Diagnostic assistance for the prediction of cardiac diseases with a non-invasive estimation of hemodynamic parameters using machine learning

Abstract

This research describes the process of the prediction for cardiac output (CO), with a non-invasive method implementing machine learning techniques. This mainly follows the Windkessel model developed and studied by several researchers who aimed to implement similar methods and variations of this thermodynamics theory for human blood flow. Cardiac output is crucial for determining the demands of tissue oxygen that meets human body’s heart demands, and it is often measured with Swan-Ganz catheters, which represents a risk in 10% of the cases for patients in clinical and ICU scenarios. The databases, contemplated in this research are compared in order to proceed with the most suitable implementation. These are as follows: The MIMIC II Clinical Database, MGH/MF Waveform Database and the Hemodynamics and Respiratory Pattern Dataset. The latter was selected for this work which contained 170 records with cardiac output values, which worked as ground truths for the prediction model. Therefore, algorithms such as XGBoost vector regression, decision trees, and random forest are implemented to forecast cardiac output based on the thermodilution theory in this research use case. Among such algorithms, random forest gave the best efficiency, as 97.77% (±4.95%) of accuracy was obtained through cross-validation, under the setup of 700 decision trees. This represents an acceptable enhancement for forecasting cardiac output in contrast between the work in the state of art which resulted in 79,91% of accuracy implementing the same machine learning technique but with different data and similar features.