Tuberculosis (TB) is among the more frequent causes of death in many countries. For pulmonary TB, early diagnosis greatly increases the efficiency of therapies. Although highly sensitive tests based on nucleic acid amplification tests (NAATs) and loop-mediated isothermal amplification (TB-LAMP) are available, smear microscopy is still the most widespread diagnostics method in most low-middle-income countries, and the true positive rate of smear microscopy is lower than 65%. Thus, there is a need to increase the performance of low-cost diagnosis. For many years, the use of sensors to analyze the exhaled volatile organic compounds (VOCs) has been proposed as a promising alternative for the diagnosis of several diseases, including tuberculosis. In this paper, the diagnostic properties of an electronic nose (EN) based on sensor technology previously used to identify tuberculosis have been tested on-field in a Cameroon hospital. The EN analyzed the breath of a cohort of subjects including pulmonary TB patients (46), healthy controls (38), and TB suspects (16). Machine learning analysis of the sensor array data allows for the identification of the pulmonary TB group with respect to healthy controls with 88% accuracy, 90.8% sensitivity, 85.7% specificity, and 0.88 AUC. The model trained with TB and healthy controls maintains its performance when it is applied to symptomatic TB suspects with a negative TB-LAMP. These results encourage the investigation of electronic noses as an effective diagnostic method for future inclusion in clinical practice.
Keywords: breath analysis; electronic nose; tuberculosis (TB); volatile organic compounds (VOCs).