Abstracts

Rationale: Timely caregiver intervention is key for mitigating seizure-related adverse events. We are developing a real-time seizure detection system that alerts caregivers to allow prompt intervention. Seizures are detected based on elevated autonomic nervous system activity which is reflected in heart and respiratory measures captured by an unobtrusive, wearable sensor. Methods: Data were collected on 50 children undergoing video EEG evaluation at Children's National Medical Center (CNMC) by attaching a Zephyr BioPatch sensor to the chest to continuously record electrocardiogram (ECG) and acceleration. Novel methods were developed for accurate extraction of cardiac metrics in real time from mobile subjects. Heart and accelerometer data were condensed into multiple parameters that differentiate seizure and non-seizure activity. A detection algorithm was developed based on a weighted combination of these parameters. The algorithm was implemented in a compact unit which can be clipped to a belt or placed on a nearby table. A microcontroller performs algorithm computations on data streamed from a Bluetooth-enabled ECG sensor to enable real-time seizure detection and alert. Results: Approximately 60% of patients had a seizure during the monitoring period. Seizures without any clinical response or those lasting less than 10 sec (e.g. single myoclonic jerks or clusters) were excluded. Additionally, subjects with multiple seizures per hour were excluded because the autonomic signals often did not return to baseline and this seizure frequency is outside of the intended use of the monitor. After exclusions, the algorithm was evaluated on 9 children (4F, 5M, mean age 11.9 years). For additional validation, the algorithm was also evaluated with the MIT Physionet database with ECG from 5 adults with partial seizures (Neurology 53:1590-2, 1999). The algorithm without motion parameters detected all seizures classified as tonic-clonic (3/3) or atonic/clonic/tonic (4/4), and 3/7 and 7/10 of partial seizures from the CNMC and MIT subjects respectively. In the CNMC dataset, false positives averaged 1 per 14 hours, however, the majority of false positives occurred in a few patients with poor sensor data quality. Over half of the subjects did not have any false positives. One false positive occurred in the 16.8 hours of MIT data. The addition of motion parameters did not increase the overall detection success rate. All of the tonic-clonic seizures, 2/4 in the atonic/clonic/tonic category, and 1/7 partials were detected with the motion algorithm. A false positive rate of only 1 per 46 hours for the motion algorithm suggests the addition of motion parameters for subjects known to have distinctive seizure movements may reduce the overall number of false positives. Conclusions: Our multi-parametric approach to seizure detection shows promise for accurate detection of a range of seizure types in real time. Alternative approaches are limited to detection of seizures with repetitive motion. Because the detection algorithm utilizes a weighted combination of many parameters, the number of false alarms is minimized. Funding: NIH Award #R01EB014742