(K-410) Daily seizure risk prediction machine learning models from intracranial EEG

One percent of the global population suffer from epilepsy, a disease characterized by unpredictable and unprovoked seizures. Forecasting the probability of a seizure on a given day would increase the patient’s quality of life by offering them more confidence and freedom, while also minimizing the severity of the upcoming seizure or completely preventing it by the appropriate medication adjustment. Intracranial EEG (iEEG) is an invasive test from which electrical brain activity is recorded with high signal to noise ratio while patients are being monitored for seizures. In this study, we tested if a measure of functional connectivity in the brain could reliably predict the occurrence of a seizure in the next day.

From a cohort of 71 patients, 58 met exclusion criteria. A 20-second long intracranial EEG recording during resting periods from a daily interval between 10am and 12pm was considered for analysis. After artifact rejection and re-referencing the signals to a bipolar montage, we filtered the signals into 6 canonical frequency bands and computed phase locking value (PLV) between all pairs of channels in 2 second windows with 1 second overlap. 

For each patient and frequency band, we fit a Support Vector Machine model on features from all previous days and a binary label of seizure occurrence. We applied Principal Component Analysis (PCA) to prune features and minimize correlated features prior to fitting the model.  The accuracy of the model’s predictions was assessed using Brier Score (BS) and Brier Skill Score (BSS). 

Results:

The Principal Component Analysis resulted in up to 500 times reduction in the number of features used in the Support Vector Machine Model, and minimizing correlations. Across patients, we computed BS and BSS to evaluate model performance (Figure 1A). A logistic regression model fit on the number of testing days and BSS revealed that at least 7 days of test data was necessary to achieve a model with good performance (BS = 0.25) (Figure 1B). For patients with 7 days of test data, we evaluated daily seizure probabilities for the best performing models (Figure 1C).  We found that models best predicted focal aware seizures (FAS) (ANOVA single factor test F = 15.56, p = 6.1e-10, post-hoc Dunn’s test) (Figure 1D). Across frequency bands, all models performed comparably (Friedman Chi-square test F = 10.57, p = 0.06).

Conclusions & Discussions:

We developed a daily seizure risk score by predicting the occurrence of seizures in the upcoming day. Model performance improved with more testing days, suggesting that this analysis could be extended to longitudinal data from implanted EEG devices. Future work should include replicating this analysis on new iEEG samples and developing a robust null model to evaluate performance. Linking this model to other data streams, including medication load, non-invasive commercial watch biometrics, and daily surveys on patient-reported seizure risk would enhance the goal of daily seizure risk reporting.