New Diffusion Model Generates Realistic Irregular Clinical Time Series

Clinical time series data, particularly laboratory test results in electronic health records, are inherently irregular. The absence of a test, or "missingness," is often not random but carries significant information, reflecting clinical decisions and patient physiological states. Traditional methods often treat this missingness as a preprocessing issue, overlooking its diagnostic value. A new diffusion-based approach has been introduced to address this by jointly modeling both the laboratory values and their observation patterns. This method extends the existing TimeDiff framework, learning continuous lab values alongside discrete missingness indicators. The data is structured into 4-hour intervals and 7-day admission windows to preserve realistic sampling. Experiments using the DACMI benchmark, derived from MIMIC-III, show that the generated synthetic data closely mirrors real patient trajectories. The model successfully captures the clinically meaningful relationships between patient physiology and the patterns of clinician testing behavior, especially under conditions where data is missing-not-at-random. These findings suggest the model's potential as a foundational component for future clinical AI models that can leverage informative missingness.