Abstract:

Objective: To develop machine learning-based diagnostic models for efficient differential diagnosis between early Mycobacterium intracellulare pulmonary disease and pulmonary tuberculosis. Methods: A total of 335 patients, including 135 cases of Mycobacterium intracellulare pulmonary disease and 200 cases of pulmonary tuberculosis admitted to Beijing Chest Hospital, Capital Medical University from January 1, 2021 to December 31, 2023, were retrospectively enrolled as the training cohort. Sociodemographic characteristics, baseline clinical parameters, and immunological indicators were collected. Univariate logistic regression was used for preliminary feature screening, followed by combined feature selection using weighted aggregated SHapley Additive exPlanations (SHAP)-based recursive feature elimination and Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify core discriminative indicators, based on which multiple machine learning differential diagnosis models were constructed. An independent external validation cohort comprising 83 patients (35 with Mycobacterium intracellulare pulmonary disease and 48 with pulmonary tuberculosis) from the Fifth People’s Hospital of Suzhou, admitted from January 1, 2024 to December 31, 2024, was used to evaluate model generalizability. Results: Univariate analysis initially identified 11 clinical and immunological indicators with statistically significant intergroup differences. After dual feature screening, five core differential indicators were finally identified, namely age, pulmonary infection, diabetes mellitus, bronchiectasis, and body mass index (BMI). Among five machine learning models constructed based on these core indicators (Random Forest, Light Gradient Boosting Machine, Extreme Gradient Boosting, Extremely Randomized Trees, and Adaptive Boosting), the Random Forest model achieved the best diagnostic performance, with an area under the receiver operating characteristic curve (AUC) of 0.946 in the training cohort. In the independent external validation cohort, the model achieved an AUC of 0.830, with a sensitivity of 71.4% (25/35) and specificity of 75.0% (36/48), demonstrating stable discrimination between the two diseases. SHAP analysis enabled individualized interpretation of model predictions, and the model exhibited satisfactory overall interpretability and clinical stability. Conclusion: The Random Forest model constructed using routine clinical indicators can effectively differentiate early Mycobacterium intracellulare pulmonary disease from pulmonary tuberculosis, providing a reference for rapid auxiliary differential diagnosis in clinical practice.

Key words: Mycobacteriaceae, Tuberculosis,pulmonary, Diagnosis, Artificial intelligence