Early diagnosis and prognostic prediction of secondary bloodstream infections caused by Acinetobacter baumannii in critically ill patients by machine-learning algorithms.

BACKGROUND

Secondary bloodstream infections (sBSI) caused by Acinetobacter baumannii (AB) are a major threat to patient safety in the Intensive Care Unit (ICU) due to their prevalence and severity. Developing accurate predictive models is crucial for enhancing clinical decision-making and improving patient outcomes. This study aimed to leverage machine learning (ML) to create a diagnostic model for predicting the risk of AB-sBSI in ICU patients and a prognostic model for assessing the associated 30-day mortality risk.

METHODS

The multicenter, retrospective study enrolled 4,267 ICU patients with AB isolated from non-blood sites. Of these, 337 patients developed bloodstream infection. The analysis included 70 patients with confirmed AB secondary bloodstream infection (AB-sBSI) and 76 age and sex matched controls with non AB-sBSI. For 30-day mortality assessment, the AB-sBSI patients were categorized into non-survivors (n=39) and survivors (n=31). Demographic, microbiological, and laboratory data encompassing hematological, coagulation, and inflammatory markers were analyzed. Fourteen machine learning models were evaluated using the Deepwise and Beckman Coulter DxAI platforms with five-fold cross-validation. Model performance was assessed using five standard metrics, and the DeLong test was applied for AUC comparison. After data preprocessing, patients were enrolled to form an external validation cohort.

RESULTS

The AB-sBSI risk diagnosis model, constructed with 11 features, identified red cell distribution width as the most significant predictor. The AdaBoost model outperformed both comparative models (Linear Discriminant Analysis, Logistic Regression, LinearSVC) and the conventional biomarker C-reactive protein (AUC = 0.66), with AUCs of 0.937 in training and 0.786 in validation. For 30-day mortality prediction, another model based on 11 features selected lymphocyte count as the most influential variable. The AdaBoost model showed prominent efficacy, surpassing other model (Multilayer Perceptron, BernoulliNB, SGD) and achieving AUC values of 0.986 in training and 0.821 in validation.

CONCLUSION

We developed two ML based models for predicting AB-sBSI risk and 30-day mortality. As a preliminary exploration, both models have been converted into accessible web tools. These tools are designed to assist clinicians in making informed decisions and promptly adjusting treatment strategies for critically ill patients.