80 Recommendation (2)

ANALYSIS OF ELECTRODERMAL ACTIVITY
SIGNALS IN AROUSAL AND VALENCE
EMOTIONAL STATES AND ITS VALIDATION
WITH EEG SOURCE LOCALIZATION
A THESIS
submitted by
G. NAGARAJAN
for the award of the degree
of
DOCTOR OF PHILOSOPHY
DEPARTMENT OF APPLIED MECHANICS

INDIAN INSTITUTE OF TECHNOLOGY MADRAS
CHENNAI
SEPTEMBER 2019
CHAPTER 5
SUMMARY AND CONCLUSION
Emotional changes influence the cognitive performance, social interaction, and decision-
making capability of a human being. The effect of emotion abnormalities such as
depression and posttraumatic stress may lead to reduce the quality of life. The study of
human emotions may provide relevant information associated with various emotional
states. This information is also essential in clinical studies, developmental disorders, and
healthcare.
The emotions are described as a combination of two basic affective dimensions namely,
valence and arousal using CMA method. The valence dimension represents the degree of
unpleasant/pleasant stimuli that is perceived by the user. On the other hand, the arousal
dimension indicates the levels of emotional intensity experienced by an individual.
Characterization of arousal and valence state covers range of spontaneous and every day
emotions.
Commonly, emotion is analyzed using non-physiological data and physiological signals.
Non-physiological data includes facial expressions, voices and body gestures. However,
the emotional responses obtained from these methods may not reflect the true emotional
state of the individuals. Further, physiological signals such as EEG, ECG, HRV, and
EMG are analyzed for emotion recognition. Among the physiological signals, EDA is
one of the widely preferred technique used in emotion classification.
EDA signal is a non-intrusive technique that records the variation in skin conductance.
The signals obtained for various emotions are complex and non-stationary in nature.
112
Conventional time and frequency domain features are unable to fully characterize and
differentiate various emotional states in EDA signal. Time-frequency domain based
features provide better performance; however, it does not account for intra-subject
variability. CNN based feature extraction provides more significant and robust features to
characterize complex EDA signals.
EEG signals of the cortical region have been used to determine different emotional states.
The combination of EEG and EDA signals were proposed to classify various audio
stimuli. Although significant contribution has been made towards EEG based emotion
recognition, EEG signal acquisition involves complex instrumentation and it requires
brain source localized electrode system for better signal acquisition. On the other hand,
EDA signals are reliable, low-cost, and non-intrusive. As an index of body’s sympathetic
nervous system, monitoring of EDA signals is equivalently reliable with EEG.
In this work, EDA signals are analyzed to characterize arousal and valance states of
emotion after validation with simultaneous EEG recordings. The publicly available
DEAP database that contains EDA signals recorded using various audio-video stimuli is
considered for this study. These signals are normalized and decomposed into the phasic
and tonic components. The phasic component is applied to STFT based time-frequency
analysis. Time, frequency, and time-frequency domain features are computed from the
phasic component and time-frequency spectrum. These features are subjected to CNN
based classification model to learn robust key features. Further, these key features are fed
to four different classifiers, and their performances are compared against all features.
The result shows that the tonic and phasic components of the decomposed EDA signal is
distinct for arousal and valance. The inter-subject variations are found to be higher in
valance than in arousal. Among the considered time, frequency, time-frequency features,
113
time-frequency features are found to be effective in differentiating valance and arousal
states. The signal characteristics obtained using CNN learned features could characterize
variations in arousal and valence states of emotions. The proposed feature learning
approach is found to have the least inter-subject variability. Thus, the features extracted
from the signals are found to be the biomarkers of emotions.
The changes in EDA signals are correlated well with the source-localized EEG signal
power, and it is in agreement with the existing literature, and thereby, it provides
appropriate validation of EDA features. Among the active brain sites, the EEG response
in the frontal region correlates well with the EDA variations, and it is observed in the
corresponding electrodes, namely, FP2 and FP1. Similar results were observed with the
response of the C4 electrode, which records the activity of the central region.
After the validation of EDA signals, advanced machine learning techniques are employed
for the classification of arousal and valence states. The signals are subjected to multiscale
CNN for automated classification. Multiscale CNN is effective in extracting and learning
complimentary features at different time scales. The results demonstrated that the
extracted features can classify different emotional states in EDA signals. The
performance of multiscale CNN is found to be better than simple CNN for emotional
state classification. Thus, multiscale CNN is found to be useful for the discrimination of
emotion states in EDA signal. Moreover, the proposed approach appears to be effective
for the development of non-invasive emotion recognition device.
114
The salient contributions of this research work are as follows:
• State of emotions have been comprehensively analysed and validated using EEG
source localization,
• Two different anatomical sites namely Fp2 and FC6 with high activation in theta and
beta3 bands have been identified for various emotions, and
• Multiscale CNN based classification is used for the automated differentiation of
arousal and valence states.
115
• CHAPTER 6
SCOPE FOR FUTURE WORK
The analysis performed in this study can also be tried out to understand other
psychological disorders such as autism, and bipolar disorders.
The framework adopted in this work could well be extended to similar complex
biomedical signals.
The proposed framework could well be extended to neurological disorders such as
seizure for better source localization
116

80 Recommendation (2)

Uploaded by

Copyright:

Available Formats

80 Recommendation (2)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

80 Recommendation (2)

Uploaded by

Copyright:

Available Formats

ANALYSIS OF ELECTRODERMAL ACTIVITY

SIGNALS IN AROUSAL AND VALENCE

EMOTIONAL STATES AND ITS VALIDATION

WITH EEG SOURCE LOCALIZATION

for the award of the degree

DEPARTMENT OF APPLIED MECHANICS

SUMMARY AND CONCLUSION

making capability of a human being. The effect of emotion abnormalities such as

dimension indicates the levels of emotional intensity experienced by an individual.

Commonly, emotion is analyzed using non-physiological data and physiological signals.

one of the widely preferred technique used in emotion classification.

differentiate various emotional states in EDA signal. Time-frequency domain based

characterize complex EDA signals.

recognition, EEG signal acquisition involves complex instrumentation and it requires

nervous system, monitoring of EDA signals is equivalently reliable with EEG.

from the signals are found to be the biomarkers of emotions.

for the development of non-invasive emotion recognition device.

beta3 bands have been identified for various emotions, and

• Multiscale CNN based classification is used for the automated differentiation of

arousal and valence states.

SCOPE FOR FUTURE WORK

psychological disorders such as autism, and bipolar disorders.

The proposed framework could well be extended to neurological disorders such as

seizure for better source localization

You might also like