AI Predicts Septic Shock Stabilization After Vasopressors

Key Takeaways

• Among US ICU adults receiving vasopressors who met an operational septic shock phenotype, stabilization occurred in 42% in MIMIC-IV and 33% in eICU.
• The prediction model transported to eICU with moderate discrimination but did not consistently outperform simpler hemodynamic benchmarks; early MAP trajectory contributed substantially.
• Non-stabilization was associated with higher ICU and hospital mortality in both cohorts, an observational finding rather than an effect of model-guided care.

In Lu and colleagues' Frontiers in Medicine prediction-model study, the US Medical Information Mart for Intensive Care IV (MIMIC-IV) cohort included 3,445 adults for retrospective development and held-out model comparison. An independent, multicenter US eICU Collaborative Research Database (eICU) cohort included 453 adults for external validation. Eligible first ICU stays involved sepsis, vasopressor initiation, elevated baseline lactate, and the required measurements and continued ICU stay through the prediction landmark.

First vasopressor initiation anchored the timeline. Observations from hours 0–6 supplied predictors at the landmark; hours 6–24 supplied the subsequent outcome. Patients who died or left the ICU before the landmark were not eligible for prediction.

Stabilization required mean MAP of at least 65 mmHg and at least 80% of MAP readings meeting that target during the outcome window. Patients also could have neither additional vasopressor classes nor a higher mean norepinephrine-equivalent dose (NEE) than in the predictor window. Failure of any component, or death after the landmark during the outcome window, counted as non-stabilization.

Prespecified early predictors covered MAP, lactate dynamics, vasopressor exposure, vital signs, laboratory markers, and organ support. A qualifying repeat lactate and complete baseline white blood cell and platelet results restricted the main cohort; remaining missing predictor values were imputed using training-set values. Investigators compared algorithms using training-only preprocessing and cross-validation. A gradient boosting machine (GBM) served as a representative nonlinear model, not a demonstrated winner over other leading algorithms. Its preprocessing, fitted model, and classification threshold selected from cross-validated training predictions were locked before eICU evaluation.

In the held-out MIMIC-IV cohort, the GBM's area under the receiver operating characteristic curve (AUC) was 0.720 (95% confidence interval [CI] 0.688–0.751); in independent eICU validation, AUC was 0.730. The held-out cohort also informed algorithm comparison, so it was not an independent final performance test. The model was not retrained for external evaluation.

The GBM's Brier score was 0.209 in MIMIC-IV and 0.192 in eICU, describing prediction error rather than clinical benefit. ICU and hospital mortality were higher without stabilization in both cohorts.

At the unchanged training-derived threshold, external sensitivity was 0.466 and specificity was 0.797. Held-out discrimination closely matched logistic regression; external performance did not consistently exceed prespecified early-MAP and MAP–NEE benchmarks. Shapley additive explanations (SHAP) attributed predictions substantially to early MAP stability, alongside lactate clearance, platelet counts, renal and inflammatory measures, and vasopressor intensity.

Early MAP and vasopressor inputs conceptually overlap with later MAP and support requirements in the endpoint, even though the measurement windows do not overlap. Transport between databases involved differing sepsis identification, a smaller external cohort, and distinct dose aggregation: time-weighted infusion intervals in MIMIC-IV versus recorded offsets in eICU. Repeat-lactate and laboratory availability, together with survival to the landmark, further selected the analyzed population. Hourly-binned MAP and reduced-predictor analyses probed endpoint definition and predictor proximity. SHAP attributions and the mortality association do not establish treatment effects; the composite does not measure pharmacologic vasopressor responsiveness.

The authors call for prospective validation before implementation, including assessment of whether the locked risk model adds practically useful information beyond simpler early hemodynamic measures. Model-guided care was not tested prospectively for effects on mortality, vasopressor use, or other patient outcomes.

Clinician Questions

Did the vasopressor-treated septic shock cohorts meet every clinical Sepsis-3 shock criterion?

Eligibility combined sepsis, vasopressor treatment, and baseline lactate greater than 2 mmol/L. Adequacy of preceding fluid resuscitation could not be reconstructed from the retrospective databases, so investigators did not claim exact replication of the full clinical Sepsis-3 shock definition.

How was early lactate clearance calculated for predicting stabilization after vasopressors in septic shock?

Baseline lactate was the closest measurement from six hours before to one hour after first vasopressor initiation. Follow-up lactate had to be measured at least two hours after baseline and no later than six hours after initiation. Clearance was the baseline-to-follow-up lactate difference divided by baseline lactate, expressed as a percentage; negative values were retained.