The Analysis of Sepsis Vascular Response Index Based on Machine Learning

Technical Summary: The Analysis of Sepsis Vascular Response Index Based on Machine Learning

Problem Statement
Sepsis and septic shock represent critical dysregulated body responses to infection, often leading to multi-organ failure and high mortality rates. A specific and challenging phenomenon in septic shock is "vascular paralysis," where patients exhibit poor vascular responsiveness to vasopressor drugs, resulting in persistent hypotension despite adequate fluid resuscitation. Currently, there is no unified clinical definition for the degree of vascular paralysis, and existing tools to identify its severity and prognosis are limited. Consequently, vascular paralysis is often detected only at terminal stages, leading to poor outcomes. While the Vascular Reactivity Index (VRI), defined as the peripheral vascular resistance index divided by the vascular muscle strength score (SVRI/VIS), has shown promise in pediatric populations, its utility in adult septic shock patients remains under-explored. This study addresses the need for new tools to monitor vascular paralysis severity and improve early prognostic assessment in adults.

Methodology
This retrospective study analyzed clinical data from 126 adult septic shock patients admitted to the Intensive Care Unit (ICU) of the Second Affiliated Hospital of Guilin Medical University between December 2017 and December 2022. All patients met the Sepsis 3.0 diagnostic criteria, underwent Pulse Indicator Continuous Cardiac Output (PICCO) monitoring, and received vasopressor or positive inotropic therapy.

• Data Preprocessing: Clinical variables included demographic data, vital signs, laboratory results (e.g., lactate, WBC, CRP), and hemodynamic parameters (e.g., MAP, SVV, ELWI, GEDI, SVRI). Missing values were imputed using the mean imputation method, outliers were handled, and data were standardized.
• Machine Learning Models: The primary analytical tool was the XGBoost (eXtreme Gradient Boosting) classifier. Hyperparameter optimization was conducted using the Optuna library to tune tree depth, learning rate, regularization parameters, and subsample ratios. The model's performance was validated using five-fold cross-validation.
• Comparative Analysis: The XGBoost model was benchmarked against Support Vector Machines (SVM) and Logistic Regression (LR). Additionally, a VRI-guided stacking ensemble model was proposed, integrating base learners (LR, SVM, Random Forest, XGBoost, LightGBM, CatBoost) with a meta-classifier to combine heterogeneous model strengths.
• Interpretability: SHAP (SHapley Additive exPlanations) values were employed to interpret model predictions, identifying feature importance and the direction of influence for specific clinical indicators.
• Clinician Comparison: Predictions from 11 ICU clinicians were collected and compared against the machine learning model's performance to evaluate human versus algorithmic prognostic accuracy.

Key Contributions

1. Application of XGBoost to VRI: The study applies advanced gradient boosting techniques to analyze the Vascular Reactivity Index in adult septic shock, moving beyond traditional linear analysis.
2. Identification of Key Predictors: Through SHAP analysis, the study identifies Mean Arterial Pressure (MAP), Stroke Volume Variation (SVV), Visual Index (VIS), and Extravascular Lung Water Index (ELWI) as the most critical physiological indicators influencing vascular response prediction.
3. Prognostic Thresholds: The research establishes a specific VRI threshold (< 28.2) associated with poor prognosis. Patients with a VRI below this threshold exhibited a 28-day mortality rate 3.857 times higher than those with VRI ≥ 28.2.
4. Model Superiority: The study demonstrates that the XGBoost model significantly outperforms both SVM and LR in terms of classification accuracy and Area Under the Curve (AUC) for predicting vascular response and mortality.
5. Human vs. Algorithm Comparison: The paper provides a comparative analysis showing that the machine learning model offers more stable and consistent prognostic assessments compared to the variable and heterogeneous predictions of individual clinicians.

Results

• Cohort Characteristics: The study included 126 patients (89 male, 37 female) with an average age of 60.61 years. The 28-day survival rate was 34.92% (44 survivors, 82 deaths).
• Model Performance: The XGBoost model achieved an AUC of 0.90, indicating high discriminative ability in distinguishing vascular response statuses. In contrast, the SVM model yielded an AUC of 0.55, and the LR model showed signs of underfitting with a flat prediction trend.
• Feature Importance: XGBoost feature ranking and SHAP plots confirmed that MAP, SVV, VIS, and ELWI were the primary drivers of the model's predictions. Higher APACHE II scores, SOFA scores, SVV, and VIS values correlated with higher mortality, as did lower 72-hour lactate clearance rates.
• Clinical Correlation: The combination of VRI with APACHE II and SOFA scores demonstrated higher clinical value and sensitivity in predicting prognosis compared to using these scores alone.
• Clinician Variability: Clinician predictions showed significant inter-rater variability (median predicted risk ranging from 0.005 to 0.750), whereas the machine learning model provided reproducible results.

Significance
The paper claims that the Vascular Reactivity Index (VRI) is a significant clinical tool for predicting the prognosis of adult septic shock patients. By utilizing machine learning, specifically XGBoost, the study suggests that VRI can serve as a quantitative indicator for assessing the severity of vascular paralysis earlier in the disease course. The findings indicate that a lower VRI level is associated with a higher mortality rate, and the integration of VRI with standard scoring systems (APACHE II, SOFA) enhances prognostic accuracy. The study concludes that strengthening the monitoring of hemodynamic indicators through such machine learning frameworks can reduce patient mortality and improve treatment guidance, particularly by identifying high-risk patients who may require more aggressive intervention before reaching the terminal stages of vascular paralysis.