Semi-supervised training data selection improves seizure forecasting in canines with epilepsy

Objective: Conventional selection of pre-ictal EEG epochs for seizure prediction algorithm training data typically assumes a continuous pre-ictal brain state preceding a seizure. This is carried out by defining a fixed duration, pre-ictal time period before seizures from which pre-ictal training data epochs are uniformly sampled. However, stochastic physiological and pathological fluctuations in EEG data characteristics and underlying brain states suggest that pre-ictal state dynamics may be more complex, and selection of pre-ictal training data segments to reflect this could improve algorithm performance. Methods: We propose a semi-supervised technique to select pre-ictal training data most distinguishable from interictal EEG according to pre-specified data characteristics. The proposed method uses hierarchical clustering to identify optimal pre-ictal data epochs. Results: In this paper we compare the performance of a seizure forecasting algorithm with and without hierarchical clustering of pre-ictal periods in chronic iEEG recordings from six canines with naturally occurring epilepsy. Hierarchical clustering of training data improved results for Time In Warning (TIW) (0.18 vs. 0.23) and False Positive Rate (FPR) (0.5 vs. 0.59) when evaluated across all subjects (p < 0.001, n = 6). Results were mixed when evaluating TIW, FPR, and Sensitivity for individual dogs. Conclusion: Hierarchical clustering is a helpful method for training data selection overall, but should be evaluated on a subject-wise basis. Significance: The clustering method can be used to optimize results of forecasting towards sensitivity or TIW or FPR, and therefore can be useful for epilepsy management.