Temporal Analysis for Epileptic Seizure Detection by Using Data Mining Approach

Epilepsy is one of the most common neurological disorder characterized by recurrent seizures that negatively affect the life of patients. ElectroEncephaloGram (EEG) signal is the gold standard for seizure detection. The process of identification of seizures on EEG is extremely time-consuming and is manually done by epileptologists by means of visual inspection. This paper describes an automatic seizure detection method based on temporal analysis of EEG signal by applying Data Mining techniques. A software tool, called Training Builder, has been used to analyze the time series data using the sliding window paradigm by extracting numerous features from each temporal window. Nine pairs of temporal parameters L and S, which represent respectively the length and the shift of the sliding window, have been considered; feature selection process for space dimensional reduction, undersampling and oversampling techniques for overcoming the class imbalance problem, Bayesian Networks approach for data modeling have been applied. For each couple L and S, the proposed approach showed excellent numerical results, considering evaluation metrics (True Positive rate, True Negative rate, Precision, AUC). The best trained classifier shows a high percentage of true positives (greater than 98%).