Machine Learning based Estimation of Aortic Pressure Curves by Electrical Impedance Tomography

This repository contains algorithms, training routines and results for EIT based estimation of central aortic pressure (CAP) curves

Abstract:
Central aortic pressure is a key hemodynamic parameter to monitor and target in clinical practice. As this gold standard method is highly invasive and conventional non- invasive methods are not long-term compatible or inaccurate, the need for alternative monitoring capabilities arises. Electrical impedance tomography (EIT) is a non-invasive monitoring technique using an electrode belt around the torso. In this paper, EIT recordings from an in-vivo animal model and simultaneously recorded central aortic pressure measurements from invasive catheters are used to train a convolutional neural network predicting aortic pressure curves from EIT voltages. Different parametric representations of the aortic pressure time series are considered to reduce network complexity. A hyperparameter tuning is conducted to optimize the network. Results demonstrate that the estimation of aortic pressure curves by a trained network is feasible even on unknown test data, however, random offsets are observed.

Overview:

• Structure
• Installation
• Evaluation Results
• Author

Structure

This repository includes: - estimteparametrization.py: A script to create parametric representations of aortic pressure curves and save them as .mat files. Parametric techniques: - Piecewise Linear Regression - Spectral Estimation with Cauchy-Lorentz Pulses - Hierarchical Approach - trainmodelXX.py : To train the with Kerastuner designed CNNs for estimation and save them. - evalnn_XX.ipynb: Notebooks that show the results on the test data sets for all three parameterization techniques.

Installation

Clone the repository: bash git clone https://github.com/EITLabworks/ML-based-Estimation-of-Aortic-Pressure-Curves-by-EIT.git

Evaluation Results

Piecewise Linear Regression

Spectral Estimation

Piecewise Linear Regression

Author

This repository is created by Patricia Fuchs, Institute of Communications Engineering, University of Rostock, Germany.
The research is explained and summarized in the paper "Machine Learning based Estimation of Aortic Pressure Curves by Electrical Impedance Tomography" for the "Annual International Conference of the IEEE Engineering in Medicine and Biology Society" (EMBC) 2025.
For questions, please contact: pat.fuchs@uni-rostock.de