Impulse Data Models for the Inverse Problem of Electrocardiography.

Area of Science:

• Biomedical Engineering
• Computational Cardiology
• Artificial Intelligence in Medicine

Background:

• Traditional electrocardiography faces challenges in accurately reconstructing cardiac electrical activity.
• Inverse problems in electrocardiography are complex and computationally intensive.
• Neural networks offer a promising approach to model complex physiological signals.

Purpose of the Study:

• To develop and train neural networks for predicting heart surface potentials (HSPs) from body surface potentials (BSPs).
• To re-frame traditional inverse problems in electrocardiography as regression problems using Gaussian impulse basis functions.
• To test the predictive capabilities of the trained neural network with synthetic and experimental data.

Main Methods:

• Generated impulse HSPs using Gaussian basis functions and projected them to BSPs via a volume conductor torso model.
• Mapped BSPs (inputs) and HSPs (outputs) to 2D surface meshes for neural network training.
• Utilized a fully connected single hidden layer neural network to map body surface impulses to heart surface Gaussian basis functions.

Main Results:

• The neural network accurately predicted synthetic pulses with a root mean squared error of 9.1±1.4%.
• Predictions remained robust to noise (up to 20 dB) and predictable errors from heart displacement/rotation.
• Successfully decomposed in-vitro pig heart HSPs and calculated activation times with a mean absolute error of 10.4±11.4 ms.

Conclusions:

• Gaussian basis function impulses provide an effective and robust method for training neural networks to reconstruct HSPs from BSPs.
• The developed method offers a non-invasive approach to identify cardiac electrical dysfunction through activation mapping.
• Predicted HSPs can guide treatment options for cardiac electrical abnormalities.