Abstract Rationale Acute kidney injury (AKI) is highly prevalent in critically ill patients and encompasses a heterogeneous spectrum of pathophysiological mechanisms. Beyond its severity at diagnosis, the temporal evolution of serum creatinine (SCr) during ICU stay carries valuable prognostic information. However, competing events such as death or early discharge may distort these trajectories, requiring analytical frameworks that jointly account for longitudinal dynamics and time-to-event processes. Latent joint class models (JLCM) offer a powerful approach to identify dynamic and prognostically coherent AKI phenotypes within such complex competing-risk settings, enabling prognostic enrichment by highlighting patients with adverse recovery trajectories. We aim to identify, externally validate AKI trajectory phenotypes and predict early membership to high-risk recovery profiles. Methods We conducted a retrospective, two-step study using the European AmsterdamUMC database for model development and the US MIMIC-IV database for external validation. Adult ICU stays with AKI 24 h defined by KDIGO SCr criteria were included. Longitudinal SCr trajectories over 14 days were modeled using JLCM linking creatinine evolution under competing risks context, evaluating multiple class solutions and retaining the optimal model based on statistical fit and clinical interpretability. A multiclass XGBoost model integrating dynamic 6h windows data up to day 7 was trained to predict class membership. Sensitivity analyses assessed the robustness of trajectories on all AKI. Results Among 3,126 AKI ICU stays in AmsterdamUMCdb, a five-class solution best captured SCr heterogeneity: rapid improvement (8%), slowly rising (28%), moderate decline (23%), sustained increase (36%), and incomplete recovery (23%). These trajectories showed distinct prognostic gradients. The rapid improvement group had the highest rate of live ICU discharge and lowest mortality, whereas the sustained increase and incomplete recovery classes concentrated most deaths and prolonged ICU stays. Intermediate classes showed slower or partial renal recovery and persistent organ support. External validation on MIMIC (n = 3,092) confirmed similar trajectory shapes and preserved outcome hierarchy across settings. The predictive model showed acceptable performance from day 3 onward (accuracy=0.78, balanced accuracy=0.75), mainly driven by SCr value and slope, with no actionable hemodynamic or therapeutic variables identified. Sensitivity analyses yielded consistent trajectories and prognostic rankings across datasets. Conclusion Dynamic modeling of SCr trajectories identifies five reproducible AKI phenotypes with coherent prognostic profiles across two distinct ICU databases. Early prediction within 72h may support future development of real-time, trajectory-based AKI stratification tools to improve clinical understanding and trial design. This abstract is funded by: None
Jamme et al. (Fri,) studied this question.