Automated markers for enhanced epilepsy diagnosis, treatment, and safety monitoring

Abstract

In specialized epilepsy care, data streams are often monitored and analyzed real-time by a human observer to detect events. Knowledge about epilepsy-related events can aid diagnosis, direct treatment, and indicate the need for immediate assistance. The visual observation usually performed to detect these events is time-consuming, subjective, and sensitive to distractions. Consequently, important events could be missed. Automated markers can help detect occurrences and characteristics of events, and may identify data streams likely to contain events. The aim of this thesis was to improve situations of data monitoring for event detection in epilepsy, by constructing and validating automated algorithms to detect markers of epilepsy. In the epilepsy monitoring unit, seizures need to be detected by staff so they can go to the patient to reduce risks arising from seizures (e.g. treat injuries) and to perform diagnostic testing. Online seizure detection algorithms might help detect seizures that were otherwise missed or recognized too late. We investigated the added value of applying algorithms that detect seizures in the electroencephalogram (EEG). The algorithms may improve staff response to seizures, since they increase the total number of seizures detected and the speed of detection. Invasive presurgical or intraoperative EEGs can be monitored for events that help delineate brain tissue that needs to be resected, in order to prevent seizures. An automated algorithm is needed that can delineate epileptogenic tissue during surgery. We presented a novel algorithm - autoregressive model residual variation (ARR) - for this purpose. High ARR values measured in the intraoperative post-resection EEG were associated with poor postsurgical outcome. We concluded that the ARR algorithm might enable intraoperative epileptogenic tissue delineation, to optimize the chance of seizure freedom. In homes of people with epilepsy, monitoring for (results of) dangerous seizures can improve safety by indicating whether someone is in need of assistance and at risk of sudden death. Real-time seizure detection systems can help alert caregivers, but wearable sensors are not always tolerated. We aimed to design a remote monitoring system to detect convulsive seizures, falls, and apneas in video registrations. First, we established performance of a convulsive seizure detection algorithm. The algorithm detected all convulsive seizures with acceptable false detection rates, and could therefore improve safety. Next, we presented a remote fall detection algorithm. The algorithm showed high sensitivity and specificity in datasets with acted and seizure-related falls, reflecting its feasibility of detecting falls. We also developed and tested an algorithm to detect apneas, using video registrations of simulated and real-life seizure-related apneas. All apnea episodes were detected from at least one camera angle, with high specificity. Integrating camera inputs capturing different angles increased sensitivity. These results show feasibility of detecting apneas using the proposed algorithm. The work in this thesis contributed new algorithms to detect tissue to be removed during epilepsy surgery, and to remotely detect falls and apneas. It also adds knowledge about the validity and added value of seizure detection algorithms. We conclude that automated markers may enhance epilepsy diagnosis, treatment, and safety monitoring.