A Pilot Study Comparing Aortic Valve Area Estimates Derived from Fick Cardiac Output with Estimates Based on Cheetah-NICOM Cardiac Output

Estimation of cardiac stroke volume from radial pulse waveform by artificial neural network

BACKGROUND AND OBJECTIVES

Stroke volume (SV) and cardiac output (CO) are the key indicators for the evaluation of cardiac function and hemodynamic status during the perioperative period, which are very important in the detection and treatment of cardiovascular diseases. Traditional CO and SV measurement methods have problems such as complex operation, low precision and poor generalization ability.

METHODS

In this paper, a method for estimating stroke volume based on cascade artificial neural network (ANN) and time domain features of radial pulse waveform (SV_ANN) was proposed. The simulation datasets of 4000 radial pulse waveforms and stroke volume (SV_meas) were generated by a 55 segment transmission line model of the human systemic vasculature and a recursive algorithm. The ANN was trained and tested by 10-fold cross-validation, and compared with 12 traditional models.

RESULTS

Experimental results showed that the Pearson correlation coefficients and mean difference between SV_ANN and SV_meas (R=0.95, mean standard deviation (SD) = 0.00 ± 6.45) were better than the best results of the 12 traditional models. Moreover, as increasing the number of training samples, the performance improvement of the ANN (R=0.94(Δ + 0.04), mean ± SD = 0.00 ± 6.38(Δ± 2.02)) was better than the other best model, namely, multiple linear regression model (MLR) (R=0.93(Δ + 0.03), mean ± SD = 0.00 ± 6.99(Δ± 1.50)).

CONCLUSIONS

A method is proposed to estimate cardiac stroke volume by the ANN with time domain features of radial pulse wave. It avoids the complicated modeling process based on hemodynamics within traditional models, improves the estimation accuracy of SV, and has a good generalization ability.

MEASURING CARDIAC OUTPUT THROUGH THERMODILUTION BASED ON MACHINE LEARNING

Cardiac output (CO) refers to the amount of blood ejected from a unilateral ventricle per minute and is an important measure of cardiac function. Thermodilution is the gold standard for CO measurement because of its accuracy. However, the traditional thermodilution method requires calibration of the correction factor before measurement, which makes its practical application difficult. Therefore, conducting CO measurement by using a machine-learning-based thermodilution method is proposed in this paper, and CO is regressed and predicted through the thermodilution curve by a machine learning model. In this paper, we constructed five cardiac vascular models, and three of them were randomly selected to simulate the thermodilution process. Nine features of the thermodilution curve from the time–frequency domains were extracted and fed into the multilayer perceptron model for training. On the basis of a cross-validation method, the accuracy of the final prediction model was 97.99% ([Formula: see text]%). Simultaneously, a trained neural network was used to predict the CO of the remaining two cardiac vascular models, and the resulting error was within 5%. In this paper, an experimental system consisting of a water pump, a three-way valve and a temperature sensor is also designed, and the thermodilution curves at different quantities of flow are tested and regressed and predicted with the above model, with the error being within 10%, which met the requirement for real-world use, and thus, a method was established for measuring CO by using machine-learning-based thermodilution.