Discrimination of scalp EEG signals in wavelet transform domain and channel selection for the patient-invariant seizure detection

Abstract

In this paper, a statistical method of classifying Electroencephalogram(EEG) data for automatic detection of epileptic seizure is carried out using a publicly available scalp EEG database. The classification is carried out to distinguish the seizure segments from the non-seizure ones. The higher order moments (specifically variance) have been calculated in various sub-bands in the wavelet transform domain and utilized as the discriminating feature in the Support Vector Machine(SVM) classifier. The method is tested on 175 hours of continuous EEG data from five patients and on an average, 99% accuracy has been achieved with very high values of sensitivity and specificity. Furthermore, on the basis of the figure of merits, for their excellent performance for all the patients, seven channels have been selected for the patient-invariant seizure detection which might help the electroencephalographers reducing their laborious job of monitoring the EEG data from all the channels.

Epileptic Seizure Detection With a Reduced Montage: A Way Forward for Ambulatory EEG Devices

“…On comparing these results with those reported in [51], in case of reduced montage of 10 channels, our algorithm detects all the seizures in these 6 patients correctly where each patient has FP/hr < 0.2. Also, our proposed method shows comparable (or better) performance relative to the performance of methods such as [47]- [49] when the number of channels and the selected patients are almost the same. Most importantly, there is no significant loss in performance caused by spatial restriction and reduction in number of electrodes.…”

“…For reduced number of channels, 75 different montages involving 3, 4, 5 or 6 channels were successively adopted, and the success rate for the best combination of channels (maximum up to 6 channels) reached to 85%. Some recent works [47]- [49] also considered the topic of channel selection and presented the performance results using CHB-MIT dataset. A summary of the performance comparison of these methods with the proposed method is given in Table 6.…”

“…A summary of the performance comparison of these methods with the proposed method is given in Table 6. It should be noted that in the prior work on the topic of EEG channel selection, including [47]- [51], the method of reduction in the number of channels is proposed so that either the best detection accuracy is obtained or sensitivity is maximized for a fixed false positive rate subject to the constraint on a fixed number of maximum channels. In contrast to these earlier works, this paper considers the reduction of channels focused on a specific region of the scalp (temporal region) so that the feasibility of an ambulatory wireless EEG device with prolonged battery life could be evaluated for long-term recording and detection of epileptic seizures.…”

Patient-dependent epilepsy seizure detection using random forest classification over one-dimension transformed EEG data

“…A rst distinction between the approaches is the specic domain of the EEG data on which they have focused: Time series, frequency data, or spectro-temporal signals (Ahmad et al, 2015;Das et al, 2015). Another signicant dierence is the set of features which each method obtained in its analysis: Some studies utilized statistical, chaos theory, or information theory parameters (Ahmad et al, 2015;Das et al, 2015;Gill et al, 2015), while other eorts applied data transformations such as singular value decomposition (SVD) or principal components analysis (PCA) (Zhao et al, 2016). A third prominent divergence among the previously proposed methods is the level of articial intelligence dependency to perform a decision.…”