Abstract
The life expectancy of the residents in developed countries has been growing in the last few decades due to advances in medicine and healthcare, but may create a complicated medical condition. In addition, the enormous aging population is expected to engender a broad aging demographic. These issues create a notable impression on the socio-economic structure of developed countries, by increasing the per capita Medicare spending and imposing an additional burden on the Medicare system. Therefore, a need to develop an efficient, cost-effective, and easy-to-use remote monitoring system to serve the increasing needs of the elderly population. The objective of this dissertation is to develop algorithms that extract meaningful information that can be used in a remote monitoring system. Two clinical applications: Parkinson Disease and cardiac arrhythmia, are investigated in this dissertation. Parkinson's Disease (PD) patients mostly exhibit freezing of gait (FOG), and it occurs in the advanced stages of the disease. FOG significantly affects the motor ability of the patients, increasing the risk of falling, and decreases the quality of life of the patients. Currently, the standard measure of FOG is lacking, and as of now, detecting FOG is accomplished by employing a patient's self-reports or motor assessment conducted by healthcare professionals in the hospital setting. In this dissertation, we develop a user-independent FOG detection and prediction model. The proposed methodology is divided into three phases. In the first phase, features are extracted from the dataset. In the second phase, the data is divided into two clusters based on FOG events. In the third phase, significant factors are chosen using a randomized block design with replication. Random Forest and Neural Network models are built using a combination of significant factors obtained from the design of experiments. The network was tested on 8 hours of recorded data from PD patients. We illustrated the system performance based on the user-dependent and user-independent models, different algorithms, and preprocessing window size. Sudden cardiac death (SCD) is a developing medical issue and a significant reason for death, particularly in developed countries like the United States. There is a need to perceive the risk at early stages to avoid potential hazards and control the overall danger to the patient's health. Classification of cardiac arrhythmias is difficult for various reasons: the presence of noise, irregularity in a heartbeat, and similarities in properties of different cardiac arrhythmias. We examine the use of the Deep Neural Network for classifying ECG recording. We trained the network to categorize the database on 17 different classes, including normal sinus rhythm, pacemaker, and 15 arrhythmias. On the dataset, we train a 6-layer bidirectional Long short-term memory (BDLSTM) neural network model, which maps a sequence of ECG samples to a sequence of classes. We further extracted a set of features- autoregressive (AR) signal parameters- and train 4-layer BDLSTM neural network model. Compared to the current state of the art research, our results are the best to date in regards to accuracy, sensitivity and specificity. In conclusion, our solution is efficient compared to contemporary techniques and can be integrated into mobile devices and cloud computing to develop a remote monitoring system. These remote monitoring systems can provide further aid to patients suffering from Parkinson’s disease and cardiac arrhythmia in out-of-hospital settings. Furthermore, it would help to address the burden imposed on Medicare by reducing the need of patients to visit hospitals, increasing quality of patient care, and improving clinical performance.