Tuberculosis (TB) is an infectious disease that has been declared a global emergency by the World Health Organization and remains one of the top ten causes of death worldwide. TB diagnosis is particularly challenging in developing countries, where limited infrastructure for detection and treatment complicates efforts to control the disease. These resource constraints are especially critical in remote areas with few mechanisms for timely diagnosis, which is essential for effective patient management. Artificial intelligence (AI) has emerged as a valuable tool in supporting health professionals by enhancing diagnostic processes. This paper explores the use of natural language processing (NLP) techniques and machine learning (ML) models to facilitate TB diagnosis in settings where robust data infrastructure is unavailable. Two distinct data sources were analyzed: text extracted from electronic medical records (EMRs) and patient clinical data (CD). Four different ML-based approaches were implemented: two models using each data source independently and two data fusion models combining both sources. The relevance of these strategies was assessed in collaboration with physicians to ensure their practical applicability in clinical decision-making. The results of the data fusion models were compared to determine which source provided more valuable diagnostic information. The best-performing model, which relied solely on CD, achieved a sensitivity of (Formula presented.), outperforming smear microscopy, which typically ranges from (Formula presented.) to (Formula presented.). These findings underscore the importance of analyzing physicians’ reports and assessing the availability of such information alongside structured clinical data. This approach is particularly beneficial in resource-limited settings, where access to comprehensive clinical data may be restricted.