13 1.3.1 PART I
PART I discusses the methods and effects of discriminative features.
Chapter 2 develops an automated algorithm that can reliably detect seizures [31] for short-term EEG recordings. The algorithm also has a low hardware complexity. In the proposed approach, only a single channel EEG signal is analyzed for seizure detection. We first filter the EEG signal by a prediction error filter, also known as a whitening filter, to compute an error signal. A 19th-order prediction error filter (PEF) computes the error signal as the difference between the current input sample and the estimate of it. A window based processing is used with a 2-second sliding window with half overlap. The predictor coefficients are recomputed every one second. A two-level wavelet decomposition of the error signal computes the approximate signal and two detail signals. The total energies in a window of the error signal and the three signals from the wavelet decomposition are extracted in two different ways. The features are input to two types of classifiers: a linear support vector machine (SVM) classifier and an AdaBoost classifier. The performance of each classifier is evaluated and compared against the other.
Chapter 3 proposes a novel frequency-domain model ratio (FDMR) test to determine how these two bands should be selected [95]. Using autoregressive modeling, this paper shows that, if two bands are selected appropriately, then the ratio of band power is amplified for one of the two states. The paper introduces a novel frequencydomain model ratio (FDMR) test to determine how these two bands should be selected. The FDMR computes the ratio of the two model filter transfer functions where the model filters are estimated using different parts of the time-series that correspond to two different states. The ratio implicitly cancels the effect of change of variance of the white noise that is input to the model. Thus, even in a highly non-stationary environment, the ratio feature is able to correctly identify a change of state.
Chapter 4 to Chapter 6 develop algorithms for seizure detection and prediction using spectral power ratios for various datasets [29, 27, 81, 96].
Chapter 4 develops a seizure detection algorithm for long-term fragmented EEG recordings [29]. In the proposed approach, we first compute the spectrogram of the input fragmented EEG signals from three or four electrodes. Spectral powers and spectral ratios are extracted as features. The features are then subjected to feature selection
using classification and regression tree (CART). The selected features are then subjected to a polynomial support vector machine (SVM) classifier with degree of 2. Since all these features can be extracted by performing the fast Fourier transform (FFT) on the signals and the classifier requires low hardware complexity [97], the proposed algorithm can be implemented by the hardware with low complexity and low power consumption.
Chapter 5 develops a patient-specific algorithm that can reliably predict seizures using either one or two electrodes [27] for short-term dataset. The proposed algorithm achieves an overall sensitivity higher than 90% and a false positive (FP) rate less than 0.125 FP/hour. The algorithm also requires a low hardware complexity in extracting features and classification. In the proposed approach, we first compute the spectrogram of the input EEG signals from one or two electrodes. A window based PSD computation is used with a 4-second sliding window with half overlap. Thus, the effective window period is 2 second. Spectral powers and spectral ratios are extracted as features and are input to a classifier. A postprocessing step is used to remove undesired fluctuations of the decision output of the classifier. The feature signals are then subjected to feature selection and classification where two strategies are used. One is the single feature selec- tion and the other is the multi-dimensional feature selection. While a seizure prediction system using a single feature requires low hardware complexity and power consumption, systems using multi-dimensional features achieve a higher prediction reliability. Multi- dimensional features are selected for patients where systems using a single feature can not achieve a predetermined requirement.
Chapter 6 develops a patient-specific algorithm that can reliably predict seizures with high area under curve (AUC) for long-term fragmented EEG recordings [81, 96]. The proposed algorithm compares the performance of different feature sets and different classifiers for different canine or human subject. In the proposed approach, we first extract two sets of features. A window based feature extraction is used, where the window size is 4 second for spectral feature set and is 10 second for the correlation feature set, respectively. The 10-second window for correlation is chosen for an accurate estimate of the correlation coefficient. The first feature set includes spectral powers and spectral ratios. The second feature set includes correlation coefficients between all possible pairs of electrodes. The two feature sets are then subjected to feature selection and classification independently. Three classifiers are used and tested on the selected
15 features, which include AdaBoost, radial basis function kernel support vector machine (RBF-SVM), and artificial neural netwroks (ANN).
1.3.2 PART II
PART II discusses feature selection methods for binary classification and multicalss classification.
Chapter 7 proposes a new feature selection algorithm based on minimum uncertainty and sample elimination (referred as MUSE) [98]. The three-step algorithm first quan- tizes features into bins, ranks the features based on an uncertainty score, selects the feature with the lowest uncertainty score, and then discards samples based on an im- purity metric. The discarded samples are not used for selection of subsequent features. The process is repeated until a stopping criterion is satisfied.
Chapter 8 proposes a new multi-class feature selection criterion based onminimum uncertainty (referred as M3U) [99]. In this chapter, we propose a three-step algorithm that first quantizes features into bins, computes an uncertainty vector for each feature and all sample in each feature, and finally iteratively selects features that achieves the
minimum mean minimum uncertainty (M3U). The proposed iterative feature selection algorithm includes two minimization steps and one expectation step, which include (1) find the minimum uncertainty (MU) score for each feature sample given a feature subset, (2) compute the mean minimum uncertainty score (M2U) for the feature subset, and (3) select the feature that achieves the minimum mean minimum uncertainty score (M3U).