Abstract

Objectives Diabetes mellitus significantly increases the risk of tuberculosis, and active tuberculosis screening of people with diabetes mellitus has been advocated by WHO and other international bodies. This systematic review aimed to evaluate the accuracy of computer-assisted detection for identifying pulmonary tuberculosis among people living with diabetes mellitus. Methods Medline, Embase, Scopus, Global Health, and Web of Science were searched from January 2010 to May 2024 using MeSH headings and keywords, supplemented with grey literature searches (Conference abstracts, Trial registries, MedRxiv.org, Google Scholar). Studies evaluating computer-assisted detection diagnostic accuracy for identifying tuberculosis in populations living with diabetes mellitus were included. Two researchers independently assessed titles, abstracts, and full texts, extracted data, and assessed the risk of bias using the QUADAS-2 instrument. Forest plots and Summary Receiver Operating Curves were generated using RevMan 5.4, and statistical pooling of studies was carried out in STATA v18 using the bivariate model. Results Five eligible studies, all conducted in Asia between 2013 and 2023, were identified, including a total of 1879 individuals with diabetes mellitus, of whom 391 were newly diagnosed with tuberculosis. Four different computer-assisted detection software algorithms were used. The pooled sensitivity was 0.94 (95% CI: 0.85–0.97) and specificity was 0.77 (95% CI: 0.68–0.84). Area Under the receiver operating curve values varied from 0.7 (95% CI: 0.68–0.75) to 0.9 (95% CI: 0.91–0.96). False positive proportions ranged from 0.24% to 30.5%, while false negative proportions were 0%–3.2%. Overall heterogeneity was high (i2 55% for sensitivity and 93% for specificity) but much lower for sensitivity among the three studies using the same computer-assisted detection software (i2 0% for sensitivity; 93% for specificity). The risk of bias of the five studies was generally very low, although detailed information about diabetes management was lacking. Conclusions Computer-assisted detection tools show potential in screening people living with diabetes for active tuberculosis and appear to show good sensitivity at the thresholds indicated, but data are scarce and performance varies across settings.