Machine learning-based analysis of non-invasive measurements for predicting intracardiac pressures

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Early detection of congestion improves heart failure (HF) patient outcomes.
• Limited access to invasive hemodynamic parameters hinders optimal HF treatment.
• Pulmonary capillary wedge pressure (PCWP) is a key hemodynamic parameter for HF management.

Purpose of the Study:

• To develop a model for estimating invasive PCWP using non-invasive measurements.
• To compare traditional statistical and machine learning (ML) techniques for PCWP estimation.
• To assess the feasibility of non-invasive hemodynamic monitoring in HF patients.

Main Methods:

• Utilized data from 853 right-sided heart catheterization procedures (2017-2022).
• Included non-invasive parameters: arterial blood pressure, saturation, heart rate variability, weight, and temperature.
• Applied various traditional and ML regression/classification models, assessing performance with R² and AUC.

Main Results:

• Heart rate variability showed the highest correlation (0.16) with PCWP.
• Regression models yielded low R² values (up to 0.04).
• Classification models achieved AUC values up to 0.59, indicating limited predictive power.

Conclusions:

• Non-invasive methods, including ML, demonstrated limited correlation with invasive PCWP.
• Current non-invasive monitoring shows weak correlation with hemodynamic parameters.
• Future research should focus on trend analysis and additional features for improved non-invasive hemodynamic monitoring.