[Constructing a predictive model for the death risk of patients with septic shock based on supervised machine...

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Septic shock poses a significant threat with high mortality rates.
• Accurate prediction of 28-day mortality is crucial for timely intervention.
• Existing predictive models may lack sufficient accuracy and discrimination.

Purpose of the Study:

• To develop and validate the optimal predictive model for 28-day mortality in septic shock patients.
• To compare the performance of various supervised machine learning algorithms.
• To identify key predictive variables for septic shock mortality.

Main Methods:

• Utilized data from 3,295 septic shock patients in the MIMIC-IV v2.0 database.
• Employed five machine learning algorithms: CART, RF, SVM, LR, and SL.
• Identified optimal predictive variables using LASSO regression, RF, and XGBoost.

Main Results:

• The Logistic regression model, incorporating 16 variables, achieved an AUC of 0.806 in the validation set.
• This model demonstrated superior performance compared to traditional scoring systems (APS III, SAPS III, SOFA).
• Key predictors included pH, albumin, temperature, lactate, creatinine, calcium, hemoglobin, WBC, age, SAPS III, APS III, Na+, BMI, and APTT.

Conclusions:

• A Logistic regression model with 16 selected variables is the best predictor of 28-day mortality in septic shock.
• The model exhibits stable performance, high discriminative ability, and accuracy.
• This enhanced predictive tool can aid in clinical decision-making for septic shock management.