Development and Validation of a Deep Survival Model to Predict Time-to-Seizure from Routine EEG

Objective

To develop and validate a deep survival model (EEGSurvNet) that analyzes routine EEG to predict individual seizure risk over time, comparing its performance to traditional clinical predictors such as interictal epileptiform discharges (IEDs).

Methods

We conducted a retrospective cohort study including 1,014 consecutive routine EEGs from 994 patients recorded at a tertiary epilepsy center. We developed EEGSurvNet, a deep learning model that predicts time-to-next-seizure over a two-year horizon from a single EEG. Model performance was evaluated on a temporally-shifted testing set of 135 EEGs from 115 patients using time-dependent area under the receiver operating characteristic curve (AUROC), AUROC integrated over two years (iAUROC), and C-index. We compared the deep survival model to a clinical Cox model incorporating standard risk factors as well as a random model based on baseline seizure risk.

Results

EEGSurvNet achieved a two-year iAUROC of 0.69 (95%CI: 0.64–0.73) and C-index of 0.66 (0.60–0.73), outperforming both clinical and random models. Performance was highest in the first months following EEG, peaking at 2 months (AUROC = 0.80). Combining EEGSurvNet to clinical predictors further improved performances (iAUROC = 0.70, C = 0.69). Notably, the model showed superior discrimination on EEGs without IEDs (iAUROC = 0.78 vs 0.53). Model interpretation revealed that the temporal-occipital regions and 6–15 Hz frequencies contributed most to risk prediction.

Significance

EEGSurvNet demonstrates that deep learning can extract prognostic information from routine EEG beyond visible epileptiform abnormalities, potentially improving patient counseling and treatment decisions. Future prospective studies are needed to validate these findings and assess their clinical impact.

Key Points

• Deep learning model predicts individual seizure risk from routine EEG over 2 years

• Model performs better on EEGs without epileptiform discharges, suggesting novel biomarkers

• Temporal-occipital regions and 6-15 Hz frequencies contribute most to risk prediction

• Combined clinical-EEG model achieves best performance for seizure risk stratification