Analysis of Drug Resistance Characteristics and Risk Factors of Cavitary Tuberculosis based on Whole Genome Sequencing and Machine Learning

Background Pulmonary cavitation, a high-burden driver of pulmonary tuberculosis (PTB) transmission, necessitates targeted interventions. This study comprehensively analyzes the drug-resistance profiles and risk factors associated with cavitary TB to inform precise prevention and control strategies. Methods This study analyzed 1,247 cavitary PTB patients from Guangxi (2020-2024) with complete strain and clinical data. Whole-genome sequencing characterized strain lineages and drug resistance. Key predictors were selected using Lasso regression, and the optimal model was identified from nine machine learning (ML) models based on AUC, SHAP elucidated feature contributions to severe cavitary tuberculosis risk. Results Lineages were primarily classified as Lineage 2 (65.20%), Lineage 4 (29.91%), and 35 cases of mixed infections (2.81%). Concordance between WGS and phenotypic drug susceptibility testing (pDST) was moderate for INH (Kappa=0.634; χ2=32.667, P<0.001) but good for RFP (Kappa=0.774; difference not significant). Predominant mutations were rpoB_p.Ser450Leu (RFP), katG_p.Ser315Thr (INH), embB_p.Met306Val (EMB), and rpsL_p.Lys43Arg (S). Lasso regression selected ten variables: fatigue, fever, history of previous tuberculosis treatment, gender, age, RFP, INH, occupation, S, and ethnicity. After evaluating nine ML models, the Gradient Boosting Machine (GBM) was selected as optimal. SHAP analysis identified fatigue, older age, history of tuberculosis treatment, fever, male, and rpoB_p.Ser450 mutation were positively associated with severe cavity formation. Conclusions The rpoB_p.Ser450 mutation is linked to severe cavitary PTB, but clinical and population studies are still needed to confirm this association. Therefore, new tools based on clinical indicators and biomarkers are needed to achieve early warning and timely intervention for the risk of severe cavitation.