Prediction Model for Delirium in Patients with Sepsis-Associated Liver Injury: An Interpretable Machine Learning...

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Neurological Complications

Background:

• Sepsis-associated liver injury (SALI) patients face high risks of delirium, leading to poor neurological outcomes.
• Current predictive tools lack specificity for this patient group, hindering early identification.
• Delirium in SALI is a significant clinical challenge requiring improved prediction strategies.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting delirium in patients with SALI.
• To identify key risk factors associated with delirium development in this population.
• To enable early risk stratification and targeted interventions for improved patient outcomes.

Main Methods:

• Retrospective cohort study using MIMIC-IV data.
• Development and testing of multiple machine learning models (SVM, LR, RF, GBM, XGBoost).
• Identification of independent predictors using logistic regression and model interpretation via SHAP analysis.

Main Results:

• 62.8% of 1461 SALI patients developed delirium.
• Key predictors included diabetes with complications, low SpO2, low hemoglobin, low Glasgow Coma Scale (GCS) score, and treatments like CRRT, vasopressin, or mechanical ventilation.
• The Gradient Boosting Machine (GBM) model achieved an AUROC of 0.811 in the testing set, with mechanical ventilation, GCS, CRRT, and hemoglobin being most influential.

Conclusions:

• An interpretable GBM model effectively predicts delirium in SALI patients.
• SHAP analysis highlights multi-organ dysfunction markers as primary drivers of delirium risk.
• This model facilitates individualized risk assessment and targeted preventive strategies for vulnerable patients.