Machine Learning Models Using Hospital Admission Characteristics Do Not Optimally Predict Nosocomial Infection...

Area of Science:

• Hepatology
• Infectious Diseases
• Data Science in Medicine

Background:

• Nosocomial infections (NI) significantly increase mortality in cirrhosis patients.
• Traditional logistic regression models have been insufficient for identifying high-risk individuals.
• Developing predictive models for NI is crucial for timely intervention.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting NI in hospitalized cirrhosis patients.
• To compare the performance of Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Neural Networks (NN) against logistic regression.
• To assess the clinical utility of ML models based on day-of-admission data.

Main Methods:

• Utilized data from the prospective CLEARED consortium, including over 120 centers globally.
• Applied RF, XGBoost, and NN models to predict NI using day-of-admission clinical data.
• Compared model performance using Area-under-Receiver-operating-characteristic curve (AUC) and Brier score.

Main Results:

• Included 8,263 cirrhosis patients; 10.5% developed NI, primarily respiratory and urinary tract infections.
• NI were associated with higher inpatient mortality and liver transplantation rates.
• The RF model achieved an AUC of 0.69, outperforming other ML models and logistic regression, but no model reached AUC ≥0.80 for clinical utility.

Conclusions:

• Machine learning models using day-of-admission data cannot reliably predict nosocomial infections in cirrhosis patients.
• Current predictive models lack sufficient accuracy for clinical decision-making.
• Protocolized infection-control measures are essential for all hospitalized cirrhosis patients.