Electrocardiographic imaging (ECGI) reconstructs heart surface electrical signals from body surface ECG recordings by solving an inverse problem that relies on a transfer matrix R, which maps internal heart activity to external measurements. In practice, R is often uncertain due to anatomical variability or modeling approximations, leading to errors in reconstruction. This project investigates a novel approach to mitigate this issue by learning a correction to imperfect R matrices and enforcing consistency across multiple approximations during model training.
Electrocardiographic imaging (ECGI) is a noninvasive technique that aims to reconstruct electrical activity on the heart surface using ECG signals measured from the body surface. This process is known as an inverse problem because it attempts to infer internal cardiac behavior from external observations. A key component of this process is the transfer matrix R, which mathematically relates the heart’s electrical potentials to those observed on the torso. This matrix is typically computed from patient-specific anatomical models using techniques like the boundary element method (BEM), and it encodes how signals propagate through the torso volume.
However, in many practical scenarios—including wearable systems or applications across patient populations— R may be only approximately known due to limited imaging, anatomical variability, or simplifications in modeling. This model mismatch introduces error in the inverse solution and can significantly degrade the accuracy of heart signal reconstruction.
The first task of this project is to investigate how the perturbance of the transfer matrix R affect the cardiac dynamic prediction. After that, the project explores a novel deep learning approach to mitigate such mismatch/uncertainty. The student will use simulated data to evaluate how model inaccuracies affect performance and will help implement a physics-informed neural network that incorporates these robustness strategies. This work will contribute to developing more accurate and generalizable ECGI systems for noninvasive cardiac diagnostics.
• Good programming capability (e.g., Python, PyTorch)
• Basic knowledge in deep learning and neural networks
• Interest in healthcare problems
• Curiosity to learn new techniques in AI and healthcare
1. Learn how inverse problems are used in noninvasive heart imaging.
2. Understand the impact of uncertainty and model mismatch in scientific computing.
3. Gain hands-on experience in deep learning.
4. Explore strategies to improve model robustness through correction and consistency.