Early creatinine dynamics for stratifying acute kidney injury trajectories in patients with systemic lupus erythematosus: an exploratory multicenter retrospective study.

Objective: Patients with systemic lupus erythematosus (SLE) exhibit heterogeneous renal courses after acute kidney injury (AKI). This study applied latent class mixed modeling to identify AKI trajectory phenotypes in SLE, developed a risk-stratification model using early post-AKI data, and explored whether hydroxychloroquine (HCQ) use is associated with AKI trajectory class membership. Methods: This multicenter retrospective study used MIMIC-IV for model development and internal assessment and eICU, NWICU, and local cohorts for exploratory external assessment. Adult SLE patients with AKI were included. The primary model was a race-free, no-imbalance-correction LASSO logistic regression model. Performance was assessed with AUC, Brier score, calibration intercept, and calibration slope. Bootstrap optimism correction was applied. A non-overlap sensitivity analysis examined the influence of structural predictor-outcome overlap. HCQ was evaluated using multivariable logistic regression and propensity score matching. Results: Among 279 MIMIC patients, three AKI trajectories were identified (progressive 9.3%, recovery 19.4%, stable 71.3%). The primary race-free LASSO model achieved an apparent AUC of 0.761 (optimism-corrected: 0.706) and a Brier score of 0.086 (corrected: 0.092). When assessed against a non-overlapping post-48-hour renal endpoint, discrimination diminished substantially (AUC 0.593). Exploratory external assessments were event-limited: the main SLE external discrimination subsets contained 6 events in eICU, 3 events in NWICU, and 1 event in the local cohort; therefore, these estimates were considered descriptive rather than definitive validation. HCQ showed an exploratory association not interpretable as causal. Conclusion: This study provides an exploratory risk-stratification framework for early AKI trajectory patterns in SLE. The apparent discrimination was substantially attenuated after removing predictor-outcome overlap and adjusting for optimism. External assessments were event-limited and should be interpreted as exploratory transportability assessments rather than definitive evidence of generalizability.