Prediction of multidrug-resistant tuberculosis in tuberculosis patients using perceptron artificial neural networks model

Abstract

Background & Aims: Diagnosis and treatment of patients with multidrug-resistant tuberculosis (MDR-TB) are very important. Hence, it is necessary to predict and diagnose these patients based on individual, demographic and clinical characteristics before starting treatment. This study aimed to predict MDR-TB in TB patients using the perceptron artificial neural networks (ANNs) model. Materials & Methods: This retrospective cohort study was conducted on 1,050 TB patients who have been treated in Masih Daneshvari Hospital, Tehran, Iran from 2005 to 2015. Data on personal and demographic information, as well as medical data such as drug therapy, final outcome of treatment, and the diagnosis of MDR-TB, were collected from the patients' medical records. Results: The results of this study indicated that the predictive power of MDR-TB for both training and testing groups was 85% and 80%, respectively. Also, the variables of marital status, education, drug use, being imprisoned, extrapulmonary TB, history of comorbidities, AIDS, patients' age, and family size were identified as very effective factors. However, variables of residence, smoking history, contact with a TB person, pulmonary TB, drug side effects, nationality, and diabetes were found as effective factors in predicting the development of MDR-TB. Conclusion: Application of the perceptron ANNs model in the study of MDR-TB is able to create new horizons in the diagnosis of these patients due to high predictive accuracy.

Keywords

• artificial neural networks
• perceptron
• tuberculosis
• multidrug-resistant tuberculosis