Measurement-Dependent Parameter Recoverability in Neural Cardiovascular LPN Inference

A new study explores the recoverability of cardiovascular lumped-parameter network (LPN) parameters from limited hemodynamic data. Inverse problem estimation of LPN parameters often remains non-unique when working with incomplete observations. Researchers developed Mada, a probabilistic neural framework that delivers parameter estimates and predictive uncertainty even with missing measurements.

Using a synthetic population of 650,000 cases with 20 chamber-mechanics and vascular parameters, and up to 30 hemodynamic and timing observations, Mada showed macro R2 values ranging from 0.066 to 0.842. The framework's ability to recover LPN parameters depended significantly on the availability of measurements. Mask-aware Mada performed best with a rich clinical set but struggled with basic clinical measurements, while uncertainty increased as informative measurements were removed.

Direct optimization yielded smaller hemodynamic residuals but failed to recover the generating parameters consistently. Profiling and local-sensitivity analyses revealed case-specific parameter compensation and weak observational directions, while fine-tuning specific targets provided little improvement for difficult atrial parameters. Adding measurement noise worsened both parameter recovery and predictive-interval coverage, mirroring similar trends with a simpler probabilistic ResMLP.

This research highlights that accurately reconstructing observed hemodynamics does not always guarantee accurate recovery of the underlying LPN parameters. Emphasizing parameter recoverability and predictive uncertainty is crucial when seeking biologically meaningful estimates.