Automated Computation of Hemodynamic Metrics Based on Non-invasive Electrophysiological and Biomechanical Features

Heart failure (HF) is a serious condition wherein the heart is unable to supply sufficient amount of blood into circulation to meet the body’s demand. It has been recognized as an emerging epidemic, and its prevalence is growing. Therefore, an immense need for innovative approaches to prevent and treat HF has emerged. It is estimated that about 50% of HF patients have impaired ventricular relaxation. A critical metric to assess this condition is pulmonary capillary wedge pressure (PCWP). Another critical pressure metric is pulmonary artery pressure (PAP), which can be used to assess the risk of HF and further adverse events in patients who already have HF. The gold standard method to measure PCWP and PAP is right heart catheterization (RHC) wherein a catheter is inserted into a patient’s vein and advanced into the heart and artery to measure the pressure metrics. It is highly invasive and requires expertise, limiting its utility for routine monitoring of PCWP and PAP.Our objective is to develop a system that meets the urgent and critical need for highly available routine pressure monitoring. To achieve this, the system needs to be safe, easy to use, and capable of accurate measurement of the pressure metrics. We have developed a system that combines cardiac signals simultaneously recorded with non-invasive wearable sensors from electrophysiological and biomechanical sources. The system includes a number of signal processing methods to extract features that correlate with the pressure metrics and accurately computes these metrics via machine learning algorithms. Thus, the system does not require expertise to operate and acquire accurate pressure measurements.The signal processing methods include heartbeat segmentation, noise suppression, and extraction of signal characteristics as features to train the machine learning algorithms. We have developed a novel hierarchical algorithm consisting of several regression models to accurately compute PCWP and PAP values. In addition to noise suppression, we have developed signal quality assurance methods to reject low quality signals in order to maintain accurate computation of PCWP and PAP.The system has been validated via leave-one-out analysis with clinical trial data acquired from a diverse patient population. During the trial, RHC was performed to establish the ground truth, and our system was applied immediately before or after to record signals. Bland-Altman analysis on the predicted PCWP and the ground truth PCWP yields a bias of 0 and 95% limits of agreement of �4.76 mmHg. The same analysis on PAP yields a bias of 0 and 95% limits of agreement of �11.88 mmHg.As a potential expansion of the system, systolic time intervals (STI) have been explored. STI have been widely studied as an indirect measurement of ejection fraction, which is another important metric to assess HF, and are measured as the time intervals between electrical and mechanical cardiac events: the onset of ventricular depolarization and the aortic valve opening and closure events. We have developed a flexible algorithm to detect the onset of ventricular depolarization in challenging electrocardiography signals with various morphologies and noise levels. We have also developed a method to detect a surrogate feature for the aortic valve opening in low-frequency phonocardiography signals. Combining the onset of ventricular depolarization, the surrogate feature, and the onset of second heart sound as the aortic valve closure, we have derived STI that have shown strong correlation with ejection fraction