Objective
Patients with deep second- and third-degree burns are at high risk of bloodstream infections (BSIs) due to skin barrier disruption and immune suppression, with poor prognosis. Early risk identification is crucial for improving outcomes. This study aimed to construct and validate a machine learning model using multidimensional clinical indicators to accurately predict BSI risk in such patients.
Methods
A retrospective cohort study enrolled 301 patients with deep second- and third-degree burns (75 with BSIs) from Yongchuan Hospital Affiliated to Chongqing Medical University between January 2020 and January 2025. Multidimensional data on burn characteristics, laboratory indicators, and therapeutic measures were collected within 72 h of admission. After data preprocessing and feature screening, four models were built: logistic regression (LR), support vector machine (SVM), naive Bayes (NB), and back propagation artificial neural network (BP-ANN). Model performance was evaluated via stratified sampling and 5-fold cross-validation.
Results
Eight key predictors were identified: total body surface area, lymphocytes (LYM, most important), platelet crit, total bilirubin, creatinine, C-reactive protein, procalcitonin, and 24-hour rehydration. The BP-ANN model performed best in the test set, with accuracy, recall, precision, F1 value, and AUC all reaching 0.857, good calibration (Hosmer-Lemeshow test, P = 0.142), and significant net benefit in the 0–0.3 risk threshold interval (decision curve analysis). The LR model had an AUC of 0.891 and high generalization stability (0.999) but less balanced indicators. SVM was overfitted (limited practical value), and NB had insufficient generalization (test set AUC=0.775).
Conclusion
The BP-ANN model based on multidimensional clinical indicators accurately predicts BSI risk in patients with deep second- and third-degree burns, with good differentiation, calibration, and clinical utility, providing a reliable tool for early intervention.
扫码关注我们
求助内容:
应助结果提醒方式:
