

Conclusions: The seizure algorithm agreed with expert scorers at a rate akin to inter-rater reliability (70%-80%), and so performs similarly to humans in its ability to detect seizures.  The spike and seizure algorithms were able to detect spikes and seizures despite missing channels and different montages.  These algorithms could take the burden off EEG readers, especially in low resource areas as well as in areas where persons may not be able to easily have their EEGs read.  In conjunction with neurology experts, we propose that the BioSerenity A.I. spike and seizure algorithms could be used to accurately detect seizures and interictal events (such as spikes and sharp waves) that would allow for diagnosis of patients with epilepsy. 

Funding: This study was funded by BioSerenity.