Patient-specific seizure detection in long-term EEG using wavelet decomposition

This paper presents a novel patient-specific seizure detection approach using wavelet decomposition of multi-channel EEG data and hand-engineered features extracted from the decomposed data. EEG data of all channels of each patient are segmented into four second segment lengths, and these segments are decomposed using discrete wavelet transform into four frequency bands corresponding to the δ, θ, α and β EEG rhythms. Three features are then extracted from each of these bands, which are used to classify the seizure and non-seizure segments. The proposed approach does not require any feature processing, or any post-processing for obtaining the seizure detection results. The CHB-MIT database with data of 23 pediatric patients is used for validation of the proposed seizure detection approach using five classifiers, and accuracy, sensitivity and specificity values of 99.6%, 99.8% and 99.6% respectively, averaged over all 23 patients, are obtained using five-fold cross-validation method. The obtained seizure detection results are compared with the results of other studies using the same database, and shown to out-perform the state-of-the-art. Furthermore, the obtained results are consistent over the data of all the patients, thereby demonstrating the robustness of the approach. The computational efficiency of the proposed approach, which is another highlight of the approach, is also illustrated in the form of metrics. The suitability of the approach for seizure detection in long-term multi-channel EEG recordings is also discussed.