Patient-aware adaptive ngram-based algorithm for epileptic seizure prediction using EEG signals

This work proposes a novel patient-aware approach that utilizes an n-gram based pattern recognition algorithm to analyze scalp electroencephalogram (EEG) data and predict epileptic seizures. The method addresses the major challenge of extracting distinctive features from EEG signals through a detection of spatio-temporal signatures related to neurological events. By counting the number of occurrences of amplitude patterns with predefined lengths, the algorithm generates a probabilistic measure (anomalies ratio) that is used as a prediction marker. These extracted ratios are classified using state of the art machine learning algorithms into seizure and non-seizure windows. The efficacy of the prediction model is tested on patient records from the Freiburg database with more than 100 hours of recordings per patient and for a total of 145 seizures. The proposed algorithm is further optimized to obtain the n-gram parameters for enhanced feature extraction. Results demonstrate an average accuracy of 93.83%, sensitivity of 96.12%, and false alarm rate of 8.44%.