Crowdsourcing Algorithms To Predict Seizures

65 million people are estimated worldwide to be affected by epilepsy which is different among each patient. Researchers used data and intelligence from international scientists to achieve advances in epileptic seizure prediction performance for patients with the hardest seizures to predict.

The AES-MathWorks-NIH Seizure Prediction Challenge was ran on online data science competition platform on Kaggle.com, focusing on prediction using long term electrical human brain activity recordings obtained in the clinical trial of NeuroVista Seizure Advisory implant system. The top algorithms were rigorously evaluated in this study with more than 646 participants, 478 teams, and upwards of 10,000 algorithms submitted globally.

Algorithms were developed that could distinguish between 10 minute pre-seizure data clips vs inter-seizures, with top algorithms being tested on patients with lowest seizure prediction performance based on previous trials. Evaluation revealed improvements on average of 90% in seizure prediction performance, with different algorithms being shown to perform better for different patients, findings which support use of patient specific algorithms and long term monitoring.

Epilepsyecosytem.org was developed building on findings, which is an online ecosystem for algorithm and data sharing for further development and improvement of seizure prediction.

Greater accuracy in prediction of seizures may greatly improve epilepsy management via offering early warnings triggering interventions. Benefits of crowdsourcing a massive database of algorithms highlighted benefit that can be trained for individual patients to choose the best algorithm for prospective real time seizure prediction by bringing together the best scientists and pooling the greatest algorithms to advance epilepsy research.

The goal is to make seizures less like earthquakes that can strike without warning and more like hurricanes where there is enough advanced warning to seek safety.