Neural Network Guided Calibration for Fast Virtual Twin Generation in Cardiovascular ODE Models

Abstract

Calibration of closed-loop lumped-parameter cardiovascular models remains a major bottleneck for scalable digital-twin generation because inverse estimation is ill-conditioned and typically requires computationally expensive iterative forward simulation. This study investigates whether a supervised neural network (NN) can provide a fast inverse estimator for a paediatric sepsis cardiovascular ODE model by learning a direct mapping from prescribed haemodynamic target vectors to calibrated parameter sets. Training data are generated by sampling model parameters at random, forward-simulating the closed-loop system to steady state, and pairing the resulting target summaries with the corresponding parameters; the same target definitions and evaluation populations are used throughout for consistency. We evaluate NN inference by forward re-simulation to steady state and benchmark performance against a simulator-constrained calibration reference (Embedded Gradient Descent, EGD) using relative-error statistics, distributional similarity of achieved outputs and inferred parameters (median shift, IQR ratio, Wasserstein distance, KS statistic), and target-space localisation of parameter-space disparity (cosine distance). The NN reproduces the prescribed targets with predominantly small errors for most samples, while the largest discrepancies are confined to a well defined set of target configurations that also yield high residuals under the reference method, indicating feasibility limits of the target/model combination. Overall, NN-guided calibration provides a computationally efficient accelerator for virtual-twin generation and target-space screening, with simulator-based refinement and forward re-simulation retained to handle infeasible regimes and enforce mechanistic plausibility.