Background: Electrocardiographic Imaging (ECGI) reconstructs cardiac electrical activity non-invasively from body surface potentials, but it is an ill-posed inverse problem. Multivariate Adaptive Regression Splines (MARS) provide a data-driven alternative to traditional regularization, yet become computationally expensive on high-resolution clinical data. This study proposes a patient-specific adaptive strategy that dynamically scales MARS model complexity to improve efficiency while preserving localization accuracy. Methods: Clinical data from 10 patients with premature ventricular contractions (PVCs) were analyzed. Training data were simulated using a cellular automaton and Boundary Element Method on inhomogeneous geometries. Parametric sweeps assessed key MARS parameters (threshold-eps and term limit-Lmax). An adaptive Lmax formulation was developed using a patient-specific saturation model based on spatiotemporal volume (V) to balance data availability and model complexity. Results: High-complexity baseline model (Lmax=128, eps = 0.001) yielded accurate results (LE: 17.2 mm) but required impractical runtimes (43.4 hours). The proposed adaptive strategy (Lmax) outperformed static constraints, reducing the mean localization error (LE) in the calibration set from 18.85 mm to 15.99 mm. In a representative clinical validation case, the adaptive model achieved an LE of 14.4 mm while reducing runtime from 49.8 hours to 9.5 hours. Conclusions: Directly constraining Lmax provides more reliable control over model complexity than threshold-based stopping, which can prematurely halt fitting. Scaling Lmax with patient-specific data volume reduces overfitting while preserving physiological signals, improving the feasibility of MARS-based ECGI for pre-procedural planning. Further reductions in runtime are still needed.