Abstract
Background/Objectives: Treatment adherence challenges affect 10-20% of tuberculosis patients globally, contributing to drug resistance and continued transmission. While artificial intelligence approaches show promise for identifying patients who may benefit from additional treatment support, most models lack the interpretability necessary for clinical implementation. We aimed to develop and validate an integrated artificial intelligence framework combining traditional machine learning (interpretable algorithms like logistic regression and decision trees), explainable AI (methods showing which patient characteristics influence predictions), deep reinforcement learning (algorithms learning optimal intervention strategies), and natural language processing (clinical text analysis) to identify tuberculosis patients who would benefit from enhanced treatment support services. Methods: We analyzed 103,846 pulmonary tuberculosis cases from São Paulo state surveillance data (2006-2016). We evaluated models using precision (accuracy of positive predictions), recall (ability to identify all patients requiring support), F1-score (balanced performance measure), and AUC-ROC (overall discrimination ability) while maintaining interpretability scores above 0.90 for clinical transparency. Results: Our integrated framework demonstrated that explainable AI matched traditional machine learning performance (both F1-score: 0.77) while maintaining maximum interpretability (score: 0.95). The combined ensemble delivered superior results (F1-score: 0.82, 95% CI: 0.79-0.85), representing a 6.5% improvement over individual approaches (p < 0.001). Key predictors included substance use disorders, HIV co-infection, and treatment supervision factors rather than demographic characteristics. Conclusions: This multi-paradigm AI system provides a methodologically sound foundation for identifying tuberculosis patients who would benefit from enhanced treatment support services. The approach delivers excellent predictive accuracy while preserving full clinical transparency, demonstrating that the accuracy-interpretability trade-off in medical AI can be resolved through the systematic integration of complementary methodologies.