Dynamic Trajectories of Sepsis-Induced Coagulopathy and Machine Learning-Based Early Prediction of High-Risk Phenotypes: A Multicenter Cohort Study

Technical Summary: Dynamic Trajectories of Sepsis-Induced Coagulopathy and Machine Learning-Based Early Prediction

Problem Statement

Sepsis-induced coagulopathy (SIC) is a dynamic process involving inflammation, endothelial injury, and impaired fibrinolysis, yet it is currently assessed using static, single-time-point measurements (the SIC score). This approach fails to capture the rapid evolution of coagulation and organ dysfunction during critical illness. Patients with similar baseline SIC scores may follow divergent clinical courses—rapid improvement, persistent abnormality, or gradual resolution—which are indistinguishable using cross-sectional data alone. The study addresses the need to identify reproducible longitudinal SIC-score trajectories and determine their prognostic relevance across heterogeneous critical-care populations.

Methodology

This multicenter retrospective cohort study utilized four distinct critical-care databases:

• Development Cohort: MIMIC-IV (6,416 patients).
• Validation Cohorts: MIMIC-III CareVue (1,157 patients), eICU-CRD (2,077 patients), and NWICU (641 patients).

Data Processing and Modeling:

1. Cohort Selection: Adults (≥18 years) with ICU stays >24 hours meeting Sepsis-3 criteria and diagnosed with SIC (SIC score ≥4 within 24 hours of sepsis diagnosis) were included. Patients required at least three calculable SIC scores within the first 7 days of ICU admission.
2. Trajectory Modeling: Group-based trajectory modeling (GBTM) was applied to serial SIC scores (calculated at 6-hour intervals) over the first 7 days. Quadratic models with 1–6 latent classes were evaluated using Bayesian Information Criterion (BIC), entropy, average posterior probability (AvePP), and odds of correct classification (OCC).
3. Statistical Analysis:
   • Survival Analysis: Associations with 28-day all-cause mortality were assessed using multivariable Cox proportional-hazards models (adjusted for demographics, comorbidities, vital signs, and labs).
   • Bias Mitigation: Analyses included multiple imputation for missing data, calibrated inverse probability of treatment weighting (IPTW) to balance covariates, and a day-7 landmark analysis to address immortal-time bias (restricting analysis to patients alive at day 7).
   • Subgroup Analysis: Interaction terms were tested across strata (e.g., age, SOFA score, comorbidities) with false-discovery-rate (FDR) correction.
4. Machine Learning (ML) Prediction: A nested ML pipeline was developed to predict the "Persistently High" trajectory phenotype using baseline variables available within the first 24 hours.
   • Feature Selection: Boruta and LASSO methods were used within a 10-fold nested cross-validation framework.
   • Algorithms: Elastic net, random forest, XGBoost, SVM, and k-nearest neighbors were compared.
   • Validation: The final model (XGBoost) underwent temporal validation (MIMIC-IV 2020–2022) and external validation (MIMIC-III, eICU, NWICU). Performance was evaluated via AUROC, AUPRC, calibration curves, decision-curve analysis, and SHAP (Shapley Additive exPlanations).

Key Results

1. Identification of Three Trajectories:
GBTM identified three distinct, reproducible SIC-score trajectories across all four databases:

• Low-Rapid Decline (19.25%): Started with a median SIC of 4.0, declined rapidly, and stabilized.
• Persistently High (32.00%): Started with a median SIC of 6.0 and remained elevated with minimal improvement.
• Moderate-Decline (48.75%): Started at a median SIC of 5.0 and declined gradually.
  The three-class model demonstrated high classification quality (Entropy = 0.947; AvePP > 0.97).

2. Prognostic Associations:

• Mortality Risk: The "Persistently High" trajectory was consistently associated with the highest 28-day mortality risk across all cohorts. In fully adjusted models, the hazard ratio (HR) for Class 2 vs. Class 1 ranged from 2.45 (eICU) to 3.93 (NWICU).
• Robustness: These associations remained significant after IPTW calibration and in the day-7 landmark analysis (e.g., HR 3.09 in MIMIC-IV for Class 2).
• Incremental Value: Adding trajectory class to the baseline SOFA score modestly improved discrimination for 28-day mortality (e.g., ΔAUC = 0.028 in MIMIC-IV), whereas adding a single baseline SIC score provided negligible improvement.

3. Machine Learning Performance:

• Predictors: The XGBoost model identified 10 baseline variables, with baseline SOFA score and the SIC platelet component being the most influential predictors.
• Discrimination: The model achieved AUROCs between 0.830 and 0.862 across temporal and external validation cohorts.
• Calibration & Utility: The model showed acceptable calibration (Brier scores 0.141–0.154) and positive net benefit in decision-curve analysis across clinically relevant thresholds. SHAP analysis confirmed consistent feature importance rankings across cohorts.

Significance and Claims

The paper claims that longitudinal SIC measurements during the first week of ICU admission reveal reproducible prognostic subphenotypes that static assessments miss. Specifically:

• Prognostic Subphenotypes: The "Persistently High" trajectory identifies a distinct group of patients with sustained excess mortality risk, while "Low-Rapid Decline" marks a favorable course.
• Temporal Dimension: Trajectory information adds a modest but meaningful temporal dimension to standard severity assessments (SOFA), distinguishing patients whose coagulation dysfunction resolves from those with persistent abnormalities.
• Early Prediction: The study demonstrates that the "Persistently High" phenotype can be predicted early (within 24 hours) using routinely available clinical data via machine learning, with high generalizability across different institutions and time periods.

Limitations and Cautions:
The authors explicitly state that these are observational findings and do not establish causality or define specific treatment protocols. The study does not claim that trajectory-informed interventions improve outcomes; rather, it proposes these trajectories as targets for future prospective validation. The authors caution that the retrospective design leaves room for residual confounding and that the inclusion criteria (requiring multiple SIC scores) may introduce selection bias toward patients with longer ICU stays. The findings require prospective validation before clinical implementation.