Reliable Seizure Prediction Using Physiological Signals and Machine Learning

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant): For most individuals living with epilepsy, seizures are relatively infrequent events occupying a small fraction of their life. Despite spending as little a 0.01% of their lives having seizures (typically only minutes per month), people with epilepsy take anti-epileptic drugs (AED) daily, suffer AED related side effects, and spend their lives dreading when the next seizure will strike. The apparent randomness of seizures is associated with significant psychological consequences. In addition, despite daily AED approximately 1/3 of patients continue to have seizures. We hypothesize that epilepsy can be more effectively treated, both the seizures and their psychological impact, by providing patients with real-time seizure forecasting. Periods of low seizure probability would not require AEDs, or at least lower doses of AEDs, thus reducing AED exposure and their side effects. Periods of high seizure probability may respond to acute AED and patients could alter their activities to avoid injury. Patients would be empowered to manage their medications and life activities using reliable seizure forecasts. In this grant we investigate the hypothesis that seizures are predictable events, and pursue accurate, clinically relevant seizure forecasting using recent advances in support vector machines (SVM), data-analytic models, and Universum-SVM applied to continuous intracranial EEG (iEEG) in focal canine epilepsy. This is an initial step in establishin a new treatment paradigm for focal epilepsy, whereby the probability of seizure occurrence is continuously tracked for patient warning and intelligent responsive therapies. Naturally occurring focal canine epilepsy is an excellent model for investigation of seizure forecasting because of the clinical and electrophsyiological similarity to focal human epilepsy. This study provides a unique opportunity to study seizure forecasting in naturally occurring canine epilepsy under uniform conditions (the same environment). Importantly, dogs are large enough to accommodate devices designed for human use. The hypotheses driving this proposal are that focal seizures are not random events and there are brain states associated with low or high probability of seizure occurrence, and that these states can be reliably classified using machine learning approaches (SVM & Universum-SVM) that combine features from iEEG, behavioral state tracking, and electrocardiogram (ECG) heart rate variability. The goal of this proposal is to develop reliable seizure forecasting (when possible) and improved understanding (data characterization) when good forecasting is not possible.