Pakistan has the fifth highest incidence of Tuberculosis (TB) in the world. The current tests for TB diagnosis are smear and culture tests which have accuracies of about 40% and 70% respectively. They are inefficient, complicated and relatively expensive to perform in developing countries. In this paper, we present a novel computational predictive algorithm by modifying the standard decision tree that allows us to efficiently detect TB with high accuracy. It employs supervised learning to classify the samples into patients and healthy groups based on MFI values of different antibodies with an accuracy of about 94%, outperforming the traditional classifiers including decision trees, kNN, random forests and support vector machines (SVM). Our algorithm allows simultaneous, bench-top analysis of thousands of samples per day, does not rely on highly skilled or technical staff and provides easily interpretable results. Our work strongly suggests that developing, testing and implementing automated diagnostic algorithms such as ours can be helpful in overcoming infrastructural and human-resource constraints in poorly resourced countries.
Hussainy, Syed Fahadulla; Zaffar, Fareed; Zaffar, Muhammad Adeel; Khaliq, Aasia; Khan, Imran H.; and Ahmad, Raza, "Decision-tree inspired classification algorithm to detect Tuberculosis (TB)" (2017). PACIS 2017 Proceedings. 182.