Early identification of gram-negative bacteremia in intensive care units (ICUs) remains challenging at the time of blood culture sampling, when clinical signs are often nonspecific and existing diagnostic approaches typically rely on single-timepoint measurements. We conducted a retrospective cohort study of adult ICU patients admitted between July 2022 and January 2024 to investigate whether short-term longitudinal patterns in routinely collected clinical and laboratory data contain diagnostically relevant information for gram-negative bacteremia. Clinical and laboratory variables were extracted at three consecutive timepoints (Day -2, -1, and 0 relative to blood culture collection), and diagnostic models incorporating this temporal information were developed using complementary statistical and machine-learning approaches. Model performance was evaluated on a held-out test set using discrimination, calibration, and decision curve analysis. Among 568 patients, models incorporating short-term longitudinal data demonstrated good and consistent discrimination for gram-negative bacteremia (AUC range 0.81-0.83) with good calibration after recalibration. Diagnostic performance was stable across modeling approaches, indicating robustness of the underlying signal rather than dependence on a specific algorithm. Decision curve analysis suggested higher net benefit for model-based risk stratification compared with treat-all or treat-none strategies across clinically relevant threshold probabilities. Hemoglobin, creatinine, and albumin consistently emerged as influential contributors. These findings indicate that short-term longitudinal clinical trajectories contain diagnostically meaningful information for gram-negative bacteremia at the time of blood culture sampling and support further external validation and prospective evaluation prior to clinical implementation.
Dilken et al. (Wed,) studied this question.