Abstract
This thesis aims to develop physiological indices using modern signal processing approaches. The developed indices can be used for the classification and detection of different cardiovascular and cerebrovascular diseases in patients. This thesis comprises four studies analysing various physiological signals such as arterial blood pressure, electrocardiogram, intracranial pressure, and mean arterial pressure obtained from four different physiological signal data sets. The first study is aimed to analyse the arterial baroreflex role in cardiovascular control. The baroreflex is a significant modulator of cardiovascular control and maintains suitable blood pressure under orthostatic stress, which could otherwise lead to severe hypotension. The aim of this study is to develop new physiological indices for the analysis of baroreflex control that follow active and passive postural changes. To achieve this, an innovative principal dynamic mode (PDM) system identification approach is applied to find data-adaptive frequency components arising from closed-loop interactions between beat-to-beat intervals and systolic blood pressure recorded from 10 healthy subjects. The gain of global PDMs of both cardiac and mechanical arms may be able to distinguish between active and passive orthostatic testing in healthy subjects. The trends of nonlinear functions associated with global PDMs of both cardiac and mechanical arms may function as potential clinical markers for postural changes. Second study investigated the role of intracranial hypertension in traumatic brain injured patients. Intracranial hypertension is a fatal neurological condition that carries a high mortality risk. Its timely identification and treatment are important to the functional recovery and resuscitation of the patients. The purpose of this study is to identify quantiiii Abstract tative measures for the early detection of intracranial hypertension episodes in traumatic brain injured (TBI) patients and to investigate the association between intra-individual variability and intracranial hypertension (IH) episodes. To achieve this, the Granger causal (GC) analysis was applied to quantify the bi-directional informational flow patterns between MAP, ICP, and HR. The coefficient of variation of GC values was also estimated to measure the intra-individual variations. The findings of this study suggest that the onset of intracranial hypertensive events appears to be associated with directional communications between cerebrovascular (ICP) and cardiovascular (MAP, HR) mechanisms. Moreover, these derived GC measures can also be used as functional biomarkers in physiological diagnostics. The third study sought to detect and classify diabetes mellitus using wavelet transform and convolutional neural networks. Diabetes mellitus, a metabolic disorder, is caused by a high concentration of glucose in the blood. Its early detection and classification can help to control the disease. Deep learning advances have contributed significantly to improve the quality of health care. We investigated various deep learning architectures and proposed a continuous wavelet transform (CWT) based model for diagnosis and classification of diabetes mellitus using an electrocardiogram (ECG) signal. We trained and evaluated our proposed method on the diabetes mellitus data set and realized that the VGG-16 model achieved the highest performance parameters. The developed model can be used as potential indices for the detection and classification of diabetes mellitus. The fourth and final study is focused on detection and classification of atrial fibrillation (AF) in ECG signals. The automatic and early detection of atrial fibrillation in ECG signals can be helpful for its efficient treatment and to avoid any irreversible damage. In this study, we developed an efficient method for atrial fibrillation detection and classification, based on dynamic features extractions using signal processing (discrete wavelet transform and mel frequency cepstrum coefficient) and machine learning (support vector machine and convolutional neural networks) approaches. The proposed method exhibited state-of-the-art results, and its performance is validated by analysing the recently released AF dataset on physionet. The developed model can be considered a potential biomarker for the detection and classification of AF in electrocardiogram signals. This thesis developed novel physiological indices for classification and detection of different cerebrovascular diseases such as traumatic brain injury and cardiovascular diseases such as atrial fibrillation and diabetes mellitus using modern signal processing and Abstract machine learning approaches.
D.Phil (Electrical and Electronic Engineering)