Development of a predictive model for sepsis‑induced immunosuppression using neural networks

Technical Summary: Development of a Predictive Model for Sepsis-Induced Immunosuppression Using Neural Networks

Problem Statement
Sepsis is characterized by a dysregulated host response to infection, leading to a dual state of hyperimmunity and hypoimmunity. When the hypoimmunity state dominates, it results in sepsis-induced immunosuppression, a condition associated with secondary infections, multiple organ dysfunction, and increased mortality. While the clinical community recognizes the importance of immune homeostasis, accurately predicting the onset of immunosuppression remains in an exploratory stage. Current markers, such as monocyte HLA-DR (mHLA-DR), are often too complex or expensive for large-scale database analysis. Conversely, lymphocyte counts are more accessible but lack a clearly defined threshold for monitoring in existing literature. The study aims to address the need for early, accurate prediction of immunosuppression in sepsis patients to facilitate precision treatment and dynamic prognosis assessment.

Methodology
The study utilized a retrospective cohort design using data from the MIMIC-IV database.

• Data Source & Cohort: The analysis included 6,954 adult patients (≥18 years) admitted to the ICU for the first time with a Sepsis-3.0 diagnosis.
• Outcome Definition: Immunosuppression was defined as a lymphocyte count < 1.0 × 10⁹/L.
• Predictive Variables: The study collected 14 variables, including demographic data (age, gender, ethnicity), clinical metrics (BMI, SOFA score, Charlson comorbidity index, maximum lactate value), and treatment factors (emergency admission, vasoactive drug use/duration, mechanical ventilation/duration, corticoid use, fungal infection, and malignant history).
• Data Processing: Missing data were handled via multiple imputation (specifically for lactate data, which had a 22.16% missing rate). The BMI variable was excluded due to a 51.13% missing rate. Data were split into training (80%) and validation (20%) sets.
• Modeling Approaches: Three machine learning algorithms were constructed and compared:
  1. Multilayer Perceptron (MLP): A neural network approach.
  2. Logistic Regression (LR): A traditional statistical baseline.
  3. XGBoost: A gradient boosting decision tree model.
• Evaluation Metrics: Model performance was assessed using Area Under the Curve (AUC), Precision, Recall, F1 Score, calibration curves (Hosmer-Lemeshow test), and Decision Curve Analysis (DCA).

Key Results

• Cohort Characteristics: The incidence of immunosuppression in the cohort was 28.5% (1,983 out of 6,954 patients).
• Univariate Analysis: Significant associations were found between immunosuppression and age, gender, emergency admission, SOFA score, Charlson comorbidity index, fungal infection, vasoactive drug use, duration of vasoactive drugs, mechanical ventilation, and lactate levels.
• Multivariate Analysis: The final logistic regression model identified eight independent predictors: age, emergency admission, Charlson comorbidity index, fungal infection, vasoactive drug use, total vasoactive duration, mechanical ventilation, and lactate levels. Notably, gender, SOFA score, malignant tumor, and mechanical ventilation duration were not statistically significant in the multivariate logistic model, though they remained relevant in the machine learning context.
• Model Performance:
  • MLP Model: Demonstrated superior performance with an AUC of 0.82, Precision of 0.7208, Recall of 0.8441, and F1 Score of 0.7776.
  • XGBoost Model: Performed poorly with an AUC of 0.602, a Recall of 0.113, and an F1 Score of 0.177.
  • Logistic Regression: Showed moderate performance with an AUC of 0.686.
• Feature Importance: In the MLP model, the most influential factors for predicting immunosuppression were, in order: emergency admission, SOFA score, age, maximum lactate value, mechanical ventilation duration, total vasoactive duration, Charlson comorbidity index, gender, malignancy, race, corticoid use, and fungal infection.

Significance and Claims
The authors claim that the developed MLP model serves as a reliable tool for the early identification of sepsis-induced immunosuppression risk. The study posits that the model's ability to integrate diverse clinical variables allows for:

1. Early Intervention: Identifying high-risk patients to enable timely adjustments in treatment, such as the use of immune enhancers (e.g., thymosin peptides, granulocyte colony-stimulating factors).
2. Precision Medicine: Facilitating individualized treatment plans, such as avoiding the overuse of immunosuppressants or optimizing the timing of hormone discontinuation.
3. Resource Optimization: Improving the assessment of patient prognosis to better allocate intensive care resources.

The paper concludes that while the model shows potential clinical value, it requires further validation in multi-center and large-sample datasets to improve accuracy and universality before widespread clinical implementation. The authors acknowledge limitations inherent to retrospective studies, including selection bias and the exclusion of variables like BMI due to data loss, noting that causal relationships require further verification.