CARDIOKOOP: an in-silico multivariate cardiovascular hemodynamic dataset with control perturbations

This dataset contains 500 multivariate cardiovascular hemodynamic trajectories generated with a validated closed-loop lumped-parameter model, each subjected to a controlled preload perturbation (a step change in venous unstressed volume). Model parameters were drawn by Latin-hypercube sampling; each simulation was integrated at Δt = 0.01 s for 1500 steps (15 s) with the control step applied at t = 5 s, and 12 hemodynamic signals (atrial/ventricular pressures and volumes, ventricular in/outflows, aortic pressure) were recorded. Data are provided as a fixed train/validation/test split (400/50/50 trajectories, seed 42): z-scored signal files (*_x.csv), the physical-unit control input (*_u.csv), per-signal normalization statistics, and the per-trajectory parameter table. It provides a fully-specified, reproducible benchmark for control-aware time-series forecasting, surrogate/reduced-order modeling, Koopman/operator learning, and model-predictive-control research on physiological systems. See README.md for the complete schema and a Python loader.