Development and External Validation of a 72-Hour Dynamic Trajectory-Based Machine Learning Model for Mortality Prediction in ICU Patients with Sepsis and Pre-Existing Hypertension

Technical Summary: 72-Hour Dynamic Trajectory-Based ML Model for Sepsis Mortality in Hypertensive ICU Patients

Problem Statement
Sepsis remains a leading cause of mortality worldwide, with management complicated by the presence of pre-existing hypertension. Patients with chronic hypertension exhibit altered vascular autoregulation and reduced vascular compliance, potentially requiring higher perfusion pressures to maintain organ blood flow compared to non-hypertensive patients. While guidelines suggest a mean arterial pressure (MAP) target of ≥65 mmHg, the optimal target for hypertensive septic patients remains uncertain, with lower targets potentially linked to acute kidney injury. Furthermore, conventional severity scores (e.g., SOFA, SAPS II) rely largely on static baseline variables and may fail to capture the evolving physiological trajectories and heterogeneous treatment responses during the critical early resuscitation phase in the Intensive Care Unit (ICU). There is a lack of prognostic tools specifically designed to integrate early dynamic hemodynamic and perfusion data for this specific high-risk population.

Methodology
This study employed a retrospective cohort design utilizing two publicly available critical care databases: MIMIC-IV (development cohort) and eICU-CRD (external validation cohort).

• Cohort Selection: The study focused on adult patients (18–89 years) with sepsis and pre-existing hypertension. Inclusion criteria required a first ICU admission, an ICU length of stay ≥72 hours, and the use of vasoactive agents within the first 72 hours. Patients with >20% missingness in candidate predictors were excluded.
• Feature Engineering:
  • Static Variables: Demographics, comorbidities, initial vital signs, laboratory results, and severity scores recorded within the first 72 hours.
  • Dynamic Variables: Continuous and repeated measurements were reorganized into six consecutive 12-hour time windows within the first 72 hours. For each window, time-series features were extracted, including the first value, last value, mean, slope, standard deviation, and change amplitude.
  • Preprocessing: Missing values were handled using Multiple Imputation by Chained Equations (MICE). SOFA and SAPS II were excluded from the feature selection process to serve solely as benchmark comparators.
• Feature Selection: A two-step approach was applied:
  1. Boruta Algorithm: Used for broad screening to identify variables associated with in-hospital mortality.
  2. LASSO Regression: Applied to the Boruta-selected variables to further reduce dimensionality and select the final feature set.
• Model Development: Seven machine learning algorithms were trained and compared using 5-fold cross-validation: Logistic Regression, Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest, Backpropagation Neural Network (BPNN), LightGBM, and XGBoost. The best-performing model underwent sigmoid probability calibration.
• Evaluation Metrics: Performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), Brier score, calibration slope, and Decision Curve Analysis (DCA). SHAP (Shapley Additive Explanations) values were used for model interpretability.

Key Results

• Cohort Characteristics: The development cohort (MIMIC-IV) included 1,215 patients (403 deaths, 33.2% mortality). The external validation cohort (eICU-CRD) included 1,747 patients (516 deaths, 29.5% mortality). Non-survivors in the development cohort exhibited higher rates of AKI, mechanical ventilation, and vasoactive agent use, along with higher lactate and norepinephrine requirements.
• Model Performance (Internal Validation):
  • XGBoost_Sigmoid (calibrated XGBoost) achieved the best performance with an AUC of 0.7819 ± 0.0493, accuracy of 0.7364, and specificity of 0.8758.
  • The pooled out-of-fold AUC for XGBoost_Sigmoid was 0.778, outperforming SOFA (0.540) and SAPS II (0.638).
  • Calibration analysis showed a slope of 1.0490 and a Brier score of 0.1783.
• External Validation:
  • In the eICU-CRD cohort, the model achieved an AUC of 0.713 with a Brier score of 0.1854 and a calibration slope of 0.7440.
  • Decision Curve Analysis indicated a net benefit over "treat-none" strategies across threshold probabilities of 0.05–0.30.
  • Risk stratification demonstrated a stepwise increase in observed mortality: 13.7% (low risk), 27.5% (medium risk), and 47.3% (high risk).
• Feature Importance (SHAP Analysis): The most influential predictors identified were Glasgow Coma Scale (GCS), mean urine output, mean heart rate, last norepinephrine requirement, age, Charlson index, last urine output, platelet count, mean lactate, and base excess slope. Lower GCS, lower urine output, higher norepinephrine requirements, older age, and higher lactate were associated with increased mortality risk.

Significance and Claims
The authors claim that this study successfully developed and externally validated a 72-hour landmark machine learning model specifically for ICU patients with sepsis and pre-existing hypertension. The primary significance of the work lies in:

1. Integration of Dynamic Trajectories: Unlike conventional scores that rely on static snapshots, this model incorporates early dynamic hemodynamic and perfusion trends (over 72 hours), capturing the patient's response to resuscitation.
2. Superior Discrimination: The XGBoost_Sigmoid model demonstrated better discrimination than traditional severity scores (SOFA and SAPS II) in the internal validation cohort.
3. Clinical Utility for Reassessment: The authors posit that the model is not intended as an admission-only tool but rather as a mechanism for early risk reassessment after the initial resuscitation phase. It allows clinicians to stratify patients into low, medium, and high-risk groups based on their physiological trajectory, potentially aiding in the identification of patients who may require more aggressive management or closer monitoring.
4. Interpretability: Through SHAP analysis, the model provides transparent insights into key drivers of mortality (e.g., neurological status, renal perfusion, and vasopressor dependence), aligning with clinical reasoning regarding the pathophysiology of sepsis in hypertensive patients.

The study concludes that while the model retains moderate performance in external validation, further prospective validation and local recalibration are necessary before it can be implemented as a standard bedside clinical decision support tool.