| dc.description.abstract | This study investigates to improve the stress levels computation and its reliability using multiple physiological signals. In which, stress inducement, physiological signal acquisition, preprocessing, feature extraction, classification, optimization of features
from multiple physiological signals, significant feature estimation, decision boundary optimization, and fusion are the major process. Mental arithmetic task stimulus is used to induce stress on the subjects and sixty healthy subjects with a mean age of 22.5 ± 2.5 years were used. This investigation considered the five physiological signals (electrocardiogram (ECG), heart rate variability (HRV) signal, electromyogram (EMG), galvanic skin response (GSR), and skin temperature (ST)) to measure the effect of stress induced on the subject. The acquired ECG and EMG signals were
preprocessed using wavelet denoising method to remove the noises in the frequency
range of signals and 4th order IIR elliptic filter to remove the noises in GSR and ST
signals. The ectopic beat removal algorithm was used to eliminate the presence of noise
peaks and artifacts in HRV signal. In the feature extraction, ECG and EMG signals
features were computed using discrete wavelet packet transform (DWPT), Lomp-
Scargle (LS) periodogram is used to extract the low and high frequency band's power
spectrum in the short- term HRV signal. The startle detection algorithm was
implemented to extract and analyze the feature related to GSR tonic response, and
finally the skin temperature features were extracted directly in the time domain. The
obtained features classified in to four levels of stress including normal using three
nonlinear classifiers (K nearest neighbor (KNN), probabilistic neural network (PNN),
and support vector machine (SVM)). Average classification rate and F1 score above
50% and 0.5 are considered as the dominant features respectively in this work. Result
indicates that, 20 features as dominant features among the 244 features investigated
over various frequency bands of five physiological signals. The maximum average
classification accuracy of four levels was obtained as 74.20% in mean feature of ECG,
76.69% in third cummulant feature of HRV, 74.67% in mean of EMG, 66.84% in startle
frequency feature of GSR, and 63.63% in mean feature of ST in subject-independent
study. The results also indicate a significant improvement of classification results in the
four class of subject-independent study over the earlier highly subject-dependent
studies. In order to improve the classification rate on stress levels and its reliability,
the optimization of decision boundary based on physiologically significant feature
vectors estimation is required. The variable-order hidden Markov model (HMM) based dynamic Bayesian network (DBN) was constructed to extract the dynamic changes of
each physiological signal feature and capable to identify the significant feature vectors
and decision boundaries corresponding to the different levels of stress. The DBN
networks generalized the three decision boundaries of the 20 different dominant
features processed, and the result shows that the maximum average Bayesian
probability of each boundary is 0.544, 0.61, and 0.75 in all the states with respect to
normal state. Finally, these optimized feature vectors belongs to different boundaries
fused to make the global decision to ensure the reliability. The result shows that, an
excellent agreement of reliability measure with improved classification accuracy while
the significant components only presents in the fusion. | en_US |
