Abstracts

Rationale: Seizures are a symptom of underlying brain disease, and recurrent unprovoked of seizures leads to have epilepsy. These occurrences are unanticipated which has significant impact on the quality of life of people with epilepsy. Recording electrical activity in the brain using intracranial electroencephalography (iEEG) empowers detection and prediction of brain wave patterns with the help of machine learning algorithms. Although many predictive machine-learning algorithms have been developed, standardised and benchmarked algorithms will improve the clinical and commercial viability of seizure-prediction devices.

Methods: Using the iEEG data from the first-in-human long-term trial of a seizure prediction device, the ‘Melbourne University AES-Math Works-NIH Seizure Prediction Challenge’ globally crowdsourced predictive algorithms in a standardized framework but improvements in prediction performance are still needed. For ongoing development of these and new algorithms, Epilepsyecosystem.org hosted the contest iEEG data of 3 patients in the form of 10-minute windows without overlap containing preictal (duration of 1 hour) and interictal data for development of machine learning algorithms to distinguish between the two classes. The algorithms were ranked on the contest data based on AUC (area under ROC curve) performance that took into account the number of true and false positives.

Via our Epilepsyecosystem.org website we have also launched an on-going independent evaluation (https://github.com/epilepsyecosystem/CodeEvaluationDocs/tree/master/CodeEvaluationInstructions) of the top ranked algorithms. This evaluation focuses on the full iEEG data record of the 15 patients from the long-term trial. We have independently evaluated top algorithms on the contest data, and we selected the second-place algorithm as the fastest algorithm to be evaluated against full trial of contest dataset. We tested the effect of class imbalance on the training mode of the full trial of the contest data. The average ratio of class imbalance for all 3 patients of contest data in full trial were 2% (preictal) to 98% (interictal). This class imbalance was the indication of the bad behaviour of the algorithm on the full trial dataset with the average performance of 0.51. We encountered this problem by under-sampling the majority class (interictal) to 90% and increased average performance up to 0.75.

Results: The average performance of the algorithm after moderating the class distribution improved by 24%. Detailed AUC performance of full trial before and after class imbalance for the same patients of contest data reported in table 1.

Conclusions: Current results suggest viable algorithms are being developed with the contest data and ongoing work through the independent evaluation on the full trial dataset should yield a new set of high-performing and efficient seizure prediction algorithms. This will help make seizure prediction both clinically and commercially viable and help alleviate the stress associated with seizures.

Funding: Please list any funding that was received in support of this abstract.: NHMRC GNT1160815.