Bridging Deterministic and Probabilistic Deep Learning Imputation via Post-Hoc Uncertainty Quantification
Deep learning models such as BRITS and SAITS have established strong benchmarks for missing value imputation in critical care data. However, these models are deterministic, i.e., they provide no mechanism for uncertainty quantification (UQ). This limits the applicability of these methods in clinical settings, where uncertainty stemming from measurement error, signal noise, and the use of surrogate markers is intrinsic to the data-generating process. On the other hand, probabilistic approaches, such as variational autoencoders (VAEs), incorporate UQ by design, but their reliance on fixed distributional priors constrains their ability to model the heterogeneous structure of medical time series, leading to comparatively higher imputation error. We address this trade-off by applying model-agnostic post hoc UQ methods, specifically Monte Carlo Dropout and Last-Layer Laplace Approximation, to high-performing deterministic imputation models. This formulation decouples imputation accuracy from uncertainty estimation, retaining the predictive performance of state-of-the-art models while producing well-calibrated confidence estimates suitable for downstream clinical decision-making.