Abstracts

Rationale: Although treatment for epilepsy is effective for nearly 70 percent of patients, many remain in need of new therapeutic approaches. Predicting seizures in these patients could enhance their quality of life if the prediction performance is clinically practical. In this study, we investigate the improvement in the performance of a recent seizure prediction algorithm (Gadhoumi et al., Clinical Neurophysiology. 2013;124:1745-54) using a novel measure of scale-free dynamics of the intracerebral EEG. The performance of the algorithm is evaluated on continuous multi-day depth-electrode EEG of 17 patients with intractable mesial temporal lobe epilepsy who underwent presurgical evaluation. Methods: Scale-free dynamics of a system are characterized by the invariance of statistical properties along different time scales. We investigate such dynamics in the intracerebral EEG by means of a robust multifractal analysis based on wavelet leaders and bootstraps (Wendt et al., Signal Processing Magazine, IEEE. 2007;24:38-48). Scale-free properties are quantified through estimates of the scaling exponents — the first and second cumulants ― that characterize the power laws describing the inter-scale relations. The cumulants are first investigated for their ability to discriminate preictal and interictal epochs. The patient-specific classifier-based seizure prediction algorithm is then trained using combinations of cumulants and features originally proposed with the algorithm and referred to as thermodynamic features. Its performance was out-of-sample tested on 1446 h of continuous data containing 128 seizures. The sensitivity, the false prediction rate and the proportion of time under warning were evaluated for seizure prediction horizons ranging between 5 and 60 min. The sensitivity of the algorithm was statistically compared with that of an analytical Poisson random predictor to evaluate its superiority to chance. Results: Using a combination of the first cumulant and the thermodynamic measures, seizures were predictable above the chance level (p-value < 0.05) in 13 of 17 patients (76 %; an increase from 7/17 patients compared to our original results). The average sensitivity reached 80.5 % across patients and the proportion of time under warning was 25.1 % for the critical false prediction rate of 0.15/h (fig. 1). Preictal changes were detected between 27.2 and 90 min. Surgery outcome and seizure above-chance predictability showed no significant association. Conclusions: A substantial improvement in seizure prediction performance was possible by using descriptors of scale-free properties of intracerebral EEG in patients with mesial temporal lobe epilepsy. The results suggest that scale-free dynamics may have different properties in the preictal state compared to the interictal state. Robust quantifiers of these dynamics carry a predictive power useful in seizure prediction and possibly in characterizing the preictal state. Supported by CIHR MOP-10189, RSC-NSERC CHRPJ 323490-06.