Noradrenaline-trajectory phenotypes in septic shock: derivation and external validation in two independent cohorts.

BACKGROUND

Septic shock is heterogeneous, and noradrenaline (NA) requirements evolve over time in ways that reflect vascular responsiveness and shock biology. Whether long-horizon NA dose-trajectory phenotypes are reproducible across healthcare systems and identifiable early in the clinical course remains uncertain.

METHODS

We retrospectively analyzed 1111 adults with septic shock at Sheba Medical Center (Israel) and 9343 adults from MIMIC-IV. Hourly NA infusion trajectories were reconstructed for 10 days. A fully prespecified clustering pipeline combining static-feature K-means with dynamic time warping refinement was derived in Sheba and applied unchanged to MIMIC-IV. The primary outcome was 90-day mortality; 30-day mortality was analyzed as a supportive secondary outcome. Secondary analyses included feature interpretability, multivariable landmark Cox models (24-144 h), and early phenotype prediction using exposure features available at 24-96 h.

RESULTS

Five stable phenotypes emerged in Sheba and six in MIMIC-IV. Despite this numerical difference and somewhat more prolonged high-dose exposure in Sheba sustained/late-escalating patterns, both cohorts exhibited the same core families of trajectory shapes with preserved mortality gradients (90-day mortality 37-89% in Sheba; 16-69% in MIMIC-IV). Mortality gradients were preserved at both 30 and 90 days across cohorts. Across cohorts, exposure-persistence features-particularly time to maximal NA dose and cumulative NA burden during days 2-4-were key determinants of phenotype structure and mortality. Early-prediction models identified final phenotypes using 24-h data with 85% accuracy in Sheba and 86% in MIMIC-IV, improving by 48 h to 89% and 88%, respectively; high-confidence assignments achieved approximately 91%-95% accuracy.

CONCLUSIONS

Long-horizon NA dose-trajectory phenotypes are clinically interpretable hemodynamic exposure/response patterns, prognostically coherent, and externally reproducible. Their defining features were detectable within the first 24 h and showed stronger discrimination by 48 h. At present, these phenotypes should be interpreted as descriptive hemodynamic-response patterns that may support dynamic risk stratification, prognostic enrichment for future trials, and hypothesis generation; whether early phenotype identification can improve management or outcomes requires prospective evaluation.