Xiaoyu Zhang ¹ Xifan Zhang ¹ Deng Zhang ² Jing Xu ¹ Jingping Zhang ^1* Xin Zhang ^1*

• ¹ Department of Infectious Diseases, The First Affiliated Hospital of China Medical University, Shenyang, China
• ² Department of Infectious Diseases, The First Affiliated Hospital of Xiamen University,Xiamen,China, Xiamen, Fujian Province, China

Objective: To develop a machine learning-based prediction model to assist clinicians in accurately determining whether the detection of Klebsiella pneumoniae (KP) in sputum samples indicates an infection, facilitating timely diagnosis and treatment. Research Methods: A retrospective analysis was conducted 286 cases with sputum cultures yielding only KP were included, comprising 67 cases in the colonization group and 219 cases in the infection group. Antimicrobial susceptibility testing was performed on the included strains, and through univariate logistic regression analysis, 15 key influencing factors were identified. These factors were used to construct the model, which was evaluated using accuracy, precision, recall, F1 score, AUC value, and Brier score. Results: Antimicrobial susceptibility testing indicated that the resistance rates for penicillins, cephalosporins, carbapenems, and quinolones were significantly higher in the infection group compared to the colonization group (all P < 0.05). Six predictive models were constructed using 15 key influencing factors, including Classification and Regression Trees (CART), C5.0, Gradient Boosting Machines (GBM), Support Vector Machines (SVM), Random Forest (RF), and Nomogram. The Random Forest model performed best among all indicators (accuracy 0.93, precision 0.98, Brier Score 0.06, recall 0.72, F1 Score 0.83, AUC 0.99).The importance of each factor was demonstrated using mean decrease in Gini. "Admitted with a diagnosis of respiratory infectious disease" (8.39) was identified as the most important factor in the model, followed by "Hypoalbuminemia" (7.83), then "ESBL" (7.06), "Electrolyte Imbalance" (5.81), "Age > 62 years" (5.24), "The number of Positive Sputum Cultures for KP > 2" (4.77), and being bedridden (4.24). Additionally, invasive procedures (such as history of tracheostomy, use of ventilators for > 96 hours, and craniotomy) were also significant predictive factors. The Nomogram indicated that CRKP, presence of a nasogastric tube, admission to the ICU, and history of tracheostomy were important factors in determining KP colonization. Conclusion: The Random Forest model effectively distinguishes between infection and colonization status of KP, while the Nomogram visually presents the predictive value of various factors, providing clinicians with a reference for formulating treatment plans.