Machine learning-derived multivariate renal function trajectories in acute kidney injury in critically ill patients: a multicentre retrospective study.

Background: Acute kidney injury (AKI) exhibits considerable heterogeneity. The objective of the current study was to identify AKI subphenotypes in intensive care unit (ICU) patients using multivariate renal function trajectories.

Methods: A retrospective study was performed on two independent datasets: the MIMIC-IV and eICU datasets. Using group-based multivariate trajectory modelling (GBMTM), we identified AKI subphenotypes based on trajectories of estimated glomerular filtration rate (eGFR) and urine output (UO). Multivariate Cox regression was applied to quantify the risk associated with the AKI subphenotype concerning clinical outcomes.

Results: Our study enrolled a total of 17 113 ICU patients diagnosed with AKI, revealing four AKI subphenotypes via GBMTM. Subphenotype 1, characterized by a swift decrease in eGFR coupled with a gradual increase in UO, exhibited the highest mortality rates (24% in the MIMIC-IV cohort, 20% in the eICU cohort). Conversely, subphenotype 4, featuring a marked increase in eGFR alongside stable UO levels, demonstrated the most favourable prognosis (15% mortality in the MIMIC-IV cohort, 8% in the eICU cohort). Subphenotypes 2 and 3 shared similar eGFR trends, however, subphenotype 2 experienced a rapid decrease in UO, whereas subphenotype 3 maintained stability in this regard. Across both datasets, subphenotype 4 showed a significantly reduced risk of mortality compared with subphenotype 1 {hazard ratio [HR] 0.72 [95% confidence interval (CI) 0.60-0.81],  < .001 in the MIMIC-IV cohort; HR 0.49 [95% CI 0.37-0.64],  < .001 in the eICU cohort}.

Conclusions: This study differentiated and stratified AKI subphenotypes among ICU patients by leveraging multivariate renal function trajectories, laying a foundation for personalized therapeutic strategies.