An early sepsis prediction model utilizing machine learning and unbalanced data processing in a clinical context.

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Sepsis Research

Background:

• Sepsis diagnosis in China relies on traditional methods, potentially delaying treatment.
• Early and accurate sepsis diagnosis is vital for improving patient outcomes and reducing mortality rates.

Purpose of the Study:

• To develop and validate a predictive model for early sepsis diagnosis.
• To identify key clinical features associated with sepsis development.

Main Methods:

• Utilized data from 2,385 patients (364 with sepsis) and external validation on MIMIC-III and eICU databases.
• Employed Random Forest and SHapley Additive exPlanations (SHAP) for model development and risk factor analysis.
• Applied data preprocessing techniques including Multiple Imputations and Synthetic Minority Oversampling (SMOTE).

Main Results:

• The Random Forest model achieved an Area Under the Curve (AUC) of 87% and an F1-score of 77%.
• Identified 18 diagnostic features for early sepsis prediction.
• SHAP analysis results align with current clinical understanding of sepsis risk factors.

Conclusions:

• Established the relationship between 18 clinical features and sepsis diagnosis.
• Systolic Blood Pressure, Albumin, and Heart Rate are significant indicators for predicting sepsis likelihood.