Development and validation of machine learning models for predicting pancreatic injury observed after tigecycline initiation: a retrospective real-world study

Background Tigecycline is widely used for severe infections and infections caused by multidrug-resistant pathogens. Although pancreatic injury observed after tigecycline initiation is relatively uncommon, it may adversely affect patient prognosis. Currently, clinical management mainly relies on post-treatment monitoring, and reliable tools for early risk prediction are lacking. This study aimed to develop and internally validate an interpretable machine learning model for predicting pancreatic injury observed after tigecycline initiation using real-world clinical data. Methods This single-center retrospective study included patients who received tigecycline for at least 3 consecutive days at Nanfang Hospital, Southern Medical University, between January 2020 and March 2024. The primary outcome was pancreatic injury observed after tigecycline initiation, defined as serum amylase >405U/L and/or serum lipase >180 U/L within 7 days after tigecycline initiation. Demographic characteristics, comorbidities, concomitant medications, tigecycline treatment information, and baseline laboratory parameters were collected. Missing values were imputed using K-nearest neighbors imputation, and class imbalance was addressed using SMOTE. The dataset was randomly divided into training and test sets at an 8:2 ratio. Eight machine learning algorithms were developed and compared. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, positive predictive value, negative predictive value and F1 score. SHAP analysis was used for model interpretation. Results A total of 623 patients were included, of whom 146 (23.4%) developed pancreatic injury. Among the evaluated algorithms, LightGBM had the highest point-estimate AUC. The final 7-feature LightGBM model achieved an AUC of 0.726, accuracy of 0.728, sensitivity of 0.379, specificity of 0.833, PPV of 0.407, NPV of 0.816, and F1 score of 0.393 on the held-out test set. These results indicate moderate discrimination but limited sensitivity. Conclusion With relatively high specificity and negative predictive value, this model may provide preliminary risk stratification information and assist targeted pancreatic enzyme monitoring during tigecycline therapy, particularly for identifying patients at lower predicted risk. However, prospective external validation is required before routine clinical implementation.