Abstracts

Rationale: It is important to know the frequency of seizures that patients experience in their home environment. Existing devices are useful mainly for identifying tonic-clonic (TC) seizures and have limited ability to detect other seizure types. Nelli is a remote camera-based device employing machine learning that is designed to identify seizures with a positive motor component, including but not limited to focal or generalized TC seizures, clonic, tonic, hyperkinetic seizures, automatisms, and myoclonic seizures. We prospectively tested the device’s ability to detect seizure events in 116 subjects in the epilepsy monitoring unit (EMU) at Thomas Jefferson University Hospital.

Methods: Nelli’s cameras were mounted in 8 rooms in the EMU. Subjects admitted to the EMU for routine seizure characterization or pre-surgical evaluation were eligible for enrollment. Following informed consent, Nelli cameras and conventional video-EEG data were collected simultaneously. Video-EEG was evaluated per routine clinical care and the category and time of seizures were documented. Nelli’s machine models analyzed video during the same periods, evaluating three custom classes designed to quantify the level of motor involvement of seizures. Video-EEG and Nelli results were compared to determine the sensitivity and specificity of the three hierarchical seizure detection models’ ability to identify events as compared to the clinical gold standard.

Results: By conventional video-EEG review criteria, 18 subjects experienced 44 tonic-clonic seizures and 14 subjects experienced 54 other seizures with prominent motor behaviors (clonic, tonic, hyperkinetic, or repetitive myoclonic) during simultaneous monitoring. The Nelli device recorded 14,098.5 hours of video. 1,126.4 hours of data from 5 subjects (including 11 tonic-clonic seizures and 5 seizures with prominent motor behaviors) were excluded due to visual artifacts caused by fluorescent lighting. Overall, Nelli detected seizures with a positive motor component with a sensitivity of 90.0%, 95% CI [84.3%, 94.0%]. Nelli detected tonic-clonic seizures with a sensitivity of 97.0%, 95% CI [84.7%, 99.5%] and a false detection rate (FDR) of .051/h; other prominent motor seizures were detected with a sensitivity of 91.7%, 95% CI [80.4%, 96.7%] and a FDR of 2.92/h.

In addition to the seizure types above, 24 subjects experienced 88 seizures with automatisms or other more subtle motor behaviors. Nelli detected at least one of these typically subtle events in all but 3 subjects. These seizures were identified by the system with a sensitivity of 86.4%, 95% CI [76.2%, 92.6%] and a FDR of 19.4/h.

Conclusions: The machine learning algorithm showed excellent sensitivity in detecting seizures with a variety of motor behaviors. The ability to detect seizures other than tonic-clonic seizures represents a significant advance in seizure detection, and demonstrates broad utility for detecting seizures in a home or hospital setting.

Funding: Please list any funding that was received in support of this abstract.: Funding for this study was provided by Neuro Event Labs.