The global rise in deaths caused by pulmonary tuberculosis (TB) has placed increased pressure on overburdened healthcare systems to provide TB diagnostic services. Artificial intelligence-based computer-aided diagnosis (AI-based CAD) promises to be a powerful tool in responding to this health challenge by providing actionable outputs which support the diagnostic accuracy and efficiency of clinicians. However, these technologies must first be extensively evaluated to understand their impact and risks before pursuing wide-scale deployment. Yet, health technology assessments for them in real world settings have been limited. Comprehensive evaluation demands consideration of technical safety, human factors, and health impacts to generate robust evidence and understand what is needed for long-term sustainable benefit realisation. This work-in progress study presents a three-stage methodological approach that will be used to guide the data collection and analysis process for evaluating the impact of implementing a commercial AI-based CAD system for TB diagnosis in a real-world radiological setting.

Backgrounds: Pulmonary tuberculosis (PTB) is a major health and economic burden. Accurate PTB detection is an important step to eliminating TB globally. Interferon gamma-induced protein 10 (IP-10) has been reported as a potential diagnostic marker for PTB since 2007. In this study, a meta-analysis approach was used to assess diagnostic value of IP-10 for PTB. Methods: Web of Science, PubMed, the Cochrane Library, and Embase databases were searched for studies published in English up to February 2019. The pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), the area under the curve (AUC) and hierarchical summary receiver operating characteristic (HSROC) curve were estimated by the HSROC model and random effect model. Results: Eighteen studies including 2836 total participants met our inclusion criteria. The pooled sensitivity, specificity, PLR, and NLR of IP-10 for PTB detection were 86, 88%, 7.00, and 0.16, respectively. The pooled DOR was 43.01, indicating a very powerful discriminatory ability of IP-10. The AUC was 0.93 (95% CI: 0.91-0.95), showed the accuracy of IP-10 was good. Meta-regression showed that there was no heterogeneity with respect to TB burden, study design type, age, IP-10 assay method, IP-10 condition and HIV-infection status. Conclusions: Our results showed that IP-10 is a promising marker for differentiating PTB from non-TB.