OBJECTIVE  To analyze the influencing factors of medication adherence in patients with multidrug-resistant tuberculosis based on machine learning methods, and to construct and evaluate a prediction model.

METHODS  A total of 259 patients with multidrug-resistant tuberculosis admitted to Jiangxi Chest Hospital from Jan. 2020 to Dec. 2022 were selected as the modeling set, and another 112 patients with multidrug-resistant tuberculosis from Jan. to Dec. 2023 were selected as the validation set. The modeling set was divided into a good adherence group (MMAS-8≥6) (167 cases) and a poor adherence group (MMAS-8＜6) (92 cases) according to the Morisky Medication Adherence Scale (MMAS-8) score. Prediction models were constructed with logistic regression, decision classification regression tree, back propagation neural network (BPNN) algorithm and support vector machine algorithm, and their predictive values were compared through receiver operating characteristic curves.

RESULTS  The logistic regression model showed that age (OR=2.250), disease duration (OR=2.473), educational level (OR=0.274), family supervision of medication (OR=0.330), health education status (OR=0.397), adverse drug reactions (OR=2.579), comorbidities (OR=2.236), regular and timely follow-up visits (OR=0.235) and monthly family income (OR=0.367) were influencing factors for medication adherence in patients with multidrug-resistant tuberculosis (P＜0.05). Among the four models constructed by machine learning algorithms, the BPNN model demonstrated the best comprehensive predictive performance, with an area under the curve (AUC) of 0.855(95%CI: 0.805－0.905), sensitivity of 0.772 and specificity of 0.850. The validation set showed that all models had good predictive performance (AUC＞0.7).

CONCLUSIONS  Among the prediction models for medication adherence in patients with multidrug-resistant tuberculosis constructed based on machine learning algorithms, the BPNN model exhibits the best predictive performance. Regular and timely follow-up visits, educational level, disease duration, family supervision of medication and health education status are the main predictive factors.