The increasing availability of human electrophysiological data has enabled the development of more refined myocyte models. However, these models often rely on stiff systems of ordinary differential equations (ODEs) to accurately describe the complex ionic dynamics across the cell membrane, requiring very small time steps to ensure numerical convergence. In particular, the voltage-gated sodium channel Nav1.5, which exhibits rapid state transitions on the order of milliseconds, necessitates time steps as small as 0.1 μs when using explicit integration schemes. In this study, we propose a hybrid approach using physics-informed neural networks (PINNs) to model the gating kinetics (m, h, j) of the sodium channel, thereby replacing the corresponding ODEs in the model by Skibsbye et al. for the human atrial action potential. The hybrid PINN-based model reduced the overall computational time by a factor of 2.1× for a time step of 10 μs. Furthermore, it allowed for a tenfold increase in time step size (from 10 to 100 μs), while maintaining a root mean square error below 0.25 mV during the repolarization phase over a train of 100 stimuli delivered at a pacing rate of 1 Hz. Importantly, since the refractory period is largely governed by the inactivation kinetics of sodium channels, we evaluated the restitution properties of the hybrid model. The proposed approach preserved key action potential features, with a maximum deviation of 25 ms in the restitution curve pacing rates below 2.5 Hz. This deviation in the hybrid PINN model is attributed to errors in predicting the kinetics of the inactivation j gate of the voltage-gated sodium channel. These results demonstrate that PINNs can accurately capture the fast dynamics of ion channel gating while significantly improving computational efficiency, offering a promising alternative for high-fidelity cardiac simulations.