Abstracts

Rationale: Hypermotor or hyperkinetic seizures typically have a frontal lobe origin. They occur more commonly in children than in adults and often occur at night when supervision and care may be reduced. Motor manifestations during these seizures may be violent and can result in serious consequences such as injuries due to uncontrolled movements, dizziness, headaches, and even death. Detection of these seizures is challenging because the manifestation on scalp EEG may be subtle or absent altogether. Moreover, the clinical manifestations of hypermotor seizures may resemble those of non-epileptic parasomnias. In this context, extra-cerebral detection of hypermotor seizures using accelerometry may constitute an alternate solution. The objective of this study is to assess the performance of seizure detection using limb-based accelerometers (ACMs) in pediatric patients (Ages 5-15 years) with hypermotor seizures. Methods: Simultaneous acquisition of video-EEG and 12 channels of accelerometer (ACM) during hypermotor seizures was performed in 7 patients. Four 3-dimensional accelerometers were used, placed respectively on the left arm (LA), right arm (RA), left leg (LL) and right leg (RL). The raw ACM data was calibrated and fed into a feature extraction block. Time and frequency domain features were extracted. The seizure detection thresholds on these features were determined using a leave-one-out cross-validation technique after which the seizure detection decision was made. The seizure detection algorithm was run prospectively using data from each ACM placed in LA, RA, LL and RL location separately as well as taking all the ACM data from all limbs together. A detection window of ±2 minutes from the annotated seizure onset was used to determine True and False Positive detections. The sensitivity and false positive rate of seizure detection were then determined. Results: Results from prospective application of this algorithm to 7 patients (53 seizures) show a high sensitivity (mean sensitivity: 97.02%) of seizure detection with a false positive rate of 2.1 detections/hour. The mean latency of seizure detection across all patients was 2.85 seconds. Such short latencies to detect convulsive seizures are critical to aid in enabling fast responses to detected seizures. The use of additional ACM sensors on more limbs does not appear to improve the false positive rate (mean FPR for left arm: 2.13/hr vs. Mean FPR for both arms and legs: 2.72/hour). Conclusions: The performance of real-time, prospective seizure detection using limb-based ACMs has a high sensitivity (mean Se: 97.02%) of detecting hypermotor seizures. However, the false positive rate (mean FPR: 2.13/hour), although encouraging, may not be adequate for chronic long-term seizure monitoring, especially during the night time. Further improvements to the performance of seizure detection can be achieved via use of multi-modal sensors, placement of sensors in body locations other than the limbs, and patient-specific tuning of algorithm parameters over time.