Related Experiment Video
Updated: Jan 9, 2026
Microscopy-based Assays for High-throughput Screening of Host Factors Involved in Brucella Infection of Hela Cells
Published on: August 5, 2016
Machine learning-based prediction model for chronic brucellosis: a multi-feature approach using clinical and
Rong Wang1,2, Bin Niu1,2, Chenming Zhang1,3
1Department of Infectious Diseases, The First Hospital of Shanxi Medical University, Taiyuan, China.
Background:
Chronic progression is a major clinical challenge in human brucellosis (HB), affecting nearly one-third of patients and leading to long-term disability. Reliable early prediction tools are lacking, hindering timely risk stratification and individualized management. This study aimed to develop and validate machine learning (ML) models to predict chronic progression using routinely available clinical and laboratory data.
Methods:
We retrospectively analyzed 555 patients with confirmed brucellosis admitted between 2019 and 2024. Clinical characteristics and laboratory indicators at admission were collected. Feature selection was performed using Boruta and recursive feature elimination. Six supervised ML models (random forest [RF], LightGBM, XGBoost, logistic regression [LR], multilayer perceptron [MLP], and support vector machine [SVM]) were constructed and evaluated by discrimination, calibration, clinical utility, and predictive metrics. Model interpretability was assessed using SHapley Additive exPlanations (SHAP), and a web-based prediction tool was developed.
Results:
Of 555 patients, 144 (25.9%) progressed to chronic brucellosis. Compared with the recovery group, chronic cases presented more frequently with arthralgia and arthritis and showed distinct biochemical profiles, including lower alanine aminotransferase (ALT), aspartate aminotransferase (AST), triglycerides (TG), and higher high-density lipoprotein cholesterol (HDL-C), albumin (ALB), blood urea nitrogen (BUN), and uric acid (UA). Among the six models, RF consistently demonstrated the most robust performance across metrics, achieving the highest AUC in the test set (0.782, 95% CI: 0.701 - 0.856), superior calibration (Emax = 0.155), and the greatest net clinical benefit in decision curve analysis. SHAP analysis identified TG, HDL-C, UA, eosinophil count, PA, ALT, BUN, and GLB as the most influential predictors, with biologically plausible associations.
Conclusion:
Using eight routinely available variables, the RF model demonstrated moderate discrimination with well-calibrated probability estimates but limited sensitivity. The tool may assist early risk stratification of chronic brucellosis when combined with clinical judgment; however, its predictive performance should be interpreted cautiously until validated in external, multicenter, and prospective studies.
