Professor Department of Neurology, Tampere University and Tampere University Hospital Tampere, Finland
Rationale: Nelli™ is a software-based system developed by Neuro Event Labs to aid clinicians in the detection of epileptic events through the selection of relevant epochs based on biomarkers derived from 3D video and audio signals. The purpose of this study is to assess the performance of a medical team using Nelli in their seizure annotation workflow, including the reduction in total viewing duration as compared to full-length EMU recordings.
Methods: A cohort of 90 patients with suspected epilepsy admitted to the Danish Epilepsy Center and Aarhus University Hospital for standard video EEG monitoring are being recruited for simultaneously recording with Nelli during the night. The software was configured for maximum sensitivity, yielding an event for every period of visible motion. A pair of medical experts trained in clinical neurophysiology was tasked to independently (and blinded from both the reference standard and the patients’ medical history) to label the Nelli-detected events containing epileptic seizures. The independent annotations were compared with one another to establish inter-rater agreement. Aftwards, the experts were invited to collaborate to establish consensus on disputed events. All three annotation sets were evaluated against the gold standard provided by the video EEG laboratory. The institution was then provided with the expert consensus set for re-evaluation, and the reference standard was updated accordingly for any seizure events originally missed by the laboratory. Results: These results are preliminary, with 42 out of 90 subjects evaluated. For these 42 patients, the mean monitoring period was 2.6 nights (range: 1-9). The mean nightly duration was 9.5 h (range: 5.8-13.6). Nelli registered a mean of 208 events/night per patient (range: 41.5-614), and the mean cumulative event duration was 1.6 h/night (range: 0.2-6.5). The mean fraction of event time compared to the monitored period was 16% (range: 3-63%).Out of 42 patients, eight adults (aged 17-45), two adolescents (aged 11-12), and four small children (aged 2-5) experienced epileptic motor seizures (focal to bilateral tonic-clonic, tonic, myoclonic, spasms, and unclassified focal motor) during their EMU stay. All seizure epochs from the reference standard were detected by the Nelli software, and the expert workflow yielded a test sensitivity of 13 out of 14 (93%) patients experiencing epileptic seizures during the monitoring. The adult and adolescent group (n=10; 46 seizures) reached a performance of 94% sensitivity, 92% PPV, and a mean inter-rater agreement of 0.76. In small children (n=4; 135 seizures), the experts performed at 61% sensitivity, 71% PPV, and a mean inter-rater agreement of 0.29. Notably, 101 of these 135 subtle motor seizures were confirmed in the re-evaluation phase of the study protocol, meaning that they were not originally registered in the video EEG examination. Conclusions: A workflow using the Nelli system is capable of recognizing epileptic seizures of varied severity, with high sensitivity, and from video only. A test sensitivity >90% suggests that a vast majority of motor seizures can be detected without EEG. In some cases, this workflow even enables detection of subtle seizures not originally registered by the laboratory. Despite a low discard threshold resulting in more false detections, the software pruned 84% (per-patient average) of the recorded video. This study demonstrates that software like Nelli offers additional diagnostic insight to aid clinicians to diagnose and treat epilepsy, while potentially saving a substantial amount of annotation effort. Funding: Please list any funding that was received in support of this abstract.: Neuro Event Labs provided funding for the data collection and machine processing performed in this study, as well as salary for the medical experts performing blinded seizure annotation.