72-hour hemodynamic and urine output trajectories predict renal recovery in continuous kidney replacement therapy: a machine learning approach

Predicting renal recovery in patients receiving continuous kidney replacement therapy (CKRT) remains challenging. Conventional single-time-point hemodynamic assessments at CKRT initiation have shown limited prognostic value. We investigated whether 72-hour hemodynamic and urine output (UO) trajectories, analyzed through machine learning, could improve prediction of 28-day renal recovery. This single-center retrospective cohort study included 331 CKRT episodes (318 patients). Renal recovery was defined as successful CKRT discontinuation with no kidney replacement therapy for at least 72 h within 28 days of initiation. Phase 1 evaluated day-0 variables with Cox and Fine-Gray models. Phase 2 developed machine learning models (logistic regression, XGBoost, bidirectional LSTM) using 104 engineered features from 72-hour MAP, SBP, DBP, heart rate (HR), and UO time-series data. Performance was assessed by 5-fold stratified cross-validation with bootstrap 95% confidence intervals, with SHAP analysis for interpretability. Of 331 episodes, 71 (21.5%) achieved renal recovery. In Phase 1, day-0 UO was non-significant in both univariable (HR = 1.010 per 100 mL/day; p = 0.135) and multivariable Cox regression (HR = 1.004 per 100 mL/day; p = 0.573) and in a Fine-Gray competing-risks model (sHR = 1.006; p = 0.385); CCI (HR = 0.715) and SOFA (HR = 0.783) were the strongest predictors (C-index = 0.776). In Phase 2, XGBoost Combined (day-0 + 72 h features) achieved an AUC of 0.848 (95% CI, 0.765–0.890), versus 0.762 for baseline-only XGBoost (DeLong p = 0.035). SHAP analysis identified late-phase (36–72 h) UO variability (UO_late_std) as the dominant predictor. Because 37 recoveries and 70 deaths occurred within 72 h, a landmark sensitivity analysis was restricted to 224 episodes still on CKRT (34 recoveries): UO_late_std retained discrimination (AUC 0.742) and a parsimonious model (SOFA, CCI, age, UO_late_std) reached 0.801, whereas the combined (0.682) and day-0 (0.690) models could not be distinguished at this sample size. Day-0 hemodynamic variables fail to predict renal recovery, whereas late-phase UO variability over the first 72 h is a strong predictor. This signal persisted among patients still on CKRT at 72 h, although the landmark cohort was underpowered to assess the incremental value of the full model. Temporal UO patterns warrant prospective validation.