Heart failure is the leading cause of mortality worldwide and its complications result in over one million hospitalizations annually in the United States alone. Increasing number of hospitalizations along with escalating health care costs are resulting in a mounting burden on the healthcare system and warrant novel practical approaches to patient management. Remote monitoring of ambulatory patients using wearable and invasive sensors is gaining acceptance as one of the solutions to this problem. Ambulatory sensors that can make physiologic measurements continuously and provide a comprehensive picture of the patient's status over time form the basis of any successful remote monitoring system.
This thesis focuses on one such physiologic sensor, namely heart sounds, for remote ambulatory monitoring. The aim of this research is to develop novel algorithms for accurate measurement and tracking of clinically useful heart sound parameters and the application of these measurements to detect cardiovascular perturbations. This thesis achieves these goals in three main parts.
First, a novel low complexity framework is developed to accurately measure the different heart sound components. For a given heart sound (e.g. S1) a dynamic programming based algorithm is applied to select and track the largest most consistent peak. To establish the value of the proposed framework, it is tested on acute and chronic pre-clinical data collected during heart failure deterioration and compared to a traditional non- tracking algorithm. In all these pre-clinical experiments, the performance of the proposed tracking framework is found to be superior to the traditional non- tracking approach. These results validate that heart sounds measured using the proposed framework contain clinically relevant information about heart failure status that has historically not been available due to the use of a non- tracking approach.
Second, the clinical utility of heart sounds based parameters measured at a non- traditional pectoral location is evaluated in an acute hospitalized setting. In particular the heart sounds ejection time (HSET) which is an indicator of changing LV systolic performance is studied. Heart sound based ejection time measured using our tracking framework is compared to the stroke volume (SV) measurements recorded during hospitalization. In 20 patients with changes in SV > 10 ml the mean correlation coefficient between HSET and SV is found to be R = 0.6762. Also, the HSET is shown to have 70% sensitivity at 80% specificity to detect periods of low stroke volume (SV <50 ml).
Finally, the utility of heart sounds based measurements for the detection of cardiovascular perturbations is evaluated. In particular a system capable of detecting episodes of obstructive sleep apnea (OSA) in heart failure patients with pacemakers and cardiac resynchronization therapy devices is developed. Features for OSA detection are generated by optimally extracting information in the S1 measurements using wavelet decomposition and adaptive dyadic time segmentation. Linear discriminant analysis and support vector machine (SVM) classifiers are trained using different feature selection schemes and tested on an independent test dataset. The tracking based classification is found to consistently outperform non-tracking based classification, emphasizing the importance of tracking. SVM with recursive feature elimination scheme and tracking is shown to have the highest (91.8%) accuracy yielding an improvement of 7% over the non-tracking based approach. The output of the best classifier is used as an OSA severity score which is shown to be correlated significantly (R = 0.72, p<0.05) with the gold standard apnea hypopnea index.