Epileptic seizures and sleep states detection using hybrid approaches

Since couple of decades, biomedical signal processing plays an important role to the improve the quality of human life. The recent research in the field of biomedical signal processing is carried out by using important biomedical signals like EEG, EMD, ECG, blood pressure and nasal signals. Thesis is divided into 5 chapters including summary. Chapter 1 gives the introduction about Epilepsy, sleep states, piecewise linear functions and metrics used for performance evaluation. In Chapter 2, we proposed a hybrid approach for seizure detection using EEG signals. We used two piecewise linear models namely Halfwave and Franklin transformation. The reason of preferring the linear models over other models is that linear models are simple, computationally fast, and efficient. The algorithm is tested on 23 different subjects having more than 100 hours long term EEG in the CHB-MIT database in several respects. It showed better performance compared to the state-of-the art methods for seizure detection tested on the given database. In 2017 Bhattacharyya et al. [68] proposed best method for seizure detection based on multichannel, which may not be used for real time applications efficiently, achieved average sensitivity, specificity, and accuracy of 97.91%, 99.57%, 99.41%. Whereas, our proposed method achieved an average sensitivity, specificity, accuracy, false alarm rate and kappa of 99.45%, 99.75%, 99.01%, 0.0039, 0.964 respectively and these results are comparatively good than [68] and other state of the art methods. In Chapter 3, various sleep states are detected by using the extended version of the method proposed in Chapter 2. The aim of the algorithm proposed in Chapter 3 is to detect the various sleep states by combining different biomedical signals like EEG, blood pressure, and nasal signals. Algorithm is tested on MIT-BIH Polysomnographic database having more than 70 hours long term EEG, blood pressure and respiratory (nasal) signals with six different sleep classes. Proposed method shows better performance than state of the art methods. Proposed algorithm achieved an average sensitivity, specificity, accuracy, and false alarm rate of 98.35% and 97.32%, 96.96%, 0.029 respectively for two randomly picked classes, 96.62% and 97.10%, 93.94%, 0.030 for randomly picked any 4 classes, 96.13% and 98.33%, 93.84%, 0.016 for all six classes, which is higher than the existing state of the art methods. To use the algorithm in real time scenario, and to increase the further accuracy of the method proposed in Chapter 3 we proposed a new approach for sleep states in Chapter 4. We used two biomedical signals i.e., blood pressure signal, in time domain and EEG signal, in frequency domain. In time domain, statistical and morphological features are extracted from the blood pressure signal and in frequency domain, a piecewise linear reduction namely Franklin transformation is applied on EEG signal. The Franklin coefficients are used as discriminatory features in frequency domain. The novelty of the proposed method is that we considered two cases, the blood pressure signal by itself, and the combination of it with EEG signal. The motivation behind the first one is that in certain cases, e.g., smart personal mobile devices, only the blood pressure signal is available. In both cases the algorithm is tested on MIT-BIH Polysomnographic database having more than 80 hours long term EEG and blood Pressure signals. In both cases we performed comparison tests with relevant state-of-the-art methods, and our algorithm showed better or equal performance in terms of sensitivity, specificity, accuracy, and false alarm rate. Our proposed algorithm in case of using only blood signal, achieved an average sensitivity, specificity, accuracy, false alarm rate and kappa of 95.45%, 98.27%, 93.78 %, 0.0170, 0.0224 respectively which is good or comparatively equal to the state-of-the-art methods. Whereas, an average sensitivity, specificity, accuracy, false alarm rate and kappa of 99.45%, 99.75%, 99.01%, 0.0039, 0.964 respectively is achieved using blood and EEG signals, which is higher than the existing state of the art methods so far.