Abstract:
Epilepsy is one of the most common and serious neurological disorders that affects a
significant amount of people around the world. It is characterized by sudden occurrence
of massive seizure attack, which is unpredictable in nature. The treatment of epilepsy is
often carried out through continuous monitoring of the patient using electroencephalogram
(EEG) signals. Since the EEG records are generally of long duration
and the number of patients is huge, an automatic system for diagnosis of epilepsy and
detection is necessary. In addition, it may aid in focal drug delivery and generating
alarm through an implantable device.
Various methods are available in the literature for automatic seizure detection from
EEG signals. The most promising performances are reported by those using timefrequency
transform domain techniques. Recently, the empirical mode decomposition
(EMD) has emerged as a simple and effective method for the analysis of time-series
data. Unlike time-frequency transforms, the EMD is data adaptive, not requiring any
basis function or assumption in regard to data linearity and stationarity. This is
particularly important given that the EEG signals are highly nonlinear and nonstationary.
However, limited amount of work is available in the literature that use the
EMD analyzing EEG signals to classify them for epilepsy diagnosis and seizure
detection.
In this thesis, efficient EMD-based methods are developed classification of EEG signal
for subsequent diagnosis of epilepsy and seizure detection. A comprehensive database
of EEG records, publicly available online is used for analyzed using statistical and
chaotic features extracted from the decomposed intrinsic mode functions. The ability of
these features in discriminating the EEG signals is extensively studied. Classification
systems are then developed using the statistical and chaotic features in an artificial
neural network (ANN). The performance of these classification systems is investigated
in terms of sensitivity, specificity and accuracy for various problems of classification
regarding real-life medical scenario of epilepsy diagnosis and detection of seizure
activity. The results show that the features extracted in EMD domain can classify the
EEGs with 100% sensitivity, 100% specificity and 100% accuracy in most of the cases,
while requiring reduced computational cost and fewer features. It is further observed
that the statistical features play the major role in improving the overall performance
compared to the chaotic ones. Finally, an extensive study is conducted to determine
whether statistical priors can capture the underlying statistics of EEG signals.
