GA-SVM modeling of multiclass seizure detector in epilepsy analysis system using cloud computing

In this paper, we present an epilepsy analysis system, referred to as EAS, for long-term electroencephalography (EEG) monitoring of patients with epilepsy. In our previous works, a high accuracy seizure detection algorithm had been devised. Six support vector machines (SVMs) had been trained to collaboratively classify EEG data into four types, i.e., normal, spike, sharp wave, and seizure. The EAS had initially extracted a total of 980 features from raw EEG data of patients, and then, for each SVM, it used a naive genetic algorithm (GA) to determine a feature subset of the 980 features. However, the feature subsets still included some low-impact features for the EEG classification, and the training process of the seizure detector was time consuming. In this study, the GA is enhanced to further exclude low-impact features from the feature subsets and MapReduce parallel processing is adopted to speed up the training process. In the experiment, a 363-h clinical EEG records were acquired from 28 participants, 3 of which were normal, and 25 were patients with epilepsy. The experiment results show that average size of the feature subsets is reduced from 133.5 to 92.5 and the overall classification accuracy increases from 88.8 to 90.1 %. The new seizure detector processes a 10-s EEG record within 0.6 s, meaning that it meets the real-time requirement for online EEG monitoring gracefully. When the number of servers increases from 1 to 15, the training time of the detector is reduced from 38.3 to 4.9 h. Our new approach improves the EAS significantly.