Machine Learning Models Using Hospital Admission Characteristics Do Not Optimally Predict Nosocomial Infection Development in a Global Cirrhosis Cohort.

Abstract

BACKGROUND Nosocomial infections (NI) in cirrhosis are associated with high mortality but could be preventable. Logistic regression (LR) models have failed to identify high-risk patients. We aimed to develop machine learning (ML) models to predict NI. METHODS The CLEARED consortium consists of prospectively enrolled cirrhosis inpatients from >120 centers. Using day-of-admission clinical data, three ML approaches[Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Neural Networks (NN)] were used to predict NI. Data were split 80:20 for training and testing stratified by the outcome. Models were compared using Area-under-Receiver-operating-characteristic curve (AUC). RESULTS 8,263 patients (55.90±13.34 years; 64.1% men) from 127 centers in 37 countries were included. NI developed in 869 (10.5%) a median of 6 (4-11) days post-admission. Major NIs were respiratory (29.6%) and urinary tract infection (15.7%), and spontaneous bacterial peritonitis (13.5%). NIs occurred more frequently in patients from low/low-middle income countries and those with severe liver disease, alcohol etiology, and admission infections. NIs were associated with inpatient mortality (31.9% vs. 8.1%, p<0.001) and liver transplantation (4.8% vs. 1.9%, p<0.001). While the RF model (AUC 0.69) showed good calibration (Brier score 0.09), outperforming XGBoost, NN, and LR models (AUC 0.66 for all; LR comparison p = 0.043), no model achieved AUC ≥0.80 for clinical utility. At 10% predicted probability threshold, RF model demonstrated only 75.4% sensitivity, 52.9% specificity, and 15.9% PPV. CONCLUSIONS NIs cannot be accurately predicted from day-of-admission data using ML models, even in a large, prospective, global cirrhosis cohort. Every hospitalized patient with cirrhosis should receive protocolized infection-control measures.