Project

RF-SCG: Contactless heart recording

signalkinetics

Groups

RF-SCG is a new system that can capture seismocardiogram (SCG) recordings without requiring any contact with the human body. Such an unobtrusive approach would enable lay users to routinely monitor their SCG signals (e.g., on a daily basis), and may provide early warnings of cardiovascular conditions. This capability could be particularly helpful for monitoring high-risk populations – like the elderly, neonates, or patients with arrhythmia – in their everyday environments. It may also enable on-the-spot heart recordings in the event of a cardiovascular emergency. For example, if someone suspects they may be suffering from a heart attack, they could use such a system to immediately measure their SCG.

What is the seismocardiogram (SCG) ?

Seismocardiography was first studied in the late 1950s by scientists who were inspired by the technology used in seismology to register underground vibration and predict earthquakes. They adapted the technology to measure fluctuations of the cardiac movements using accelerometers. The medical community has invested significant effort in studying and understanding SCG recordings since their discovery. Clinical studies have demonstrated that the SCG is more sensitive and specific than the ECG in detecting coronary heart disease during stress exercise testing. Multiple projects have been dedicated to understanding the peaks and valleys of SCG recordings and to mapping these fiducial points to micro-cardiac events. Various studies have demonstrated that SCG recordings can be used to diagnose and monitor various cardiovascular conditions including arrhythmia, myocardial infarction, ischemia, and hemorrhage.

How does RF-SCG work?

Our approach relies on low-cost millimeter-wave radars. The radar transmits a signal and captures its reflection off the user’s chest to sense cardiac micro-vibrations. Because millimeter-wave signals can traverse clothes, this approach neither requires users to take off their clothes nor requires affixing a sensor to their chest.

RF-SCG’s first component is a 4D Cardiac Beamformer that zeroes in on the reflection coming from the apex of the heart. At a high level, this component combines 3D filters with a time-domain CNN to discover the 3D location of the heart while estimating the heart rate. This component also filters out various sources of noise and interference in space and time.
RF-SCG’s second component is an RF-to-SCG Translator that aims to learn a transformation of the function between wireless reflections off the human chest and standard SCG recordings. Our intuition for why learning such a transformation is possible stems from the fact the radar reflections (used in RF-SCG) and accelerometer measurements (which are used in standard SCG recordings) capture chest vibrations that arise from the same underlying micro-cardiac events. Hence, during the training phase, this component uses recordings from an accelerometer placed at the apex of the heart in order to learn the transformation between radar reflections and standard SCG recordings. Once it has been trained, RF-SCG does not need the accelerometer anymore, and it uses the learned translation filters to transform the reflections to SCG recordings.
The final component RF-SCG’s pipeline performs automatic labeling of the SCG recordings in order to extract the timing of five fiducial points of interest: mitral valve closing, isovolumetric contraction, aortic valve opening, aortic valve closing, and mitral valve opening. This component modifies and adapts that the U-Net model – which is typically used in Computer Vision to identify salient features in images – to identify the fiducial points of interest in 1D SCGs.

How well does the system work?

Empirical evaluation with 40,000 heartbeats from 21 healthy subjects demonstrates RF-SCG’s ability to robustly time five key cardiovascular events (aortic valve opening, aortic valve closing, mitral valve opening, mitral valve closing, and isovolumetric contraction) with a median error between 0.26%-1.29%

This project is funded by the National Science Foundation CAREER Award (CNS-1844280) and the Office of Naval Research YIP Award.