Recent Advancements in Machine Learning-Based Bloodstream Infection Prediction: A Systematic Review and Meta-analysis of Diagnostic Test Accuracy

Abstract

Abstract. Purpose: Bloodstream infections (BSIs) present significant public health challenges. With the advent of machine learning (ML), promising predictive models have been developed. This study evaluates their performance through a systematic review and meta-analysis. Methods: We performed a comprehensive systematic review across multiple databases, including PubMed, IEEE Xplore, ScienceDirect, ACM Digital Library, SpringerLink, Web of Science, Scopus, and Google Scholar. Eligible studies focused on BSIs within any hospital setting, employing ML models as the diagnostic test. We evaluated the risk of bias with the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) checklist and assessed the quality of evidence using the Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach. Models reporting the area under the receiver operating characteristic curve (AUROC) were included in the meta-analysis to identify key performance drivers. Results: After screening, a total of 30 studies were eligible for synthesis, from which 41 models and 8 data types were extracted. Most of the studies were carried out in the inpatient settings (n=17; 56%), followed by the emergency department (ED) settings (n=7; 23%), and followed by the ICU settings (n=6; 20%). The reported AUROCs in the hospital inpatients settings, ranged from 0.51-0.866, in the ICU settings AUROCs ranged from 0.668-0.970, and in the emergency department (ED) settings the AUROCs of the models ranged from 0.728-0.844. One study reported prospective cohort study, while two prospectively validated their models. In the meta-analysis, laboratory tests, Complete Blood Count/Differential Count (CBC/DC), and ML model type contributed the most to model performance. Conclusion: This systematic review and meta-analysis show that on retrospective data, individual ML models can accurately predict BSIs at different stages of patient trajectory. Although they enable early prediction of BSI, a comprehensive approach to integrate data types and models is necessary. Systematic reporting, externally validated, and clinical implementation studies are needed to establish clinical confidence.