Machine Learning Models for Predicting Tuberculosis Treatment Adherence, Nonadherence, and Loss to Follow-up in Low- and Middle-income Countries: A Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020-compliant Systematic Review and Illustrative Meta-analysis.

BACKGROUND: Suboptimal adherence to tuberculosis (TB) treatment continues to undermine global TB control efforts, particularly in low- and middle-income countries. In recent years, machine learning (ML) approaches have been explored as tools to predict adherence-related outcomes and guide targeted interventions. OBJECTIVE: The objective of this study was to systematically review and synthesize the predictive performance of ML models for TB treatment adherence, nonadherence, and loss to follow-up. METHODS: We conducted a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines. Multiple databases were searched from inception to March 2026. Studies that used ML models to predict TB adherence-related outcomes and reported the area under the receiver operating characteristic curve (AUC) were included. We performed a random-effects meta-analysis of logit-transformed AUC values, along with assessments of heterogeneity, publication bias, sensitivity analysis, and risk of bias using the Prediction model Risk of Bias Assessment Tool. RESULTS: Seven studies met inclusion criteria. The reported AUC values ranged from 0.72 to 0.95. The pooled AUC was 0.84 (95% confidence interval: 0.76-0.90), indicating good predictive performance. Substantial heterogeneity was observed (I2 = 99.9%, P < 0.001). Sensitivity analysis demonstrated stability of results. Egger's test showed no evidence of publication bias (P = 0.915). Most studies were judged to be at high risk of bias, particularly in the analysis domain. CONCLUSION: ML models demonstrate promising ability to predict TB treatment adherence-related outcomes. However, variability in study methods and a high risk of bias underscore the need for standardized approaches and rigorous external validation.