CAIR Improves Physiological Time-Series Imputation for Clinical Data.

This research introduces CAIR, a novel two-stage framework designed to accurately impute missing data in physiological time series, such as arterial blood pressure or blood glucose. Traditional imputation methods often fail in clinical settings because they don't account for the unique characteristics of physiological missingness, where gaps can occur during extreme clinical events and vary widely in length. CAIR addresses these issues by first learning a coarse base curve using a bidirectional-GRU interpolator. This initial estimate is then repeatedly corrected by a Transformer refiner over three successive passes. The entire system is trained jointly using a broad, signal-agnostic random-gap curriculum. Evaluations demonstrate CAIR's superior accuracy across various missingness mechanisms (MCAR, MAR, NMAR) and gap lengths on datasets like continuous glucose monitoring (AI-READI) and intensive care arterial pressure (MIMIC-III). Notably, CAIR not only achieves low reconstruction error but also effectively preserves crucial clinical burden metrics, a challenge where other high-accuracy imputers often fall short.