Deep second- and third-degree burns become high-risk triggers of bloodstream infections due to severe disruption of the skin barrier and extensive tissue necrosis [1]. Due to high stress and consumption, along with trauma exposure and frequent invasive procedures, these patients have a 20–30 % incidence of developing bloodstream infections [2,3]. The event not only significantly prolongs the hospital stay and heightens the medical burden, but also raises the risk of sepsis and multi-organ failure by 3 to 5 times [4,5]. With the advancement of burn care technology, early identification of infection risk and implementation of precise interventions have become the core breakthroughs to improve patient prognosis. How to establish an efficient risk prediction tool based on easily accessible clinical indicators has always been a research hotspot in the field of burns.

Currently, there has been some progress in studies on risk prediction of bloodstream infections in burn patients. Traditional studies have focused on univariate analysis or multivariate regression models to identify classic risk factors such as total body surface area (TBSA), combined inhalation injury, white blood cell count (WBC), and C-reactive protein (CRP) [[6], [7], [8]]. However, these models are limited by the dimensionality of the features and fail to capture nonlinear relationships, making it difficult to integrate dynamically changing laboratory indicators and clinical features. Machine learning techniques have been gradually applied to the field of infection prediction by virtue of their advantages in processing high-dimensional data, such as logistic regression (LR), support vector machine (SVM), and back propagation artificial neural network (BP-ANN) that have demonstrated their potentials in the fields of pneumonia and urinary tract infections [9,10], but there remains a deficiency in studies concerning patients with deep second- and third-degree burns: first, the inclusion indicators are mostly limited to basic burn characteristics, lacking systematic integration of dynamic indicators such as inflammatory factors, coagulation function, electrolytes, etc.; second, the model validation link is weak, with most studies relying only on the logic model and failing to assess the generalization ability through cross-validation [11]; and third, the suitability of different machine learning algorithms for burn infection prediction has not been clarified to guide practical clinical applications. Therefore, the construction of optimal models based on multidimensional clinical indicators and rigorous validation has become the key to fill the gaps in this field.

In this study, a retrospective cohort design was used to systematically collect multidimensional indicators of burn characteristics, laboratory tests, and therapeutic measures within 72 h of admission, using patients with deep second- and third-degree burns as the study population. Based on the feature screening, four machine learning models, namely, LR, SVM, naive Bayes (NB), and BP-ANN, were constructed, and the differentiation, calibration and clinical utility of the models were comprehensively evaluated through stratified sampling and 5-fold cross-validation, and the optimal predictive model was finally screened out. It is expected to provide early intervention targets for physicians, optimize the allocation of anti-infective resources, and promote precise management.