Applicability of quantitative seizure movement analysis technology in a Portuguese Epilepsy Monitoring Unit

Abstract

Introduction: One third of epilepsies are drug-resistant and potential candidates to surgical treatment. Seizure semiology is essential for determining the epileptogenic zone but it relies on subjective analysis. Quantitative motion analysis during motor seizures can be a more accurate indicator and it has been successfully developed and laboriously implemented in a specific Epilepsy Monitoring Unit (EMU). This study aims to evaluate the feasibility of integrating quantitative seizure movement analysis technology into the routine of a Portuguese EMU, by determining DeepInBed system's ability to extract motion quantification metrics from standard 2D video recordings, its effectiveness in distinguishing between automotor and hypermotor seizures, and the ease of use of seizure analysis software by clinicians. Methods: A single-center, cross-sectional study was conducted with patients with automotor or hypermotor seizures that underwent non-invasive Video-EEG monitoring between January 2019 and February 2024 in an EMU at a Portuguese refractory epilepsy center not originally involved in the development of the software. Video recordings utilized 1080p HD cameras positioned perpendicularly to the patients' bed. Distinct marking points were established for seizure onset, seizure end and for the beginning and ending of automotor or hypermotor semiology. DeepInBed, based on pre-trained vision transformer pose estimation and Python programming, was employed to analyze motion patterns. Movement extent and maximum and mean velocity metrics from 10 automotor and 10 hypermotor seizures reviewed and classified by two experienced epileptologists were assessed to determine their capability to distinguish between the two types of seizures. To assess if Kinect Seizure Analyzer (KiSA), a software for quantitative movement analysis, is clinician-friendly and easy to use, two experimental tests were performed and qualitatively analyzed. Results: Twenty-three automotor (seven patients) and twelve hypermotor (five patients) seizures were included. Aggregated metrics from 17 joints showed a higher sum of displacements (hypermotor=17758 pixels vs automotor=13064 pixels, p value=0.001), movement extent (hypermotor=832162 pixels2 vs automotor=499955 pixels2, p value<0.001), mean velocity (hypermotor=627 pixels/s vs automotor=336 pixels/s, p value<0.001), mean acceleration (hypermotor=7137 pixels/s2 vs automotor=3969 pixels/s2, p value<0.001) and mean jerk (hypermotor=129287 pixels/s3 vs automotor=77870 pixels/s3, p value<0.001), in hypermotor than in automotor seizures with a statistically significant difference. Separate analysis of movement extent and maximum and mean velocity of the trunk and the most active wrist key points in 10 automotor and 10 hypermotor seizures demonstrated no statistically significant difference between these metrics, preventing us from yielding distinct cutoff parameters. The experimental tests using KiSA showed that although useful and able to perform semi-automatic motion tracking, the program is time consuming for non-experts and still has a few glitches. Conclusion: DeepInBed quantitative seizure movement analysis is capable of extracting objective motion metrics from 2D videos routinely recorded in a Portuguese EMU, making it a promising tool in pre-surgical evaluation. Even with a low number of seizures, objective metrics seemed to differ between hypermotor and automotor seizures. We could not, however, find a reliable cutoff value for wrist and trunk, which are usually the most clearly seen key points during seizures, to distinguish between the two seizure types, limiting its application in clinical practice. Additionally, seizure analysis software is useful, but needs to be improved to become clinician-friendly for everyday practice.