Seizure forecasting with multiple timescales and features.

Area of Science:

• Neurology
• Machine Learning
• Biomedical Engineering

Background:

• Epileptic seizure forecasting remains a significant clinical challenge.
• Existing seizure prediction methods often rely on short-term intracranial EEG (iEEG) data, limiting clinical applicability.
• There is a need for reliable seizure forecasting models using long-term iEEG recordings.

Purpose of the Study:

• To propose and evaluate a machine learning algorithm for forecasting epileptic seizures 2-4 minutes prior to onset.
• To identify optimal EEG features and analysis methods for reliable seizure forecasting on long-term iEEG data.
• To enable timely patient intervention and automated device activation.

Main Methods:

• Implementation of a multiple long-time scale cycle feature analysis framework for seizure forecasting.
• Inclusion of state-of-the-art time series features: critical slowing down (autocorrelation, variance), interictal epileptiform discharge (IED) spike rate, High Frequency Activity (HFA), univariate features, and Neural Mass Model (NMM) features.
• Analysis of seizure phase histograms and synchronization indices (SI) on fast and slow time scales using iEEG data from 14 epilepsy patients.

Main Results:

• The combination of 'autocorrelation + variance + NMM + spike rate' features demonstrated the highest average Area Under the Curve (AUC) of 0.83.
• This feature set showed superior performance in forecasting seizures across multiple patients and time scales.
• Comparative analysis highlighted the performance of various seizure forecasting features on long-term recordings.

Conclusions:

• A novel machine learning model for seizure forecasting achieves performance comparable to state-of-the-art methods.
• The proposed model does not require pre-selection of the optimal EEG channel.
• The study provides valuable insights into the comparative performance of different seizure forecasting features on long-term iEEG data.