Integrated Artificial Intelligence Framework for Tuberculosis Treatment Abandonment Prediction: A Multi-Paradigm...

Area of Science:

• Artificial Intelligence in Medicine
• Public Health Informatics
• Machine Learning for Disease Management

Background:

• Treatment adherence issues impact 10-20% of global tuberculosis patients, fueling drug resistance and transmission.
• Existing artificial intelligence (AI) models often lack the interpretability required for clinical use.
• There is a need for AI systems that can identify tuberculosis patients requiring enhanced support while maintaining transparency.

Purpose of the Study:

• To develop and validate an integrated AI framework for identifying tuberculosis patients who would benefit from enhanced treatment support.
• To combine traditional machine learning, explainable AI, deep reinforcement learning, and natural language processing.
• To ensure clinical transparency and interpretability in AI-driven patient identification.

Main Methods:

• Analysis of 103,846 pulmonary tuberculosis cases from São Paulo state surveillance data (2006-2016).
• Development of an integrated AI framework using traditional ML, explainable AI, deep reinforcement learning, and NLP.
• Model evaluation using precision, recall, F1-score, and AUC-ROC, with a focus on maintaining interpretability scores above 0.90.

Main Results:

• Explainable AI achieved performance comparable to traditional ML (F1-score: 0.77) with maximum interpretability (score: 0.95).
• The integrated AI ensemble significantly outperformed individual approaches (F1-score: 0.82), showing a 6.5% improvement (p < 0.001).
• Key predictors for needing support included substance use disorders and HIV co-infection, not demographics.

Conclusions:

• The developed multi-paradigm AI system offers a robust method for identifying tuberculosis patients needing enhanced support.
• The AI approach achieves high predictive accuracy while maintaining full clinical transparency.
• This study demonstrates that the accuracy-interpretability trade-off in medical AI can be resolved through integrated methodologies.