Gram-negative bloodstream infection is a life-threatening condition in intensive care units, frequently associated with severe sepsis, high mortality and rising antimicrobial resistance. Early identification of pathogens and timely antibiotic therapy are critical to improving outcomes, yet conventional blood culture methods remain time-consuming and may delay intervention. Although rapid molecular diagnostics can accelerate detection, their high cost and technical demands limit widespread use in many settings. A retrospective analysis conducted in the ICU of the West District of the First Affiliated Hospital of Anhui Medical University evaluated whether machine learning techniques could support earlier prediction of Gram-negative bloodstream infection using routinely available clinical and laboratory data. The resulting model prioritised parsimony, interpretability and feasibility within time-sensitive ICU workflows.
Cohort Design and Feature Selection
The study analysed data from ICU admissions between January and July 2025. Eligible patients were aged 18 years or older, had an anticipated ICU stay exceeding 48 hours and had at least one blood culture performed. Exclusion criteria removed patients with non-Gram-negative bloodstream infection, concurrent non-Gram-negative infection around the index episode and those who were pregnant or lactating. From 596 ICU admissions, 405 patients were included. Ninety-four patients, representing 23.21%, were identified with Gram-negative bloodstream infection.
Clinical and laboratory variables were collected within a 24-hour window surrounding blood culture sampling. Data preprocessing excluded patients with more than 30% missing values and variables with over 15% missing data. Remaining missing values were addressed using multivariate imputation by chained equations under a missing at random assumption. The cohort was randomly divided into training and validation sets in a 7:3 ratio.
Feature selection applied LASSO regression with 10-fold cross-validation to determine the optimal penalty parameter. Seven variables were initially identified and subsequently entered into multivariable logistic regression. Four variables were retained as optimal predictive features: deep vein catheterisation, continuous renal replacement therapy, procalcitonin and C-reactive protein. These variables formed the basis for model development.
Must Read: Interpretable ML Improves Post-Transplant Risk Prediction
Model Development and Validation
Seven machine learning algorithms were trained using the selected variables: logistic regression, decision tree, random forest, eXtreme Gradient Boosting, Light Gradient Boosting Machine, support vector machine and artificial neural network. Hyperparameter tuning was conducted using grid search with five-fold cross-validation in the training set. Model performance was assessed using area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, F1-score, positive predictive value and negative predictive value.
In the validation cohort, all models achieved accuracy and AUC values of 0.85 or above. The eXtreme Gradient Boosting model demonstrated favourable overall performance. It achieved an AUC of 0.898 with a 95% confidence interval of 0.807 to 0.959, an accuracy of 88.43%, a sensitivity of 60.71% and a specificity of 96.77%. The F1-score was 70.83%, with a positive predictive value of 85.00% and a negative predictive value of 89.10%. Calibration assessment produced a Brier score of 0.101, indicating good agreement between predicted and observed outcomes. Decision curve analysis showed a positive net benefit across relevant threshold probabilities, supporting potential clinical utility.
The emphasis on a limited number of routinely available ICU variables distinguished the framework from models relying on extensive predictor panels. The approach sought to balance discrimination performance with clinical feasibility in real-world ICU practice.
SHAP-Based Interpretability and Clinical Implications
To enhance transparency, Shapley Additive Explanations were applied to the eXtreme Gradient Boosting model. Feature importance ranking demonstrated that procalcitonin had the highest mean absolute SHAP value, followed by C-reactive protein, deep vein catheterisation and continuous renal replacement therapy. Beeswarm plots illustrated the direction and magnitude of each feature’s contribution to model output. Positive SHAP values were associated with increased predicted risk, while negative values indicated decreased risk.
Dependency plots showed that higher procalcitonin and C-reactive protein values were associated with increased predicted risk of Gram-negative bloodstream infection. Deep vein catheterisation and continuous renal replacement therapy were also associated with elevated risk when present. Force plots provided patient-level explanations by demonstrating how individual feature values contributed to predictions in representative cases with and without infection.
The model was positioned as a clinical decision-support tool rather than a diagnostic replacement. Its function is to support risk awareness and prioritisation, complementing microbiological confirmation and clinical judgement. Calibration curves and decision curve analysis in both training and validation cohorts demonstrated stable discrimination and meaningful clinical net benefit. The pathogen-oriented modelling strategy focused specifically on Gram-negative bloodstream infection rather than treating bloodstream infection as a homogeneous entity. Despite reliance on only four variables, discrimination performance was broadly comparable to that reported in other bacteremia prediction research, while maintaining simplicity and interpretability.
A machine learning framework based on four routinely available ICU variables achieved strong discrimination for early prediction of Gram-negative bloodstream infection. Among the evaluated algorithms, eXtreme Gradient Boosting demonstrated the most favourable performance, supported by calibration and decision curve analyses. SHAP-based interpretability clarified both feature-level and patient-level contributions to risk estimation. Although developed and internally validated in a single centre with a limited sample size, the model highlights the feasibility of parsimonious, interpretable prediction tools to support timely intervention in critically ill patients.
Source: BMC Medical Informatics and Decision Making
Image Credit: iStock
References:
Zhou YL, Da HT, Wang TT et al. (2026) A machine learning model for the early prediction of Gram-negative bloodstream infection in ICU patients. BMC Med Inform Decis Mak: In Press.
