A novel ANN-based classification of spike-wave activity in 24-hour EEG recordings in rats using spectrograms: Spike-Wave Discharge Artificial Neural Network (SWAN)

Abstract

Background

Electroencephalographic (EEG) detection of spike-wave discharges (SWDs) is essential for diagnosing absence epilepsy. Automated tools for long-term wearable EEG are needed, but current methods relying on basic signal variability metrics inadequately capture SWD complexity.

New method

We developed the Spike-Wave discharge Artificial Neural Network (SWAN), a shallow ANN classifier analyzing STFT spectrograms. SWAN addresses two critical dimensions of absence epilepsy: 1) spontaneous SWDs in WAG/Rij rats, and 2) drug-induced SWD transformations mediated by alpha2-adrenoreceptor agonists (xylazine, dexmedetomidine).

Results

Trained on baseline EEG from 3 rats and tested on baseline/pharmacological recordings from 4 rats, SWAN achieved high precision (0.96) and sensitivity (0.79) across both conditions. It incorporates a novel "certainty" metric quantifying detection confidence.

Comparison with existing methods

SWAN surpasses amplitude-based variability measures (e.g., standard deviation) by directly evaluating complex spatiotemporal SWD patterns in spectrograms, enabling more reliable detection. Its shallow architecture facilitates mathematical interrogation of SWD features.

Conclusions

SWAN accurately identifies both spontaneous and pharmacologically transformed SWDs in a validated rat model. High precision minimizes over-diagnosis in prolonged recordings, while automation supports unattended monitoring via wearable devices. Future work requires expanded datasets to optimize sensitivity under pharmacological challenge. SWAN provides a robust tool for epilepsy research and therapeutic assessment.