CISepsis: a causal inference framework for early sepsis detection.

Area of Science:

• Medical Informatics
• Machine Learning
• Causal Inference

Background:

• Early sepsis prediction models often use structured electronic medical record data.
• Sepsis involves complex physiological interactions, leading to mixed data types and confounding factors.
• Confounders in data can mask true causal relationships, reducing model generalizability and interpretability.

Purpose of the Study:

• To develop an early sepsis prediction approach that removes confounding effects and captures causal relationships.
• To enhance model generalizability, robustness, and interpretability in sepsis prediction.

Main Methods:

• Proposed a causal inference approach (CISepsis) to identify and remove confounding effects.
• Constructed a causal structure diagram to analyze relationships between observations, confounders, and labels.
• Employed back-door adjustment and instrumental variable methods, optimizing mutual information to eliminate confounder influence.

Main Results:

• CISepsis demonstrated significant improvements in Area Under the Curve (AUC) compared to XGBoost, LSTM, and MGP-AttTCN on the MIMIC-IV dataset.
• Achieved high AUC values (0.921-0.926) across multiple prediction time points.
• Ablation experiments confirmed the effectiveness of the proposed method.

Conclusions:

• Causal inference effectively removes confounding factors, enhancing the accuracy of early sepsis prediction.
• CISepsis offers improved generalizability, robustness, and interpretability over traditional methods.
• Future work includes exploring counterfactual adjustments for clinical applications and intervention analysis.