Deep learning methods have recently shown promising results in electrocardiographic imaging (ECGI), but the availability and representativeness of training data remain major challenges. This study evaluates the clinical transfer of PULSE, a pretrained neural network model originally developed using canine torso-tank data, to a human clinical dataset for premature ventricular contraction (PVC) localization. Tikhonov inverse solutions were computed from body surface potentials and used as inputs to the pretrained model without retraining. Performance was assessed using localization error (LE), and uncertainty was estimated via Monte Carlo dropout. Results showed heterogeneous patient-dependent behavior: PULSE improved localization in some cases, matched Tikhonov in others, and degraded performance in the remaining cases. Global uncertainty summaries showed a weak association with localization performance, whereas uncertainty evaluated near clinically relevant regions showed a stronger relationship with localization error. These findings suggest that uncertainty in ECGI should be interpreted as a spatially localized quantity rather than through global summaries alone.