This paper presents a novel approach to creating a comprehensive surrogate model for cardiac electromechanics using Latent Neural Ordinary Differential Equations (LNODEs). The goal is to develop a fast and accurate whole-heart digital twin for personalized medicine, which is currently hindered by the high computational costs of high-fidelity multi-scale cardiac models. By training LNODEs on 400 3D-0D closed-loop electromechanical simulations of a heart failure patient, the model learns the pressure-volume dynamics while accounting for 43 parameters describing cardiac electrophysiology, mechanics, and hemodynamics. The trained LNODEs enable global sensitivity analysis and robust parameter estimation with uncertainty quantification (UQ) on a single CPU in just 3 hours. The model's accuracy is validated through testing errors ranging from 2% to 6%, and it captures complex dynamics with a compact representation. The study also highlights the importance of certain model parameters in influencing different quantities of interest (QoIs) and provides insights into the interplay between parameters across different cardiovascular compartments. The proposed method opens the door to extending the model to incorporate geometric variability and multiple pathological conditions, potentially leading to a universal whole-heart simulator for clinical practice.This paper presents a novel approach to creating a comprehensive surrogate model for cardiac electromechanics using Latent Neural Ordinary Differential Equations (LNODEs). The goal is to develop a fast and accurate whole-heart digital twin for personalized medicine, which is currently hindered by the high computational costs of high-fidelity multi-scale cardiac models. By training LNODEs on 400 3D-0D closed-loop electromechanical simulations of a heart failure patient, the model learns the pressure-volume dynamics while accounting for 43 parameters describing cardiac electrophysiology, mechanics, and hemodynamics. The trained LNODEs enable global sensitivity analysis and robust parameter estimation with uncertainty quantification (UQ) on a single CPU in just 3 hours. The model's accuracy is validated through testing errors ranging from 2% to 6%, and it captures complex dynamics with a compact representation. The study also highlights the importance of certain model parameters in influencing different quantities of interest (QoIs) and provides insights into the interplay between parameters across different cardiovascular compartments. The proposed method opens the door to extending the model to incorporate geometric variability and multiple pathological conditions, potentially leading to a universal whole-heart simulator for clinical practice.