The impact of EEG preprocessing parameters on ultra-low-power seizure detection

Closed-loop neurostimulation is a promising treatment for drug-resistant focal epilepsy. A major challenge is fast and reliable seizure detection via electroencephalography (EEG). While many approaches have been published, they often lack statistical power and practical utility. The use of various EEG preprocessing parameters and performance metrics hampers comparability. Additionally, the critical issue of energy consumption for an application in medical devices is rarely considered. Addressing these points, we present a systematic analysis on the impact of EEG preprocessing parameters on seizure detection performance and energy consumption, using one to four EEG channels. We analyzed in 145 patients with focal epilepsy the impact of different sampling rates, window sizes, digital resolutions and number of EEG channels on seizure detection performance and energy consumption. Focusing on clinically relevant, event-based metrics, we evaluated seizure detection performance of a state-of-the-art convolutional neural network (CNN) via the Seizure Community Open-Source Research Evaluation (SzCORE) framework. Statistical relevance of parameter changes was assessed using linear mixed-effects models. Energy consumption was analyzed using an ultra-low-power microcontroller. Reducing the sampling rate from 256 to 64 Hz led to a decrease in sensitivity (p = .015) and false detections per hour (FD/h; p = .002). Larger window sizes reduced sensitivity between 1 and 8 second windows (p = .033) and FD/h between one second and all other sizes (all p values < .001). Average detection delays increased between one second and 4 and 8 second windows (both p < .01). Lower digital resolutions decreased sensitivity between 16 and 8 bits (p = .007). Compared to four channels, using only one EEG channel resulted in a decrease in sensitivity (p < .001) and an increase in the average detection delay (p = .020), but showed less FD/h (p = .005). CNN energy consumption decreased from 49.15 to 17.26 µJ/s when the sampling rate was reduced from 256 to 64 Hz. Lowering the number of channels from four to one reduced the CNN energy consumption from 79.04 to 31.63 µJ/s. This study provides guidance on choosing EEG preprocessing parameters for innovative developments of closed-loop neurostimulation devices to further advance the treatment of drug-resistant focal epilepsy. A major challenge in closed-loop neurostimulation is fast and reliable seizure detection, ideally with minimal energy consumption. Many published seizure detection approaches are not ready for application and lack statistical power and comparability. The impact of EEG preprocessing parameters was systematically evaluated in 145 patients with focal epilepsy. The chosen window size and number of channels had the strongest impact on seizure detection performance and energy consumption. Our results offer guidance for EEG preprocessing choices to advance the treatment of drug-resistant focal epilepsy.