Electrocardiographic imaging aims to reconstruct epicardial electrical activity from body-surface potential measurements, but the inverse ECG problem is highly ill-posed and becomes even more challenging under realistic clinical conditions such as missing electrodes and measurement noise. Although recent deep learning methods have shown promising performance, many of them degrade substantially when the input data are corrupted by lead loss or noise. In this study, we investigate the use of a Kolmogorov–Arnold Network (KAN)-based model for robust inverse ECG reconstruction. We evaluate the proposed approach on paced body-surface and epicardial recordings and benchmark it against prominent neural network methods, including a multilayer perceptron and a residual network. Our experiments show that while conventional deep models can achieve very high reconstruction accuracy under clean conditions, their performance deteriorates more rapidly under severe lead dropout. In contrast, the KAN-based model preserves higher temporal and spatial reconstruction fidelity in missing-lead scenarios, indicating improved robustness to incomplete measurements.