Continuous Estimation of Left Ventricular Hemodynamic Parameters Based on Heart Sound and PPG Signals Using Deep Neural Network

Continuous estimation of left ventricular hemodynamic parameters is helpful to early diagnosis of cardiovascular diseases. Current non-invasive methods are somewhat inconvenient to monitor these parameters. Here, a deep neural network is built to noninvasively estimate left ventricular systolic pressure (LVSP), left ventricular diastolic pressure (LVDP), maximum rate of left ventricular pressure rise (+ dp/dt(max)) and minimum rate of left ventricular pressure drop (- dp/dt(min)) based on heart sound and PPG signals. The model consists of residual network and bidirectional recurrent neural network. Performance is evaluated on 2 beagle dogs' experiment data with large ranges induced by epinephrine. Mean absolute errors and standard deviations between the estimated and the measured LVSP, LVDP, + dp/dt(max) and - dp/dt(min) are 7.23±8.33 mmHg, 2.12±3.0 mmHg, 298±406 mmHg/s, and 172±386 mmHg/s, respectively. The average correlation coefficients for LVSP, LVDP, + dp/dt(max) and - dp/dt(min) are 0.94, 0.86, 0.95 and 0.92. The results show that accurate intraventricular hemodynamic parameters can be achieved by non-invasive heart sound and PPG signals with deep neural networks. This technique suggests an easy way for real-time monitoring of intraventricular hemodynamics.