Year of Publication
Doctor of Philosophy (PhD)
Electrical and Computer Engineering
Dr. Himanshu Thapliyal
The advent of IoT has enabled the design of connected and integrated smart health monitoring systems. These health monitoring systems can be utilized for monitoring the mental and physical wellbeing of a person. Stress, anxiety, and hypertension are the major elements responsible for the plethora of physical and mental illnesses. In this context, the older population demands special attention because of the several age-related complications that exacerbate the effects of stress, anxiety, and hypertension. Monitoring stress, anxiety, and blood pressure regularly can prevent long-term damage by initiating necessary intervention or clinical treatment beforehand. This will improve the quality of life and reduce the burden on caregivers and the cost of healthcare. Therefore, this thesis explores novel technological solutions for real-time monitoring of stress, anxiety, and blood pressure using unobtrusive wearable sensors and machine learning techniques.
The first contribution of this thesis is the experimental data collection of 50 healthy older adults, based on which, the works on stress detection and anxiety detection have been developed. The data collection procedure lasted for more than a year. We have collected physiological signals, salivary cortisol, and self-reported questionnaire feedback during the study. Salivary cortisol is an established clinical biomarker for physiological stress. Hence, a stress detection model that is trained to distinguish between the stressed and not-stressed states as indicated by the increase in cortisol level has the potential to facilitate clinical level diagnosis of stress from the comfort of their own home.
The second contribution of the thesis is the development of a stress detection model based on fingertip sensors. We have extracted features from Electrodermal Activity (EDA) and Blood Volume Pulse (BVP) signals obtained from fingertip EDA and Photoplethysmogram (PPG) sensors to train machine learning algorithms for distinguishing between stressed and not-stressed states. We have evaluated the performance of four traditional machine learning algorithms and one deep-learning-based Long Short-Term Memory (LSTM) classifier. Results and analysis showed that the proposed LSTM classifier performed equally well as the traditional machine learning models.
The third contribution of the thesis is to evaluate an integrated system of wrist-worn sensors for stress detection. We have evaluated four signal streams, EDA, BVP, Inter-Beat Interval (IBI), and Skin Temperature (ST) signals from EDA, PPG, and ST sensors. A random forest classifier was used for distinguishing between the stressed and not-stressed states. Results and analysis showed that incorporating features from different signals was able to reduce the misclassification rate of the classifier. Further, we have also prototyped the integration of the proposed wristband-based stress detection system in a consumer end device with voice capabilities.
The fourth contribution of the thesis is the design of an anxiety detection model that uses features from a single wearable sensor and a context feature to improve the performance of the classification model. Using a context feature instead of integrating other physiological features for improving the performance of the model can reduce the complexity and cost of the anxiety detection model. In our proposed work, we have used a simple experimental context feature to highlight the importance of context in the accurate detection of anxious states. Our results and analysis have shown that with the addition of the context-based feature, the classifier was able to reduce misclassification by increasing the confidence of the decision.
The final and the fifth contribution of the thesis is the validation of a proposed computational framework for the blood pressure estimation model. The proposed framework uses features from the PPG signal to estimate the systolic and diastolic blood pressure values using advanced regression techniques.
Digital Object Identifier (DOI)
This work was supported by the Kentucky Science and Engineering Foundation under Grant KSEF-3528-RDE-019. The duration of the funding was from July 2016-June 2019.
Nath, Rajdeep Kumar, "Novel Machine Learning and Wearable Sensor Based Solutions for Smart Healthcare Monitoring" (2021). Theses and Dissertations--Electrical and Computer Engineering. 167.