Machine learning-based blood pressure estimation using impedance cardiography data.

Objective: Accurate blood pressure (BP) measurement is crucial for the diagnosis, risk assessment, treatment decision-making, and monitoring of cardiovascular diseases. Unfortunately, cuff-based BP measurements suffer from inaccuracies and discomfort. This study is the first to access the feasibility of machine learning-based BP estimation using impedance cardiography (ICG) data.

Methods: We analyzed ICG data from 71 young and healthy adults. Nine different machine learning algorithms were evaluated for their BP estimation performance against quality controlled, oscillometric (cuff-based), arterial BP measurements during mental (Trier social stress test), and physical exercise (bike ergometer). Models were optimized to minimize the root mean squared error and their performance was evaluated against accuracy and regression metrics.

Results: The multi-linear regression model demonstrated the highest measurement accuracy for systolic BP with a mean difference of -0.01 mmHg, a standard deviation (SD) of 10.79 mmHg, a mean absolute error (MAE) of 8.20 mmHg, and a correlation coefficient of r = 0.82. In contrast, the support vector regression model achieved the highest accuracy for diastolic BP with a mean difference of 0.15 mmHg, SD = 7.79 mmHg, MEA = 6.05 mmHg, and a correlation coefficient of r = 0.51.

Conclusion: The study demonstrates the feasibility of ICG-based machine learning algorithms for estimating cuff-based reference BP. However, further research into limiting biases, improving performance, and standardized validation is needed before clinical use.