Automated analysis and detection of epileptic seizures in video recordings using artificial intelligence

Abstract

Automated seizure detection promises to aid in the prevention of SUDEP and improve the quality of care by assisting in epilepsy diagnosis and treatment adjustment. In this phase 2 validation, the algorithmic model performance of a contactless, marker-free video-based motor seizure detection system is evaluated against video-electroencephalographic monitoring (VEM) from two Danish epilepsy monitoring units. Detection performance of five categories of motor epileptic seizures (tonic-clonic, hyperkinetic, tonic, unclassified motor, automatisms) and psychogenic non-epileptic seizures (PNES) with a motor behavioral component lasting for >10 seconds were assessed. 230 patients were recruited in the validation set, with 340 motor seizures identified by VEM. The control threshold was evaluated at a range of values to compare the sensitivity and false detection rate (FDR). At optimal thresholds, the performance of seizure groups in terms of sensitivity (CI) and FDR/hr (CI): tonic-clonic- 95.2% (77.3%, 99.2%); 0.084 (0.081, 0.11), hyperkinetic- 92.9% (68.5%, 98.7%); 0.62 (0.57, 0.67), tonic- 78.3% (64.4%, 87.7%); 5.81 (5.46, 6.19), automatism- 87% (74.3%, 93.9%); 3.30 (3.08, 3.54), unclassified motor seizures- 77.7% (70.8%, 83.4%); 4.79 (4.48, 5.12), and PNES- 95.6% (85.2%, 98.8%); 1.70 (1.59, 1.83). A generic threshold recommended for all motor seizures under study asserted 87.8% sensitivity and 6.48 FDR/hr. These results indicate the system’s achievable performance at pre-specified thresholds that is clinically applicable for use as a seizure screening solution in diagnostic workflows.