Biomarker-based tests may facilitate Tuberculosis (TB) diagnosis, accelerate treatment initiation, and thus improve outcomes. This review synthesizes the literature on biomarker-based detection for TB diagnosis using machine learning. The systematic review approach follows the PRISMA guideline. Articles were sought using relevant keywords from Web of Science, PubMed, and Scopus, resulting in 19 eligible studies after a meticulous screening. All the studies were found to have focused on the supervised learning approach, with Support Vector Machine (SVM) and Random Forest emerging as the top two algorithms, with the highest accuracy, sensitivity and specificity reported to be 97.0%, 99.2%, and 98.0%, respectively. Further, protein-based biomarkers were widely explored, followed by gene-based such as RNA sequence and, Spoligotypes. Publicly available datasets were observed to be popularly used by the studies reviewed whilst studies targeting specific cohorts such as HIV patients or children gathering their own data from healthcare facilities, leading to smaller datasets. Of these, most studies used the leave one out cross validation technique to mitigate overfitting. The review shows that machine learning is increasingly assessed in research to improve TB diagnosis through biomarkers, as promising results were shown in terms of model's detection performance. This provides insights on the possible application of machine learning approaches to diagnose TB using biomarkers as opposed to the traditional methods that can be time consuming. Low-middle income settings, where access to basic biomarkers could be provided as compared to sputum-based tests that are not always available, could be a major application of such models.