Significant adverse prognostic events in patients with urosepsis: a machine learning based model development and...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Urosepsis, a sepsis subtype, has a rising incidence and high mortality rate.
• Accurate prognosis prediction is crucial for effective management of urosepsis patients.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for predicting prognosis in urosepsis patients.
• Identify key prognostic factors influencing urosepsis outcomes.

Main Methods:

• Utilized the MIMIC-IV database (v3.1) with 1389 urosepsis patients, split into training/validation cohorts (7:3).
• Employed ML algorithms (RF, Lasso, Boruta, XGBoost) for variable selection, identifying 9 optimal predictors.
• Evaluated model performance using accuracy, AUC, sensitivity, specificity, MCC, and F1-score, with SHAP for interpretability.

Main Results:

• The XGBoost model demonstrated superior performance with an accuracy of 0.818 and AUC of 0.904.
• Internal validation showed strong results: accuracy 0.797, AUC 0.869, sensitivity 0.797, specificity 0.752, MCC 0.597, F1-score 0.791.
• SHAP analysis provided global explanations for the XGBoost model, highlighting critical prognostic factors.

Conclusions:

• ML models, particularly XGBoost, offer significant prognostic capability for urosepsis.
• Interpretable ML (SHAP) facilitates clinical understanding and personalized treatment planning.
• External validation in diverse populations is recommended for generalizability.