Artificial Intelligence in Drug-Resistant Tuberculosis Diagnosis. Systematic Review & Meta Analysis

Abstract

Abstract Background Artificial Intelligence (AI) models possess potential screening or diagnosing Drug-Resistant Tuberculosis. This meta-analysis aimed to elaborate on the performance of AI models in predicting drug-resistant TB compared to the gold standard. Methods A systematic search of full-paper articles from Pubmed/MEDLINE and SCOPUS published between January 2005 and April 2020 was conducted. Studies involving whole-genome sequencing or human subject data that used drug-susceptibility test results as a reference standard were retrieved. Hierarchical summary receiver-operating characteristic (HSROC) and bivariate model were performed to calculate pooled sensitivity and specificity. Heterogeneity and publication bias were also assessed. Results 25 Models form four genome-based studies, four radiology-based studies, and two clinical and demographic-based studies were included. The radiology-based models, mainly built with the Convolutional Neural Network (CNN), possessed pooled sensitivity of 58% (95%CI 50%-65%), and a pooled specificity of 75% (95% CI 47%-91%) for detecting MDR-TB. Genome-based models outperformed other models despite developed with simpler classifiers (CART/GBT/LR). The genome-based studies reached pooled sensitivity of 92% (95%CI 90-93%), pooled specificity of 98 (95% CI 96%-98%) in predicting isoniazid resistance. For predicting rifampicin resistance, the pooled sensitivity and specificity were 93% (95%CI 90%-94%) and 98% (95% CI 96%-98%) respectively. The Artificial Neural Network (ANN) demonstrated superiority rather than the Classification and Regression Tree (CART), and logistic regression (LR) on the clinical and demographic model in predicting drug-resistant TB. Conclusions Simple classifier performs better in genomic data whereas the CNN model works best in high-dimensional data, such as radiology images. The ANN model indicates a preferred model for data obtained from clinical and demographic parameters. (PROSPERO number CRD42020167439) Key messages Artificial Intelligence model through machine learning and deep learning possess a good diagnostic performance for drug-resistant tuberculosis. Artificial Neural Network model outperforms other technique. The model built from genomic data shows the best performance, followed by model from patient-based data and radiology/imaging data.