top of page

Exploring a New Frontier in Electroencephalographic (EEG) Spike Detection: A Controlled Synthetic Comparison

William Zhang
31/10/2026

Reliable spike detection requires identifying brief events without repeatedly mistaking noise for spikes. This EEG-motivated proof of concept compared a compact convolutional neural network (CNN) with methods derived from the Smoothed Teager Energy Operator, Bayes optimal template matching (BOTM), and Multi-Template Matching. The signals were simplified mathematical waveforms rather than clinical EEG. The dataset contained 230 parent scenes, each created at -3 dB and 0 dB, yielding 460 arrays divided into training, validation, calibration, and held-out testing sets. All four methods received the same 100 test arrays from 50 parents, including 10 parents without spikes. Detections were matched one-to-one to known spike times within ±15 ms. Across the 40 spike-containing scenes per condition, CNN mean F1 was 95.90% at -3 dB and 96.53% at 0 dB. BOTM-derived achieved 99.22% and 100.00%, respectively. A separate extension tested three independently trained residual CNNs on 500 new parents (1,000 arrays), with each model fixed before testing. They averaged 51.46% fewer false detections than the BOTM-derived method, with a pooled recall of 99.486% versus 99.952%. However, the CNN and BOTM thresholds were chosen using different objectives, and the residual CNN missed more events. The primary comparison favored BOTM-derived by mean F1. The extension showed a promising reduction in false detections, but further work must examine whether this can be retained without losing recall and whether it carries over to real EEG.

bottom of page