Seizure forecasting using machine learning models trained by seizure diaries.

Area of Science:

• Neurology
• Data Science
• Biomedical Engineering

Background:

• Refractory epilepsy significantly impacts patients' quality of life due to unpredictable seizures.
• Identifying cyclical seizure patterns can enable accurate forecasting and improve patient management.
• Machine learning (ML) offers potential for detecting non-intuitive seizure regularities.

Purpose of the Study:

• To develop and evaluate ML and statistical models for forecasting seizures in patients with refractory epilepsy.
• To differentiate between predictable and unpredictable seizure patterns using patient-reported data.
• To assess the impact of seizure frequency on prediction accuracy.

Main Methods:

• Utilized seizure logs from 153 patients (8337 seizures) from the Human Epilepsy Project.
• Trained and evaluated forecasting models including statistical approaches (Bayesian fusion) and ML algorithms (SVM, LSTM).
• Employed leave-one-person-out cross-validation for robust model assessment on unseen subjects.

Main Results:

• Support Vector Machine (SVM) regression and a combined statistical model demonstrated superior forecasting performance.
• SVM achieved prediction accuracies of 50% (0 days error) to 87% (3-4 days error) for unseen subjects.
• Higher seizure frequency in patients correlated with improved prediction accuracy.

Conclusions:

• ML models effectively identify non-random patterns in seizure diaries for forecasting future events.
• Diary-based seizure forecasting contributes to understanding individual and population-level seizure predictability.
• This research advances the potential for personalized epilepsy management through predictive analytics.