Tuberculosis (TB) is an infectious disease primarily impacting the lungs. It spreads through the air when an infected person coughs, sneezes, or talks. Diagnosing TB involves clinical examinations and specialized tests performed by medical professionals. Coughing is a common symptom. The diagnosis of TB involves clinical examinations and specialized tests. However, studies have shown that medical doctors can distinguish between cough sounds associated with different respiratory conditions. Therefore, using artificial intelligence to analyze cough recordings of patients to diagnose TB is a promising research direction. In this study, we propose a customized cascaded approach for diagnosing TB using cough audio. This approach involves a series of models arranged in a sequence, where the output of one model serves as the input for the next. In the first phase, we distinguish between bursts in audio signal as noise or cough. In the second phase, we classify cough as TB or non-TB. Non-TB cough includes both voluntary and non-TB reflex cough. For this study we collected a dataset consisting of cough audio recordings from TB and non-TB patients at Mayo Hospital in Lahore, Pakistan. The recordings were obtained using the AI4LYF DCT application, a fully automated phone-based system, with no manual annotation. We apply statistical classifiers based on spectral and time domain features, both with and without clinical metadata. Through a stratified grouped cross-validation approach, our results show that using cough sounds along with demographic and clinical factors yielded an accuracy of 97% when the random forest was used. Similarly, for all other classifiers, the accuracies were ≥90 % when demographic and clinical data was included (from ≤80 )^. Our findings suggest that our model based on patient data and cough auido could support community health workers and health programs in identifying TB cases more effectively and cost-efficiently.