Authors :
Presenting Author: Amir Hossein Daraie, BSc – The Johns Hopkins University School of Medicine
Joon kang, MD – The Johns Hopkins University School of Medicine; Sridevi Sarma, PhD – The Johns Hopkins University School of Medicine; Adam Charles, PhD – The Johns Hopkins University; Luis Sanchez, MSc – The Johns Hopkins University School of Medicine; Lynette Talley, BSc – The Johns Hopkins University School of Medicine
Rationale:
Accurate seizure detection is crucial for accurate diagnosis and optimal management of epileptic seizures. Monitoring seizures with long-term video-electroencephalography (EEG) in the epilepsy monitoring unit (EMU) is one of the most useful diagnostic tools in evaluating, diagnosing, and managing epilepsy in a safe and cost-effective way. Currently, clinicians rely on EEG technicians to detect events in real time or the patients to press their alert buttons; but many seizure or seizure-like events are missed by both, leading to suboptimal diagnosis and treatment. There is a need for more accurate and reliable seizure detection methods.Methods:
Locating epileptic activity in long EEG recordings is cumbersome and requires effective automated methods to precisely localize subtle EEG abnormalities. Several studies have been conducted on epileptic seizure detection using various machine learning techniques. However, there are still challenges and gaps in accurate seizure detection. Machine learning methods may be sensitive to overfitting and require a large amount of training data, while the majority of datasets do not encompass a sufficient volume of EEG signals and instead consist of fragmented signals, rendering them unsuitable for real-time EEG signal detection.
In this study, we aimed to leverage intracranial EEG data to develop a new seizure detection algorithm based on a network-based iEEG marker that quantifies region-to-region interactions in the epileptic brain to detect events in real time in an automated fashion. We use the source-sink (SS) metric computed from dynamical network models to quantify each node’s influence to and from other nodes in the network (Gunnarsdottir 2022). Sources are the nodes that continually inhibit a group of neighboring nodes and sinks are inhibited nodes themselves. The SS hypothesis postulates that the absence of clinical seizures is caused by other areas inhibiting the epileptogenic zone (EZ), which are sinks during interictal periods. However, the EZ transitions to sources before seizures and it is this change that our algorithm detects. By using the entropy of SS metric as a measure of the complexity, we track the irregularity of the distribution of sources and sinks in the network as well as patterns leading to seizure events.
Results:
We tested our seizure detection algorithm on continuous EEG data from two patients being monitored at Johns Hopkins Hospital consisting of overall ten days of iEEG data (eight and five days for patient one and two respectively). We observed that our algorithm improved the number of detected seizures by 210% and 120% for patient one and two respectively, while detecting all the annotated seizures. These results were confirmed by an expert EEG technician. Conclusions:
This algorithm will help clinicians make more informed decisions when localizing the EZ thus improving surgical outcomes. The algorithm also provides objective and accurate seizure detection, facilitating more informed decision-making, personalized treatment adjustments, and improved monitoring of patients' seizure patterns.
Funding:
R01NS125897-01 (PI Sridevi Sarma, co-PI Joon Kang)
National Institutes for Health NIH/NINDS
04/01/2022-3/31/2027