Using NLP, we identified topics with high Cᵥ coherence scores derived from embeddings and semantic search clustering, indicating that NLP-extracted topics capture meaningful semantic structures. These topics highlighted common treatment experiences and adherence barriers faced by patients. Incorporating human-in-the-loop supervision further refined keywords within topic clusters. NLP-assisted analysis of data from routine medical records presents an opportunity to support targeted intervention development.

Varying patterns of medication adherence were significantly associated with clinical DR-TB outcome. The indication that early missed doses were most strongly associated with negative treatment outcomes highlights the importance of treatment adherence support strategies, particularly at the beginning of treatment or when the patients begin in missing consecutive doses.