Abstract:
Epileptic seizure describes a recurrent abnormal but synchronized surge of electrical
activity in the brain, which is the second common neurological disease. Signal
processing methods try to model visual information into few parameters, which
can be easily detected thus decision making becomes more accurate compared to
the method based on visual observation of EEG. For seizure detection, the features
can be categorized as univariate/bivariate and linear/nonlinear types. But,
the use of composite feature set has been limitedly reported. Since, EEG is a
non-stationary signal and distribution of its energy demonstrates the seizure activities,
time-frequency distribution can perform better than the conventional frequency
analysis methods. E ectiveness of time-frequency based feature extraction depends
on the choice of a kernel and its processing time. Moreover, development of a multifeatured
set capable of detecting and classifying epileptic seizure originated from
di erent parts and state of the brain is still a challenging problem. Prior to feature
extraction, pre-processing involving Hilbert transform is performed. For the preprocessed
EEG signal, time-frequency distributions (TFDs) are obtained by using
twelve Cohen Class kernels capable of reducing in
uence of cross-terms. The TFDs
are examined to nd out nonuniform modules corresponding to dominant bands,
namely ; ; ; and
components and to form a feature set containing modular
cumulative energy at percentile frequencies and modular entropy. This feature set
when fed to each of the decision tree, KNN and ANN classi er, can produce greater
detection accuracy independent of the kernels and o ers lesser processing time. But,
it is unable to classify epileptic seizure originated from di erent parts and state of
the brain. Therefore, a high number of uniform modules are formed to compute a
mutifeatured set containing modular energy and entropy which is found e ective in
detecting as well as classifying epileptic seizure originated from ve di erent parts
and state of the brain. Simulations are carried out using standard EEG dataset to
evaluate the performance of the proposed method in terms of selectivity, sensitivity
and accuracy. It is shown that the proposed method outperforms a state-of-the-art
method with superior e cacy.
