A hybrid feature extraction and machine learning approaches for epileptic seizure detection

Epileptic seizure detection from the brain EEG signals is an essential step for speeding up the diagnosis that assists researchers and medical professionals. For this, various classification signal processing techniques have been developed in the traditional works. Still, they limit with the problems of increased complexity, reduced performance and insufficient classification rate. This motivates to design an automatic system for classifying the normal and abnormal EEG signals. Thus, an efficient machine learning approaches are implemented in this work, to overcome the existing techniques limitations. Here, an enhanced curvelet transform technique is established in order to overcome the disadvantage of Gabor and Wavelet transformations data loss and indiscriminate orientations. This method has the capacity to furnish the all the signals data required with no information loss of shearlet transformation and hence implemented to preprocess the given EEG signal, which smoothen the signal by eliminating the noise. Then, a modified graph theory, fractal dimension and novel pattern transformation techniques are employed to extract the features and patterns. The extraction of features is crucial for classification and compression of huge volume of EEG signal that possess low information. This theory improves the precision and speed of the computational technique. Most of the current research, Graph theory is reflected in the area of quantitative description of the time series data. It is predominantly employed for the reduction of noise and not affected during choosing the factors. From the patterns, the statistical features are extracted by using a standard gray level co-occurrence matrix technique that comprises entropy, homogeneity, energy, correlation and maximum probability. This method calculates the linear dependency of the adjacent samples thereby effective measurement of information loss in the transmitted signal is accomplished. Then, these extracted features are fed to the classifier named as novel random forest classification for detecting and classifying the signal as healthy, ictal and interictal. During simulation, various performance measures have been used for evaluating the results of existing and proposed classification techniques and results validate the efficacy of proposed technique.