This work presents a new framework for tuberculosis diagnosis and treatment optimization using the Moth Search Algorithm (MSA) and Stacked Decision Tree Structures. The MSA optimizes feature selection, which decreases dimensionality by 23%, resulting in quicker training times. Stacked decision trees enhance classification performance with accuracy of 96.4%, precision of 94.8%, recall of 95.2%, and F1-score of 95.0%—7–10% better than traditional machine learning models. The hybrid approach identifies 28% better convergence and 15% fewer false negatives, of utmost relevance to early diagnosis. Personalized treatment planning is facilitated in this approach for better patient outcome. Future applications are real-time execution and extending the model’s flexibility to other infectious diseases in limited-resource clinic settings.