Seizure prediction with long-term iEEG recordings: What can we learn from data nonstationarity?

TL;DR

This study investigates how long-term nonstationarity in iEEG recordings affects seizure prediction, emphasizing the importance of adaptive retraining and meta-state analysis for improved accuracy.

Contribution

It demonstrates the impact of long-term nonstationarity and meta-state switching on seizure prediction, proposing adaptive retraining strategies based on these dynamics.

Findings

Long-term iEEG exhibits weekly/monthly meta-states.

Meta-state switching influences seizure prediction accuracy.

Adaptive retraining improves prediction performance.

Abstract

Repeated epileptic seizures impair around 65 million people worldwide and a successful prediction of seizures could significantly help patients suffering from refractory epilepsy. For two dogs with yearlong intracranial electroencephalography (iEEG) recordings, we studied the influence of time series nonstationarity on the performance of seizure prediction using in-house developed machine learning algorithms. We observed a long-term evolution on the scale of weeks or months in iEEG time series that may be represented as switching between certain meta-states. To better predict impending seizures, retraining of prediction algorithms is therefore necessary and the retraining schedule should be adjusted to the change in meta-states. There is evidence that the nature of seizure-free interictal clips also changes with the transition between meta-states, accwhich has been shown relevant for seizure prediction.