Learning to generalize seizure forecasts.

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Epilepsy is defined by unpredictable seizures.
• Patient-reported seizures correlate with cyclic interictal epileptiform activity (IEA) over days.
• This multidien interictal-ictal relationship offers potential for seizure forecasting.

Purpose of the Study:

• To rigorously test the generalizability of seizure forecasting models across unseen epilepsy patients.
• To evaluate the efficacy of pretrained algorithms in predicting seizures based on multidien IEA cycles.
• To assess if seizure forecasting performance is independent of data acquisition methods.

Main Methods:

• Utilized retrospective long-term intracranial EEG (icEEG) and subscalp EEG (sqEEG) data from 159 participants.
• Extracted instantaneous multidien phases from IEA detections.
• Trained generalized linear models (GLMs) and recurrent neural networks (RNNs) for 24-hour seizure probability forecasting.

Main Results:

• Forecasting seizures above chance in 79% (GLMs) and 81% (RNNs) of unseen subjects.
• Achieved median Area Under the Curve (AUC) of 0.70 (GLMs) and 0.69 (RNNs).
• Pretrained models showed comparable performance to individualized models but for more subjects, and maintained calibration across varying seizure rates.

Conclusions:

• Seizure forecasting based on multidien IEA cycles generalizes across patients.
• This approach can significantly reduce the data required for individual seizure forecasts.
• Generalization is independent of seizure reporting method (patient-reported vs. electrographic) or IEA recording method (icEEG vs. sqEEG).