Seizure Forecasting by High-Frequency Activity (80-170 Hz) in Long-term Continuous Intracranial EEG Recordings
Zhuying Chen1, Matias I Maturana2, Anthony N Burkitt2
1From the Department of Biomedical Engineering (Z.C., A.N.B., M.J.C., D.B.G.) and Graeme Clark Institute for Biomedical Engineering (M.J.C., D.B.G.), University of Melbourne, Parkville; Department of Medicine (Z.C., M.I.M., M.J.C., D.B.G.), St Vincent's Hospital; and Seer Medical (M.I.M.), Melbourne, VIC, Australia. zhuyingc@student.unimelb.edu.au.
Background And Objectives:
Reliable seizure forecasting has important implications in epilepsy treatment and improving the quality of lives for people with epilepsy. High-frequency activity (HFA) is a biomarker that has received significant attention over the past 2 decades, but its predictive value in seizure forecasting remains uncertain. This work aimed to determine the utility of HFA in seizure forecasting.
Methods:
We used seizure data and HFA (80-170 Hz) data obtained from long-term, continuous intracranial EEG recordings of patients with drug-resistant epilepsy. Instantaneous rates and phases of HFA cycles were used as features for seizure forecasting. Seizure forecasts based on each individual HFA feature, and with the use of a combined approach, were generated pseudo-prospectively (causally). To compute the instantaneous phases for pseudo-prospective forecasting, real-time phase estimation based on an autoregressive model was used. Features were combined with a weighted average approach. The performance of seizure forecasting was primarily evaluated by the area under the curve (AUC).
Results:
Of 15 studied patients (median recording duration 557 days, median seizures 151), 12 patients with >10 seizures after 100 recording days were included in the pseudo-prospective analysis. The presented real-time phase estimation is feasible and can causally estimate the instantaneous phases of HFA cycles with high accuracy. Pseudo-prospective seizure forecasting based on HFA rates and phases performed significantly better than chance in 11 of 12 patients, although there were patient-specific differences. Combining rate and phase information improved forecasting performance compared to using either feature alone. The combined forecast using the best-performing channel yielded a median AUC of 0.70, a median sensitivity of 0.57, and a median specificity of 0.77.
Discussion:
These findings show that HFA could be useful for seizure forecasting and represent proof of concept for using prior information of patient-specific relationships between HFA and seizures in pseudo-prospective forecasting. Future seizure forecasting algorithms might benefit from the inclusion of HFA, and the real-time phase estimation approach can be extended to other biomarkers.
Classification Of Evidence:
This study provides Class IV evidence that HFA (80-170 Hz) in long-term continuous intracranial EEG can be useful to forecast seizures in patients with refractory epilepsy.
