Author(s) :

Shrey Somani.

Volume/Issue :

Abstract :

Canine epilepsy is shown to impact an estimated 0.6-5.7% of domestic dog populations worldwide, with approximately 25-33% of affected animals undergoing drug resistance despite consistent epileptic medication (1). Current wearable devices are limited to post-onset seizure detection and cannot predict seizures before they occur. We have utilized various machine learning pipelines for canine seizure classification using intracranial electroencephalographic (iEEG) recordings from Kaggle American Epilepsy Society Seizure Prediction Challenge datasets. This comprised over 3,038 ten-minute segments across four dogs. Raw EEG signals underwent bandpass filtering (0.5-70 Hz), 60 Hz notch filtering, z-score normalization, and artifact clipping. A 328-dimensional feature vector was extracted per segment, including spectral band power across five frequency bands (delta, theta, alpha, beta, and low-gamma), delta/beta ratio, inter-electrode correlation coefficients, Hjorth parameters, and statistical moments. Specifically, an XGBoost gradient classifier was trained with stratified 5-fold cross-validation and scalebased class imbalance correction. With a held-out test set, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.932, sensitivity of 0.618, specificity of 0.986, and false positive ratio (FPR) of 0.014. These results exceed previously published results on this dataset, including Howbert et al. (2) (AUC 0.890), Nasseri et al. (3) (sensitivity 0.890), and Brinkmann et al. (4) (AUC 0.720). Results have demonstrated that gradient boosting applied to spectral EEG features can be utilized to achieve accurate canine seizure prediction and can establish a computational foundation whose translation to non-invasive wearable EEG hardware represents a critical direction for future research.