Measurement-Dependent Parameter Recoverability in Neural Cardiovascular LPN Inference

Mada shows that in a synthetic four-chamber cardiovascular LPN with time-varying elastance, parameter recoverability drops sharply with fewer measurements, with macro R2 falling to 0.066 using basic clinical measurements.

• Why it matters: Cardiovascular LPN personalization is a non-unique inverse problem, so matching hemodynamics may not yield physiologically meaningful parameters. The study tests whether neural inference fails or whether limited, noisy observations prevent recovery.
• What they did: A probabilistic neural framework, Mada, was trained on 650,000 synthetic cases with up to 30 hemodynamic and timing observations, varying measurement masks and noise. It compared recoverability to bounded direct optimization and a simpler probabilistic ResMLP, using profile and local-sensitivity analyses.
• The result: With mask-aware Mada, macro R2 increases from 0.066 (basic clinical) to 0.751 (rich clinical) and 0.842 (all measurements), while uncertainty rises as informative measurements are removed. Higher noise degraded both parameter recovery and predictive-interval coverage, and better hemodynamic residuals did not consistently recover generating parameters.

AI summary of the abstract · may contain errors · not medical advice · read the paper.