Aims: The standard equivalent dipole model (EDM) was developed on a generic patient anatomy. The objective of this study is to extend this model by simultaneously learning patient-specific propagation properties and dipole characteristics from the ECG.
Methods: The standard EDM (HOM) was reformulated into a Bayesian statistical framework that allowed generalization to a patient-specific EDM (INH) that is not limited by the unbounded uniform body assumption. This is achieved by learning the propagation function from the position of the dipole to the position of each electrode using the ECG and the position of the electrodes, without the need for any additional information (geometry of the heart, position and conductivity of the organs, ...). Both models were tested on simulated data with a detailed heart-torso model with four different activation sequences and three different sets of tissue characteristics. The reconstructed ECGs for each model were compared to the simulated ECG data using the coefficient of determination score (R2 score).
Results: Through all different sets of tissue features and all different activation sequences, we observed a considerable gain in median R2 score per patient (>5%) when comparing the HOM model with the INH model.
Conclusion: The INH method developed in this work is an important improvement of the HOM model by taking into account the specific characteristics of the patient to build a propagation model. The translation of these results to real patient data will lead to a cardiac activity model that is more accurate than existing HOMs while relying on minimal imaging data.