Aims: Noninvasive electrocardiographic imaging (ECGI) enables reconstruction of epicardial electrical activity from body surface potentials (BSP). In this study, we assess the generalization capacity of a previously developed nonparametric regression framework based on Multivariate Adaptive Regression Splines (MARS), by training the model across subjects and predicting epicardial signals in an unseen animal.
Method: The MARS model was trained using BSP and corresponding electrograms (EGMs) recorded from 4 anesthetized, closed-chest pigs. In previous work, the model was trained and tested on the same individual . Here, we used a leave-one-pig-out strategy: the model was trained on all pigs except that used exclusively for testing. Epicardial and torso potentials were recorded using 239 unipolar electrodes on the heart and 184 electrodes on the torso. To standardize the spatial representation of signals across subjects, BSP and EGM data were transformed from 3D subject-specific models to standardized 2D maps before training. Performance was assessed by the correlation coefficient (CC) between predicted and true EGMs.
Results: Two pigs were selected for testing: one with limited pacing sites and another with a more extensive pacing dataset. The correlation coefficient between predicted and true EGMs was improved for the first pig (0.55±0.22 vs 0.31±0.07 with within-subject training), likely due to the broader and more diverse training set. For the second pig who already had an extensive dataset, the correlation was moderately reduced compared to same-subject training (0.62±0.08 vs 0.77±0.10) demonstrated the error remaining from generalization between pigs. Conclusion: This study highlights the robustness of the MARS method in cross-subject scenarios and its potential for generalization beyond a single-subject training population. While performance slightly decreases compared to same-subject training, results remain encouraging. Further improvements are expected by optimizing the 3D-to-2D transformation used for spatial standardization.