SPARK: Physiology-informed inference of latent cardiac autonomic dynamics from RR intervals

Cardiac autonomic activity is observed only through beat-to-beat (RR) intervals, and separating it into parasympathetic-like and sympathetic-like components is an inverse problem that RR alone is not expected to solve uniquely. We introduce SPARK (Stochastic Phase-domain Autonomic Reconstruction Kinetics), a state-space model in which two independent, mean-reverting latent drives modulate heart rate through an integrate-and-fire pacemaker. The model is combined with exact continuous-time Kalman filtering and with an explicit physiological identification structure: time-scale priors, a kinetic ordering and a spectral partition of heart rate variability. SPARK offers a model-conditional description of autonomic trajectories that complements conventional heart rate variability; its latent states are reconstructions, not measurements of neural activity. Across recordings, fitting each recording from many different starting points gave a median of 23.5 distinct solutions without physiological information versus 1.0 with it, and one process absorbed essentially all the variability in 75% versus 4% of the fits; the fit diagnostics showed residual structure in most recordings. Reconstructions were then examined: in 10 tilt-table recordings, the vagal-like state decreased and the sympathetic-like state increased in 10 of 10; exploratory comparisons between 20 young and 20 older adults showed differences in reconstructed variability and faster-component kinetics, although several latent summaries were mathematically redundant; and in 263 exercise recordings, median recovery times, which are closely tied to the fitted time constants, were 42 s (vagal-like) and 341 s (sympathetic-like). How a stressor-induced change is split between the two drives remained sensitive to the spectral anchor.