Abstract / Summary
Continuous-time intensive care monitoring requires predictive analytics systems to infer physiologic states from data streams that are inherently irregular and workflow-driven. Current clinical AI pipelines often overlook how model uncertainty accumulates between observations, creating a risk of overconfidence as recorded information becomes older. We propose a two-dimensional continuous-time state-space framework with Kalman filtering to quantify model-implied “uncertainty inflation” under irregular sampling. The framework was evaluated using 2998 ICU stays from MIMIC-IV v3.1. Standardization and state-space parameters were estimated using 2248 training stays and then held fixed for 750 test stays. Under the fitted model, posterior covariance increased with observation-gap duration even when the model-estimated posterior mean changed little during the interval. Because no measurements are available within a gap, limited movement of the posterior mean should not be interpreted as evidence of actual physiologic stability. The two-dimensional model improved held-out predictive fit relative to a one-dimensional baseline, and stay-level summaries of the inferred four-channel latent trajectories were retrospectively associated with in-hospital mortality. The framework provides a reproducible method for characterizing uncertainty propagation under irregular observation timing and highlights the need to distinguish numerical persistence of a model estimate from well-supported clinical inference.