This study presents a novel method using Latent Neural Ordinary Differential Equations (LNODEs) to create a surrogate model for whole-heart electromechanical simulations. The model is trained on 400 simulations of a heart failure patient, incorporating 43 parameters that describe cardiac electrophysiology, active and passive mechanics, and cardiovascular hemodynamics. LNODEs provide a compact representation of the 3D-0D model in a latent space, using an artificial neural network with only 3 hidden layers and 13 neurons per layer, enabling efficient numerical simulations on a single processor. The model is used for global sensitivity analysis and parameter estimation with uncertainty quantification, completing these tasks in just 3 hours on a single processor. The surrogate model is trained to learn the pressure-volume dynamics of the heart, allowing for the analysis of how model parameters influence various quantities of interest (QoIs). The model is validated against a test set of 5 simulations, showing high accuracy with errors ranging from 2% to 6%. The study also demonstrates the robustness of the model in estimating parameters from time-dependent QoIs, with true parameter values contained within 95% credibility regions. The model's ability to capture complex dynamics with a small number of tunable parameters makes it highly effective and robust, outperforming Gaussian process emulators in terms of generalization, particularly for ventricular function. The study highlights the importance of understanding the relationships between model parameters, especially in different cardiovascular compartments, and shows that some parameters have a significant impact on pressure-volume loops. The method is applied to a four-chamber heart model, with the geometry derived from a heart failure patient. The model is used to perform sensitivity analysis and parameter estimation, demonstrating its potential for clinical applications. The study also discusses the extension of the method to incorporate geometric variability and multiple pathological conditions, paving the way for a universal whole-heart simulator that could be used in clinical practice for personalized parameter calibration based on patient-specific data. The method is efficient, requiring only 13 hours of computation on a single core laptop, making it a promising tool for computational cardiology.This study presents a novel method using Latent Neural Ordinary Differential Equations (LNODEs) to create a surrogate model for whole-heart electromechanical simulations. The model is trained on 400 simulations of a heart failure patient, incorporating 43 parameters that describe cardiac electrophysiology, active and passive mechanics, and cardiovascular hemodynamics. LNODEs provide a compact representation of the 3D-0D model in a latent space, using an artificial neural network with only 3 hidden layers and 13 neurons per layer, enabling efficient numerical simulations on a single processor. The model is used for global sensitivity analysis and parameter estimation with uncertainty quantification, completing these tasks in just 3 hours on a single processor. The surrogate model is trained to learn the pressure-volume dynamics of the heart, allowing for the analysis of how model parameters influence various quantities of interest (QoIs). The model is validated against a test set of 5 simulations, showing high accuracy with errors ranging from 2% to 6%. The study also demonstrates the robustness of the model in estimating parameters from time-dependent QoIs, with true parameter values contained within 95% credibility regions. The model's ability to capture complex dynamics with a small number of tunable parameters makes it highly effective and robust, outperforming Gaussian process emulators in terms of generalization, particularly for ventricular function. The study highlights the importance of understanding the relationships between model parameters, especially in different cardiovascular compartments, and shows that some parameters have a significant impact on pressure-volume loops. The method is applied to a four-chamber heart model, with the geometry derived from a heart failure patient. The model is used to perform sensitivity analysis and parameter estimation, demonstrating its potential for clinical applications. The study also discusses the extension of the method to incorporate geometric variability and multiple pathological conditions, paving the way for a universal whole-heart simulator that could be used in clinical practice for personalized parameter calibration based on patient-specific data. The method is efficient, requiring only 13 hours of computation on a single core laptop, making it a promising tool for computational cardiology.