Multiscale predictive modeling robustly improves the accuracy of pseudo-prospective seizure forecasting in drug-resistant epilepsy.

Extensive research over the past two decades has focused on identifying a preictal period in scalp as well as intracranial EEG (iEEG). This has led to a plethora of seizure prediction and forecasting algorithms which have reached only moderate success on curated and pre-segmented EEG datasets (accuracy/AUC ≳ 0.8). Furthermore, when tested on their ability to pseudo-prospectively predict seizures from continuous EEG recordings, all existing algorithms suffer from low sensitivity (large false negatives), high time in warning (large false positives), or both. In this study we provide pilot evidence that predictive modeling of the dynamics of iEEG features (biomarkers), seizure risk, or both at the scale of tens of minutes can significantly improve the pseudo-prospective accuracy of almost any state-of-the-art seizure forecasting model. In contrast to the bulk of prior research that has focused on designing better features and classifiers, we start from off-the-shelf features and classifiers and shift the focus to learning how iEEG features (classifier input) and seizure risk (classifier output) evolve over time. Using iEEG from $n=5$ patients undergoing presurgical evaluation at the Hospital of the University of Pennsylvania and six state-of-the-art baseline models, we first demonstrate that a wide array of iEEG features are highly predictable over time, with over 99% and 35% of studied features, respectively, having $R^2>0$ for 10-second- and 10-minute-ahead prediction (mean $R^2$ of 0.85 and 0.2). Furthermore, in almost all patients and baseline models, we observe a strong correlation between feature predictability (with some features remaining predictable up to 30 minutes) and classification-based feature importance. As a result, we subsequently demonstrate that adding an autoregressive model that predicts iEEG features on 12 ± 4 minutes into the future is almost universally beneficial, with a mean improvement of 28% in terms of area under pseudo-prospective sensitivity-time in warning curve (PP-AUC). Addition of the second autoregressive predictive model at the level of seizure risk further improved accuracy, with a total mean improvement of 51% in PP-AUC. Our results provide pioneering evidence for the long-term predictability of seizure-relevant iEEG features and the vast utility of time series predictive modeling for improving seizure forecasting using continuous intracranial EEG.