Establishment of a Machine Learning-Based Predictive Model for Klebsiella pneumoniae Liver Abscess.

Purpose: To investigate the clinical and ultrasonographic characteristics of pyogenic liver abscess (PLA) caused by Klebsiella pneumoniae (K-PLA) and non-Klebsiella pneumoniae pathogens, and to develop machine learning models for the differential diagnosis of K-PLA.

Materials and methods: In this retrospective study, patients clinically diagnosed with PLA and confirmed by ultrasound-guided puncture at the Fifth Medical Center of PLA General Hospital between April 2013 and December 2020 were enrolled. Based on the causative pathogens, patients were categorized into K-PLA and non-K-PLA groups. Baseline data, including ultrasonographic features, clinical characteristics, and laboratory findings, were collected. The Boruta algorithm was employed for feature selection, and four machine learning models-Deep Learning-Fully Connected Neural Network (deeplearning), Distributed Random Forest (drf), Gradient Boosting Machine (gbm), and Generalized Linear Model (glm)-were developed to diagnose K-PLA. The models were validated using 5-fold cross-validation.

Results: A total of 201 patients with bacterial liver abscess were included (median age: 57 years; range: 49-66; 136 males), comprising 134 K-PLA cases and 67 non-K-PLA cases. The Boruta algorithm identified seven significant predictive variables: history of diabetes, history of hepatocellular carcinoma, history of biliary tract disease, history of infectious diseases, duration of fever, body temperature, and alanine aminotransferase (ALT) levels. Using these variables, the four machine learning models were constructed. In the training set, the area under the receiver operating characteristic curve (AUC) for predicting K-PLA was 0.716 (deeplearning), 0.999 (drf), 0.922 (gbm), and 0.718 (glm). In the validation set, the corresponding AUC values were 0.799, 0.763, 0.848, and 0.805, respectively.

Conclusion: This study successfully established four machine learning models for predicting the risk of K-PLA, with the gbm-based model demonstrating the highest diagnostic performance. These models may facilitate early clinical diagnosis and treatment of K-PLA, thereby reducing antibiotic misuse.