Authors :
Presenting Author: Andrew Knight, MSc – Tampere University
Pragya Rai, PhD – Neuro Event Labs
Matias Hiillos, BSc – Neuro Event Labs
Csaba Kertész, PhD – Neuro Event Labs
Kaapo Annala, MSc – Neuro Event Labs
Jukka Peltola, PhD, MD – Tampere University
Michael Sperling, MD – Thomas Jefferson University, Philadelphia, Pennsylvania, USA.
Rationale:
Video-based seizure detection systems, particularly those leveraging AI, hold promise to boost both patient safety (through the use of real-time notifications) and diagnostic capability (through the improved detection efficiency) for patients at risk of seizure, both in-hospital and at-home. This study builds upon our previously described method (from Nelli®, a marker-free video-based seizure monitoring system), refining the technique and validating it against a large, independent dataset (N=224) of gold-standard labeled recordings to achieve improved performance goals.Methods:
A dataset of continuously-recorded near-infrared (grayscale) 3D video from RealSense D435 sensors was collected from consented adult subjects (aged 19-85) undergoing routine video-EEG at Thomas Jefferson University Hospital EMU. The dataset documents 81 convulsive (generalized or focal-to-bilateral tonic-clonic) seizures and 171 other prominent motor seizures with a behavioral component lasting >10 seconds. While less-prominent seizures were also labeled, they were excluded from analysis.
Our previously-described method [0] has been iterated to improve both accuracy and computational performance. By boosting the optical flow-based signals with depth information and employing GPU-accelerated primitives, an improved model was developed for real-time use. The model utilizes decision-based machine learning trained on approximately 200 hours of various motor seizure types (tonic-clonic, clonic, hyperkinetic, tonic, and other prominent motor seizures). Seizure probability is ranked every second, using a 3-second sliding window of signal history. The system aggregates continuous segments of movement (up to 5 minutes) and applies the maximum rating within each segment to categorize events as either “low” (not likely to contain a motor seizure), “medium” (likely to contain a prominent motor seizure), or “high” (likely to contain a convulsive seizure).
[0] Rai P et al. Automated analysis and detection of epileptic seizures in video recordings using artificial intelligence. Front Neuroinform. 2024 Mar 15;18:1324981.
Results:
The clinical performance assessment for the updated system is ongoing. Acceptance criteria have been set to 90% assumed sensitivity for both convulsive and prominent motor seizures, with a false detection rate of less than 1 per 24-hour period for convulsive seizures (“high” events) and less than 1 per hour for other prominent motor seizures (“medium” events). Events marked as “low” represent general scene movement and are not assessed for seizure detection accuracy.Conclusions:
By raising the performance expectations compared to the previous generation of the model, modifying its architecture to operate in real-time, and validating on a large unseen EMU dataset, this study aims to provide substantial evidence of an improved video-only seizure detection system. Sensitivity >90% and FDR of < 1 per 24 hours ensures that convulsive seizures can be detected with acceptable performance for safety monitoring. For other prominent motor seizures, sensitivity >90% and FDR of < 1 per hour provide a suitable aid for diagnostic procedures such as seizure counting.Funding: Neuro Event Labs sponsored this study.