First-day lactate trajectories and subsequent in-hospital mortality in cardiogenic shock: a landmark multicohort...

Lactate in cardiogenic shock is commonly summarized by a single measurement or a two-point estimate of clearance, although its first-day pattern may offer a more interpretable account of early clinical evolution. We examined whether first-day lactate trajectories identify reproducible patterns associated with subsequent mortality.

This retrospective multicohort study used MIMIC-IV for trajectory discovery and internal prediction modeling and eICU-CRD to evaluate transportability. Adults with a first ICU admission for cardiogenic shock and at least two lactate measurements within 24 h were eligible for trajectory construction. Primary analyses were restricted to patients alive and still hospitalized at a 24-hour landmark.

Four patterns were identified among 2,181 eligible MIMIC-IV patients: low-stable, moderate-decreasing, high-decreasing, and persistent-high. The 24-hour landmark risk set included 2,015 patients, with subsequent mortality ranging from 24.34% to 72.88%. Compared with low-stable, adjusted odds ratios were 2.08 [95% confidence interval (CI) 1.56-2.77] for high-decreasing and 5.81 (95% CI 3.57-9.46) for persistent-high. After centering each trajectory at its initial window, the increasing pattern remained associated with mortality after adjustment for initial lactate and clinical covariates (odds ratio 2.91, 95% CI 1.75-4.85, vs. the largest-decrease pattern). In foldwise cross-validation, the trajectory model had an area under the receiver operating characteristic curve (AUROC) of 0.724 and an area under the precision-recall curve (AUPRC) of 0.572. Last lactate and initial plus last lactate performed better (AUROC 0.734 and 0.736; AUPRC 0.592 and 0.594). Among 429 landmark-eligible eICU-CRD patients assigned to fixed MIMIC-IV centroids, mortality ranged from 25.49% to 78.12%; in 366 complete cases, the adjusted risk ratio for group 4 was 2.30 (95% CI 1.47-3.62).

First-day lactate trajectories provided a clinically interpretable framework for characterizing lactate burden and evolution after repeated testing. The low-stable and persistent-high extremes showed the most consistent cross-database separation, whereas the four-group partition should be interpreted as a descriptive continuum rather than four validated clinical risk classes. Simpler last-lactate summaries provided better discrimination when prediction alone was the objective.