Machine Learning Algorithm to Predict Cardiac Output Using Arterial Pressure Waveform Analysis
Ke Liao, Armağan Elibol, Wei Xiao, Liao Cenyu, Wei Wang, Nak Young Chong
Abstract
Cardiac Output (CO) is a key hemodynamic variable that can be estimated in a minimally invasive way via using Arterial Pressure Waveform Analysis (APWA). Many models use circulation mechanics to build the relationship between arterial pressure and CO. In this study, we attempt to apply machine learning and feature engineering to analyze the Arterial Pressure Waveform (APW) and create regression models to predict the CO. We utilize the traditional APWA model knowledge and the time-domain, frequency-domain, and other characteristics of time series data for feature engineering. We present the benchmarking results for several machine learning models using the MIMICII waveform database. We compare the predicted CO values from our proposed models with the “gold standard” TCO (CO measured by intermittent pulmonary artery thermodilution). Our results show that the Random forest model has the most accurate agreement (MSE: 1.421 $\displaystyle \text{L}/\min$, bias: $-0.01\displaystyle \text{L}/\min$, 95% limits of agreement: $-2.35\displaystyle \text{L}/\min$ to $+2.32\displaystyle \text{L}/\min$, percentage error: 39.44%). Notably, the XGBoost model demonstrates good tracking ability with TCO (radius bias: 11.79o, 95% radius limits of agreement: ±28.89°), achieving the clinically acceptable level.