AI-Driven Real-Time Hyperlactatemia Prediction in ICU: A Multi-Cohort International Retrospective Study with External...

Area of Science:

• Critical care medicine
• Biomarker monitoring
• Machine learning applications in healthcare

Background:

• Blood lactate is a critical indicator of tissue hypoperfusion and metabolic distress in critically ill patients.
• Early identification of elevated lactate is crucial for timely intervention.
• Existing predictive tools are limited to specific intensive care unit (ICU) populations and lack real-time capabilities.

Purpose of the Study:

• To develop and validate machine-learning models for real-time, hourly prediction of hyperlactatemia (lactate >2 mmol/L).
• To provide 6- and 12-hour predictive horizons for hyperlactatemia in general ICU populations.
• To enable earlier recognition of tissue hypoperfusion and metabolic shock.

Main Methods:

• Developed two machine-learning models (XGBoost) using routinely collected vital signs and laboratory data.
• Utilized retrospective, multi-cohort data from AmsterdamUMCdb, MIMIC-III, eICU, and HiRID for development and external validation.
• Assessed model performance using discrimination (AUROC/AUPR), calibration, decision-curve utility, and subgroup fairness.

Main Results:

• The 6-hour model demonstrated strong internal and external validation performance (AUROC ≥0.772).
• The 12-hour model also showed consistent performance across all cohorts (AUROC ≥0.76).
• Calibration and decision-curve analyses confirmed robust generalization and clinical utility without recalibration.

Conclusions:

• Validated, continuously operating machine-learning models provide real-time, near-future lactate risk assessment in ICUs.
• These models support earlier recognition of tissue hypoperfusion and metabolic shock in unselected critically ill patients.
• Enables proactive clinical decision-making for improved patient outcomes.