Abstract
Background: Severe anaemia is common in critically ill patients, frequently necessitating urgent diagnostic evaluation, haemodynamic reassessment, and red-blood-cell transfusion. Current ICU practice heavily relies on intermittent haemoglobin (Hb) measurements, which inherently delay recognition of clinically relevant Hb decline. We aimed to develop and externally validate a machine-learning model to predict severe anaemia (Hb <8 g/dL) within the next 6 hours using only routinely collected ICU data.Methods; In this retrospective observational cohort study, we developed an Extreme Gradient Boosting model using the Vienna Anaesthesia and Intensive Care Database (ViennaAIdb; Medical University of Vienna, Austria) for ICU admissions between Jan 1, 2014, and Nov 30, 2021. The primary outcome was severe anaemia within a 6-hour prediction horizon. The cohort was split at the patient level into training and internal test sets. External validation was performed using the MIMIC-IV database (v3.1) via an identical preprocessing pipeline and entirely without model retraining. Model performance was assessed by discrimination, using the area under the receiver operating characteristic curve (AUC), and calibration, using calibration plots and Brier scores. Feature contributions were evaluated with SHapley Additive exPlanations.Findings: The development cohort included 21,731 ICU stays, contributing 3,321,230 hourly prediction samples; the external validation cohort included 821 ICU stays, contributing 11,217 samples. The model demonstrated high discrimination in both cohorts, with an AUC of 0·94 in internal testing and 0·91 in external validation. Calibration was excellent internally and remained robust externally, with close agreement between predicted and observed risks (Brier scores of 0·055 and 0·110, respectively). The most influential predictors of impending anaemia were recent Hb dynamics, cumulative urine output, heart rate, and fluid administration.Interpretation: A machine-learning model based exclusively on routinely available ICU variables accurately predicted severe anaemia within a clinically actionable 6-hour horizon, demonstrating robust discrimination and calibration during external validation without retraining. These findings suggest that automated, near real-time Hb risk stratification could support a shift from reactive laboratory surveillance towards anticipatory monitoring and earlier bedside reassessment. Prospective evaluation is warranted before clinical implementation.FundingNo specific funding was received.