Development and external validation of prediction models for a trial-informed operational composite after stage 3 acute kidney injury

Abstract

Renal replacement therapy (RRT) is considered when severe acute kidney injury (AKI) is complicated by severe electrolyte disturbance, acid-base derangement, fluid overload with respiratory compromise, or azotaemia. However, actual RRT initiation is influenced by clinician and centre practice. We developed and externally validated machine-learning models to predict subsequent fulfilment of a trial-informed operational composite within 24–72 h after creatinine- or urine-output-defined stage 3 AKI. The prediction time was 24 h after stage 3 AKI onset. At this landmark, eligible patients remained alive in the ICU and had not received RRT or met any composite criterion. Predictors were routinely collected during the first 24 h. The composite comprised potassium > 6.0 mmol/L; pH ≤ 7.20 with bicarbonate ≤ 12 mmol/L within ± 2 h; fluid balance > 10% of cleaned body weight with PaO₂/FiO₂ ≤200; or blood urea nitrogen > 112 mg/dL. MIMIC-IV included 2,457 patients (238 events, 9.7%) and eICU-CRD 3,228 (145 events, 4.5%). Gradient Boosting achieved internal/external AUROCs of 0.778/0.721, AUPRCs of 0.279/0.131, and Brier scores of 0.079/0.053. External mean predicted risk was 11.8% versus an observed 4.5%; L2-penalized logistic regression and ExtraTrees performed similarly. The models provided moderate discrimination but substantial external overprediction and should not guide RRT initiation.