Landmark-specific phenotypic heterogeneity in CRRT discontinuation timing: a multi-landmark target trial emulation using MIMIC-IV

The optimal timing for discontinuing continuous renal replacement therapy (CRRT) in critically ill patients remains uncertain. Observational studies commonly simplify discontinuation timing into "early" vs. "delayed" categories, yet the same label may correspond to different clinical states at different stages of illness. We evaluated whether the clinical meaning of the early-discontinuation label and its association with 30-day mortality varied across landmarks and analytic frameworks.

Using MIMIC-IV (2008-2019), we emulated a target trial among adults who were alive and still receiving CRRT 48 h after initiation. Early discontinuation was defined as stopping CRRT within 24 h after the 48-hour landmark, whereas delayed discontinuation was defined as continuing CRRT beyond 24 h. The primary outcome was 30-day all-cause mortality. Analyses included crude Cox regression, baseline IPTW-Cox, multi-landmark analyses at 48, 72, 96, and 120 h, a marginal structural model (MSM) using time-varying IPTW, and clone-censor-weighting (CCW) with stabilized censoring weights under a 24-hour grace period. Phenotype decomposition was also performed to assess the clinical composition of patients classified as having early discontinuation at different landmarks.

A total of 1,432 patients were included, of whom 369 (25.8%) met the definition of early discontinuation. Baseline IPTW-Cox suggested that early discontinuation was associated with higher 30-day mortality (HR 1.36, 95% CI 1.13-1.64; P = 0.001). Multi-landmark HRs ranged from 1.13 to 1.35 without a monotonic pattern; precision decreased at later landmarks, and the landmark-by-exposure interaction was not statistically significant. The MSM estimate was OR 1.16 (95% CI 0.82-1.63; P = 0.40), and the revised CCW estimate was HR 1.08 (95% CI 0.91-1.28; P = 0.38). Phenotype decomposition showed that sepsis prevalence increased from 55.6% at 48 h to 69.1% at 120 h, baseline vasopressor exposure increased from 55.8% to 71.9%, and urine output decreased from 818 mL to 524 mL.

Baseline IPTW generated a significant signal suggesting harm from early CRRT discontinuation, but this was not reproduced in multi-landmark, MSM, or CCW analyses. Phenotype decomposition indicated that "early weaning" captures different clinical profiles at different time points rather than a stable, homogeneous exposure. Observational CRRT weaning studies should consider dynamic treatment strategies, time-varying causal methods, and clinically grounded outcome definitions.