Use of artificial intelligence for predicting outcomes of severe bacterial infections in intensive care units

This study aims to evaluate the performance of artificial intelligence (AI) models in predicting outcomes of severe bacterial infections among intensive care unit (ICU) patients. The primary focus is to compare AI-based predictions with traditional scoring systems, including APACHE II and SOFA, in terms of mortality, length of stay (LOS), and therapeutic response. A retrospective cohort analysis was conducted using data from ICU patients admitted between January 2020 and December 2024. Data preprocessing involved handling missing values using the MICE method, encoding categorical variables, and normalizing continuous variables. Four machine learning models were developed: XGBoost, Random Forest, Deep Neural Networks (DNN), and Logistic Regression. Model performance was evaluated using AUROC for mortality prediction, MAE for LOS prediction, and F1-score for therapeutic response prediction. XGBoost demonstrated the highest accuracy in predicting mortality, with an AUROC of 0.93, followed by Random Forest (0.91) and DNN (0.89). Logistic Regression showed the lowest performance (AUROC 0.82). Age, SOFA score, and lactate level were identified as the most significant predictors. In predicting LOS, XGBoost and Random Forest again outperformed other models, with MAE values of 3.2 and 3.4 days, respectively. For therapeutic response, XGBoost achieved the highest F1-score of 0.81. AI models consistently outperformed traditional scoring systems. AI models, particularly XGBoost, significantly improve the accuracy of predicting mortality, LOS, and therapeutic response compared to conventional methods. Integrating these models into ICU practice could enhance clinical decision-making and patient management, promoting more personalized and timely interventions. Further validation in diverse clinical settings is recommended.