A prediction and interpretation machine learning framework of mortality risk among severe infection patients with pseudomonas aeruginosa

Abstract

Pseudomonas aeruginosa is a ubiquitous opportunistic bacterial pathogen, which is a leading cause of nosocomial pneumonia. Early identification of the risk factors is urgently needed for severe infection patients with P. aeruginosa. However, no detailed relevant investigation based on machine learning has been reported, and little research has focused on exploring relationships between key risk clinical variables and clinical outcome of patients. In this study, we collected 571 severe infections with P. aeruginosa patients admitted to the Xijing Hospital of the Fourth Military Medical University from January 2010 to July 2021. Basic clinical information, clinical signs and symptoms, laboratory indicators, bacterial culture, and drug related were recorded. Machine learning algorithm of XGBoost was applied to build a model for predicting mortality risk of P. aeruginosa infection in severe patients. The performance of XGBoost model (AUROC = 0.94 ± 0.01, AUPRC = 0.94 ± 0.03) was greater than the performance of support vector machine (AUROC = 0.90 ± 0.03, AUPRC = 0.91 ± 0.02) and random forest (AUROC = 0.93 ± 0.03, AUPRC = 0.89 ± 0.04). This study also aimed to interpret the model and to explore the impact of clinical variables. The interpretation analysis highlighted the effects of age, high-alert drugs, and the number of drug varieties. Further stratification clarified the necessity of different treatment for severe infection for different populations.