Temporal sequences in epileptic seizure detection

Epilepsy is a chronic disorder that is characterized by recurrent unprovoked and unpredictable seizures. With a prevalence of 0.8% epilepsy is one of the most frequent neurological disorders. Approximately one third of epilepsy patients suffer from difficult-to-treat epilepsy with seizures refractory to currently available antiepileptic drugs. For these patients, epilepsy surgery offers an effective treatment option. Long-term electroencephalogram (EEG) recordings represent the cornerstone of the presurgical workup for these patients. These recordings usually are performed over a period of several days and are extremely time consuming and expensive. An automatic online seizure detection system processing data from the recording system in real-time would therefore be of great benefit for epilepsy monitoring units. Such a system would relieve personnel from continuously monitoring the EEG- signals during recordings. At the same time it will alert medical staff of a beginning seizure in order to assist the patient and to perform neurological and neuropsychological tests during the seizure. Furthermore, such a system would reduce the amount of data that has to be reviewed by the medical staff, thus significantly reducing costs of long time recordings. Another important potential application is the use of a seizure detection system in intensive care units. Several studies have shown that up to 10% of comatose patients in intensive care units suffer from a non-convulsive status epilepticus. Today, this status usually remains unrecognized since EEG-monitoring is performed only on rare occasions due to the complexity of EEG signals and the difficulties of online interpretation. An automatic system that continuously monitors brain functions and immediately alerts the medical personnel of an epileptic seizure therefore could considerably improve the prognosis for patients in neurological intensive care units. Designing a seizure detection system is a difficult task. Numerous attempts have been made to develop such systems. The results are still far from optimal and need a lot of improvement in order to be accepted in the hectic every-day life in a hospital. Such a system must achieve three major requirements: First, the system must work without extensive training with offline data. Second, the system must be able to detect seizures of different patients without manual parameter adjustment. Third, the system must have a low false alarm rate while showing a high sensitivity and a low latency between the onset and the detection of the seizure. The goal of the proposed project is the development of a reliable epileptic seizure detection system suitable for the application in a clinical environment. The project will extend research done by the Vienna Epilepsy Program (ViEP) and the Austrian Research Centers (ARC) over the last three years by further improving an existing experimental seizure detection system in order to reach sensitivity close to 100% with less than 0.5 false alarms per hour. In order to achieve the goals of this project, two different major approaches will be pursued. First we want to exploit characteristics of temporal feature sequences on a short time scale, in particular during epileptic seizures. A first, preliminary analysis of the mathematical features that are already developed for the current seizure detection system showed that certain features tend to change early in the evolution of a seizure while other features tend to react later in time. Within the proposed project we want to study these temporal sequences of EEG-signal characteristics in detail. In addition, we want to study the correlation between the mathematical features and the clinical seizure semiology. This information will then be included into the seizure detection model using sequence detection algorithms based on a Hidden Markov Model (HMM). To our knowledge, no HMM-based method for automatic seizure detection has been published yet. We think that this approach will make seizure detection more reliable, since it exploits important information that has been neglected up to now. For the second major approach pursued within this project we will study the relationship between changes in the dynamics of the EEG-signals during the sleep-wake-cycle of a patient and the false alarm rate of the detector. We will study whether there is a correlation between false alarms of the existing seizure detection system and the sleep-wake-cycle and which features used in the seizure detector are particularly sensitive to artifacts that led to false alarms. This information will then be incorporated into the seizure detection system by a detector that is dependent on sleep-wake-cycles and on probability measures for signal artifacts. To our knowledge a similar approach has not yet been published in the context of epileptic seizure detection.

Epilepsy with a prevalence of 0.8% represents one of the most frequent chronic neurological disorders. The burden of disease caused by chronic epilepsy is higher than that for lung cancer for men and for breast cancer for women. Epilepsy therefore represents a significant health problem. While about 2/3 of epilepsy patients can be treated successfully with antiepileptic drugs, about 1/3 of patients suffer from medically refractory epilepsy with continuing seizures despite high doses of antiepileptic drugs. The major problems for these patients is the unpredictability of their seizures which usually occur randomly without prior warning resulting in severe restrictions in everyday life, social stigmatization, injuries and eventually death due to SUDEP (sudden unexpected death in epilepsy). For some of these patients epilepsy surgery removing that part of the brain where the seizures originate from represents a highly effective treatment option. Successful epilepsy surgery depends on an accurate presurgical evaluation. One of the cornerstones of presurgical evaluation is intensive video-EEG monitoring where long-term EEG recordings are performed for an average of 5 days in order to record the patients habitual seizures. We developed a novel seizure detection algorithm called EpiScan which allows reliable detection of epileptic seizure with high sensitivity and specificity. Our algorithm was tested both in large volume retrospective and prospective data sets. Our algorithm automatically adapts decision thresholds using a data-driven approach and thus is independent of manual parameter adjustments. Our algorithm allows online detection of seizures with short detection delay within a few seconds. Therefore our algorithm represents the first system applicable in a real world clinical setting. The scientific and clinical implications of our seizure detection system can be summarized as follows: (1) Our algorithm offers new insights in the transition from the interictal to the ictal state and thus in the mechanisms of seizure generation. (2) Our system allows exact assessment of seizure frequency and severity which is important for monitoring treatment response to any therapeutic intervention. Patients subjective reports which currently have to be relied on are rather inaccurate and many seizures occur unnoticed. (3) Our algorithm enhances patients safety during long-term video-EEG monitoring by automatic online alerting of physicians, nurses and EEG technicians. (4) Visual review and analysis of large volumes of EEG data acquired during long-term video-EEG monitoring can be significantly improved and expedited. (5) Our system alerts the physician to potentially clinically relevant EEG segments overlooked or not recognized during initial visual analysis and thus increases the yield of long-term EEG recordings. (6) Our algorithm can be used as a seizure warning device to alert relatives, friends and caregivers and thus could reduce the risk for SUDEP (sudden unexpected death in epilepsy). (7) Finally our system has the potential to be used in therapeutic closed-loop devices for seizure-triggered therapeutic interventions including electrical stimulation or local drug delivery.