Machine Learning Based Early Prediction Model for Bloodstream Infections: A Scoping Review

Bloodstream infections are the leading cause of high mortality rates among critically ill patients, making early identification and timely intervention crucial. The traditional gold standard for diagnosis, blood culture, suffers from temporal delays and insufficient sensitivity, while clinical scoring systems exhibit limited predictive efficacy. Although machine learning-based predictive models have emerged continuously, their research designs, feature engineering, and validation strategies demonstrate significant heterogeneity, lacking systematic methodological frameworks and bias risk assessments. This review systematically examines studies utilizing machine learning algorithms to predict bloodstream infections risk from 2016 to 2026, summarizing data sources, algorithm architectures, and model performance, with a focus on evaluating bias risks and clinical translational feasibility. Within the inclusion/exclusion criteria, intensive care units were the predominant research setting (52%). The included models predominantly employed extreme gradient boosting, random forests, and deep learning architectures, with receiver operating characteristic curve areas of interest ranging between 0.75 and 0.97. Dynamic vital signs and laboratory inflammatory markers served as core input features. However, 80% of the studies were single-center retrospective designs, with only a minority undergoing rigorous external validation. Common issues included non-standard data preprocessing and inadequate handling of class imbalance, posing substantial bias risks.