Prospective validation of machine learning models predicting Gram negative bacteremia in ICU versus clinical scores

Bacteremia is a life-threatening infection requiring early diagnosis and timely treatment. Although machine learning (ML)–based models may support early clinical decision-making, their validity remains uncertain. We prospectively validated ML-based models developed at our center for predicting Gram-negative bacteremia and compared their performance with established clinical scores. This observational study included adult ICU patients hospitalized for ≥48 h who met Sepsis-3 criteria for suspected infection between February and July 2025. Clinical and laboratory data from the preceding 48 h were analyzed. Gram-negative bacteremia was identified in 97 of 289 blood cultures. Several clinical scores were significantly associated with bacteremia. While ML models demonstrated modest discrimination in cohort, performance improved to a moderate level (AUROC ≈ 0.70) with higher sensitivity in a post-hoc subgroup analysis of cultures obtained between ICU days 4 and 10. These findings highlight the importance of patient stratification and culture selection when validating predictive models in clinical settings.