Noise-robust bioimpedance approach for cardiac output measurement

Congestive heart failure is a problem affecting millions of Americans. A continuous, non-invasive, telemonitoring device that can accurately monitor cardiac metrics could greatly help this population, reducing unnecessary hospitalizations and cost.

Machine learning (ML) algorithms trained on electrical-impedance tomography (EIT) data are presented for portable cardiac monitoring. The approach was validated on a simulated thorax and a measured tank experiment. A highly detailed 4D chest model (finite element method mesh and conductivity profiles) was developed utilizing the 4D XCAT phantom to provide realistic data. The ML algorithms were trained using databases that assumed the presence of poorly contacting electrodes without any assumptions of knowing which electrodes would be bad in the experiment. The trained ML algorithms were compared to EIT evaluated with and without removing bad electrodes.

A regression support vector machine and a deep neural network (DNN) were found to be the most accurate and robust to poorly contacting electrodes while not needing to know which electrodes were in poor contact in the simulated and measured experiments, respectively.

Although the ML algorithms are not always better than EIT (with bad electrodes removed), the comparable results without needing a priori knowledge of which electrodes are bad is seen as a very promising feature. An evaluation of computational costs showed that the DNN required comparable computational power to the other methods while requiring less memory, which could make the DNNs an attractive algorithm for a low-power, portable system. This work represents an important validation of the method using measured data, and model development, which is needed to apply this method on real clinical data. Additionally, the developed 4D simulated thorax model could be an important tool within the EIT community.