Modelowanie predykcyjne na wielu skalach czasowych znacznie poprawia dokładność przewidywania napadów padaczki opornej na leki
Multiscale predictive modeling robustly improves the accuracy of pseudo-prospective seizure forecasting in drug-resistant epilepsy
W skrócie
Naukowcy opracowali nową metodę do przewidywania napadów padaczki, która śledzi, jak zmieniają się sygnały mózgu w ciągu kilkunastu minut. Zamiast szukać nowych wskaźników, skupili się na modelowaniu zmian już znanych parametrów mózgu i ryzyka napadu. Wyniki pokazują, że ich metoda jest o połowę lepsza w przewidywaniu napadów padaczki, która nie reaguje na standardowe leki.
Oryginalny abstract (angielski)
Extensive research over the past two decades has focused on identifying a preictal period in scalp and intracranial encephalography (iEEG). This effort has led to numerous seizure prediction and forecasting algorithms, with moderate success on datasets consisting of curated and pre-segmented EEG. When evaluated pseudo-prospectively on continuous EEG recordings, existing algorithms often exhibit low sensitivity, high time in warning, or both. In this study, we investigate whether predictive modeling of temporal dynamics of iEEG features and seizure risk can improve pseudo-prospective forecasting performance.

Approach: Using iEEG data from n = 5 patients undergoing presurgical evaluation at the Hospital of the University of Pennsylvania and six state-of-the-art baseline models, we shift the focus from designing new features and classifiers to modeling temporal evolution of iEEG features (classifier inputs) and seizure risk (classifier outputs). We develop autoregressive models to predict iEEG features and seizure risk over timescales of several minutes and incorporate these predictions into existing forecasting pipelines.

Main results: We first demonstrate that a wide range of iEEG features are predictable over time, with over 99% and 35% of features achieving R> 0 for 10-second and 10-minute-ahead predictions (mean R> of 0.85 and 0.2), respectively. We observe a strong correlation between feature predictability and classification-based feature importance. Accordingly, we show that incorporating an autoregressive model that predicts iEEG features approximately 12 ± 4 minutes into the future improves pseudo-prospective performance, with a mean increase of 28% in the area under the sensitivity versus time-in-warning curve (PP-AUC). The addition of a second autoregressive model at the level of seizure risk yields further gains, resulting in a total mean improvement of 51% in PP-AUC.

Significance: These results provide evidence for long-term predictability of seizure-relevant iEEG features and demonstrate the value of time-series predictive modeling for improving seizure forecasting from continuous intracranial EEG.