Temporal Heterogeneity of Factors Associated with Intra-ICU Mortality: A Formal Interaction Analysis of the Multi-Centre Silesian ICU Registry (2010-2020).

Background: Severity-of-illness scores and registry-derived risk models assume that the association between a baseline characteristic and death is stable over time. Methods: A retrospective analysis of 25,344 admissions (2010-2020) to 34 intensive care units in the Silesian Registry of Anaesthesiology and Intensive Care, with unchanged variable definitions and intra-ICU death as the outcome, was conducted. For each of 48 characteristics, a characteristic-by-time interaction was tested by likelihood ratio under four implementations of time-chronological quintiles, fixed calendar periods, linear and spline continuous time, each independently Benjamini-Hochberg corrected. Results: Twenty-two of 48 characteristics showed significant heterogeneity in the primary quintile analysis (q < 0.05), with 11 under all four. Adjustment for centre left 12 significant, 6 of which also survived adjustment for case mix. An adjusted model retained discrimination out of period but over-predicted risk in the final quintile. Conclusions: For nearly half of the characteristics, the crude association with intra-ICU death is not temporally stable, although most of the movement tracks a changing population. Out of period, a fixed model loses the level of predicted risk, not the weighting of its predictors-level rather than structural drift. Multi-year registry analyses should test temporal homogeneity rather than assume it.