High frequency oscillations (HFOs) have been used for seizure prediction and are promising biomarkers of epileptogenesis. However, detecting HFOs is time consuming and subjective, prompting research into automated detection and classification pipelines. We aim to understand how different EEG filtering methods impact these pipelines and harmonize detections from the same data when preprocessed differently. We preprocessed EEG with two different filters and then detected events with the short time energy (STE) detector and compared common detections. We applied t-distributed stochastic neighbor embedding (t-SNE) to the datasets and compared embeddings then investigated if shifting commonly detected events prior to t-SNE helped standardize embeddings. The finite impulse response (FIR) and infinite impulse response (IIR) filters achieved a Cohen's Kappa coefficient of 0.8962 after shifting, reflecting a high level of agreement. The t-SNE embeddings were similar only when data were shifted prior to embedding. Feasible solutions to this shifting problem are addressed.
The Effects of Filtering on High Frequency Oscillation Classification
Marianna La Rocca;
2020-01-01
Abstract
High frequency oscillations (HFOs) have been used for seizure prediction and are promising biomarkers of epileptogenesis. However, detecting HFOs is time consuming and subjective, prompting research into automated detection and classification pipelines. We aim to understand how different EEG filtering methods impact these pipelines and harmonize detections from the same data when preprocessed differently. We preprocessed EEG with two different filters and then detected events with the short time energy (STE) detector and compared common detections. We applied t-distributed stochastic neighbor embedding (t-SNE) to the datasets and compared embeddings then investigated if shifting commonly detected events prior to t-SNE helped standardize embeddings. The finite impulse response (FIR) and infinite impulse response (IIR) filters achieved a Cohen's Kappa coefficient of 0.8962 after shifting, reflecting a high level of agreement. The t-SNE embeddings were similar only when data were shifted prior to embedding. Feasible solutions to this shifting problem are addressed.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.