Harness machine learning for multiple prognoses prediction in sepsis patients: evidence from the MIMIC-IV database

Area of Science:

• Computational biology
• Medical informatics
• Clinical data science

Background:

• Sepsis, a severe systemic response to infection, has high mortality rates.
• Accurate prognostic tools are crucial for timely intervention in sepsis patients.
• Machine learning (ML) models show promise for early sepsis prediction.

Purpose of the Study:

• To develop and validate an ML model for early prognostic identification of sepsis patients in intensive care units (ICUs).
• To compare the performance of logistic regression, random forest, and CatBoost algorithms in predicting sepsis outcomes.

Main Methods:

• Utilized the MIMIC-IV v2.2 database, splitting data into 70% training and 30% validation sets.
• Applied difference analysis and multinomial logistic regression for feature selection, identifying 26 key clinical features.
• Constructed and evaluated logistic regression, random forest, and CatBoost models using precision, accuracy, recall, F1 score, and AUC.

Main Results:

• The CatBoost model demonstrated superior performance with a weighted AUC of 0.771.
• Mortality prediction achieved the highest AUC (0.804), followed by rapid recovery (0.773) and chronic critical illness (0.737).
• Urine output, respiratory rate, and temperature were identified as the most significant predictors.

Conclusions:

• The CatBoost ML model effectively identifies early sepsis prognosis, including mortality, rapid recovery, and chronic critical illness.
• Early interventions targeting urine output, respiratory status, and temperature may positively impact sepsis patient outcomes.
• Further external validation is necessary to confirm the model's generalizability.