# Personalized machine learning–based prognostic model for ICU-acquired bloodstream infections

## TL;DR

A machine learning model was developed to predict outcomes for ICU patients with bloodstream infections, using clinical data to identify high-risk patients early.

## Contribution

A personalized ML-based prognostic model for ICU-acquired bloodstream infections was developed and validated across datasets.

## Key findings

- The XGBoost model achieved strong discrimination with AUROC of 0.92 in training and 0.85 in internal validation.
- Key variables like antibiotic duration and platelet count were strongly associated with 28-day mortality.
- The model performed well in predicting outcomes for Gram-positive and Gram-negative infections.

## Abstract

Intensive care unit–acquired bloodstream infections (ICU-BSIs) are among the most prevalent healthcare-associated infections and a major cause of mortality among ICU patients. We developed a machine learning (ML)–based model to predict the prognosis of ICU-BSIs.

Adult patients with blood cultures drawn ≥48 hours after ICU admission were included: the Second Affiliated Hospital of Anhui Medical University (AMU, China) and the Medical Information Mart for Intensive Care IV (MIMIC-IV, USA). The AMU dataset was used for model training and internal validation, and the MIMIC-IV dataset served as the external validation set. The model incorporated routinely collected, easily obtainable clinical variables, including several representing the average rate of change in laboratory indicators. After comparing multiple algorithms, eXtreme Gradient Boosting (XGBoost) was selected and optimized using cross-validation and grid search.

A total of 1,903 patients from AMU and 3,496 from MIMIC-IV were included. In both cohorts, antibiotic duration, platelet count, serum creatinine, duration of invasive mechanical ventilation, and Charlson Comorbidity Index (CCI) were significantly associated with 28-day mortality (P < 0.001). The XGBoost model using 33 variables showed strong discrimination, with an AUROC of 0.92 (95% CI 0.90–0.94) for training and 0.85 (95% CI 0.80–0.90) for internal validation. Shapley Additive Explanations (SHAP) identified the 10 most important variables; a simplified model using these maintained good accuracy, with AUROC values of 0.81 (95% CI 0.76–0.85) and 0.71 (95% CI 0.70–0.73) for the internal and external validation sets, respectively. In pathogen subgroups, the internal AUROC was 0.91 (95% CI 0.87–0.94) and 0.90 (95% CI 0.86–0.93) for Gram-positive (Gram+) and Gram-negative (Gram−) infections, with external validation AUROCs of 0.72 (95% CI 0.66–0.77) and 0.72 (95% CI 0.62–0.82), respectively.

We developed and externally validated a personalized ML-based prognostic model for ICU-BSIs using multicenter time-series data. This model may facilitate early identification of high-risk patients, enabling timely intervention and optimized ICU resource allocation.