Transportable early renal recovery phenotypes and 30-day major adverse kidney events after sepsis-associated acute kidney injury: A MIMIC-IV discovery and eICU-CRD external-validation cohort study
Early renal recovery after sepsis-associated acute kidney injury is heterogeneous. We aimed to derive clinically interpretable early recovery phenotypes, assess their transportability through fixed-centroid external assignment, and examine associations with 30-day major adverse kidney events (MAKE30). This retrospective discovery and external-validation cohort study used Medical Information Mart for Intensive Care IV and eICU-CRD. Eligible patients were adults with a first intensive care unit stay and Kidney Disease: Improving Global Outcomes -compatible acute kidney injury within 7 days of sepsis onset. K-means clustering integrated five renal, fluid, and hemodynamic domains across twelve 6-hour intervals after acute kidney injury onset. Stability was assessed by repeated 80% patient subsampling. eICU CRD patients were assigned to fixed MIMIC-derived centroids without reclustering. Modified Poisson models estimated adjusted risk ratios (RRs). Of 9725 Medical Information Mart for Intensive Care IV patients with known MAKE30 status, 2529 (26.0%) experienced MAKE30. Four phenotypes were identified in 7965 patients: low-burden rapid-recovery, preserved-diuresis intermediate-risk, oliguric kidney-failure, and shock-vasopressor. The four-cluster solution had a median adjusted Rand index of 0.939 (empirical 2.5th-97.5th percentiles, 0.892-0.990). Relative to the low-burden phenotype, adjusted MAKE30 RRs were 1.74 (95% confidence interval , 1.58-1.90), 3.80 (3.47-4.16), and 3.68 (3.34-4.06). In eICU-CRD, 4708 of 8771 patients (53.7%) could be assigned; corresponding RRs were 2.82 (2.45-3.23), 4.78 (4.00-5.72), and 4.99 (4.32-5.76). Unassigned patients had higher MAKE30 rates than assigned patients (41.1% vs 25.2%). MAKE30 associations persisted in landmark and distance-restricted analyses. Early renal recovery trajectories identified four reproducible clinical phenotypes associated with MAKE30. Fixed-centroid assignment reproduced the direction of these associations in eICU-CRD patients with sufficient trajectory data. Selective assignment and the observational design limit generalizability and preclude conclusions about phenotype-guided treatment.
Authors
- Na Li (ORCID: https://orcid.org/0000-0002-8754-2851)
- Lushan Wang
- Yang He (ORCID: https://orcid.org/0009-0003-7525-083X)
- Jinglan Liu
- Zhongyu Han
Institutions
- China Three Gorges University (CN)
- Yichang Central People's Hospital (CN)
Publication Details
- Journal
- Medicine
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1097/md.0000000000051571
- Primary Topic
- Acute Kidney Injury Research
- Type
- article
- Field-Weighted Citation Impact
- 0.00