Conditions required for the artificial-intelligence-based computer-aided detection of tuberculosis to attain its global health potential

Abstract

In 2021, WHO issued a novel recommendation within its tuberculosis screening guidelines: the approval of artificial-intelligence-based computer-aided detection (AI-CAD) to analyse chest x-rays for tuberculosis detection in place of human readers.1WHOWHO consolidated guidelines on tuberculosis: module 2: screening: systematic screening for tuberculosis disease. World Health Organization, Geneva2021Google Scholar The recommendation was largely based on evidence suggesting that the accuracy of AI-CAD approximates that of radiologists in identifying tuberculosis on chest x-rays.1WHOWHO consolidated guidelines on tuberculosis: module 2: screening: systematic screening for tuberculosis disease. World Health Organization, Geneva2021Google Scholar Global health donors and actors working to eradicate tuberculosis regard AI-CAD as an important tool for finding the so-called missing millions of people with active tuberculosis that is left undetected each year.2Tzelios C Nathavitharana RR Can AI technologies close the diagnostic gap in tuberculosis?.Lancet Digit Health. 2021; 3: e535-e536Summary Full Text Full Text PDF PubMed Scopus (1) Google Scholar Donors are also drawn to AI-CAD's potential for optimising resource allocation by reducing the use of costly confirmatory diagnostics, such as the GeneXpert MTB/RIF assay.3Qin ZZ Sander MS Rai B et al.Using artificial intelligence to read chest radiographs for tuberculosis detection: a multi-site evaluation of the diagnostic accuracy of three deep learning systems.Sci Rep. 2019; 915000Crossref Scopus (136) Google Scholar In the face of the devastating effects of COVID-19 on tuberculosis care and prevention, AI-CAD has been highlighted among the tools that can be used to make the goal of tuberculosis eradication technically and programmatically possible.4Pai M Kasaeva T Swaminathan S COVID-19's devastating effect on tuberculosis care—a path to recovery.N Engl J Med. 2022; 386: 1490-1493Crossref PubMed Scopus (87) Google Scholar This excitement around AI-CAD for tuberculosis detection has emerged from an evidence base that is nearly singularly focused on estimating accuracy. However, ongoing experiences with implementation of AI-CAD for tuberculosis detection invite the global health community to consider more multifaceted critical assessments. AI-CAD has often been tested in the framework of pilot projects and research partnerships, but these tests have not always led to uptake by state authorities because many considerations have yet to be addressed. In this Comment, we identify and discuss technical, economic, and political considerations surrounding the use of AI-CAD for tuberculosis detection to provide a framework for its implementation in a manner that enhances health equity. Although there are 17 AI-CAD competitors in the market, only two are in the catalogue of the Global Drug Facility of the Stop TB partnership, the world's leading procurement and supply mechanism for tuberculosis control devices.5Stop TB PartnershipGlobal Drug FacilityDiagnostics, medical devices & other health products catalog.https://www.stoptb.org/sites/default/files/gdfdiagnosticsmedicaldevotherhealthproductscatalog_0.pdfDate: September, 2022Date accessed: September 1, 2022Google Scholar Many developers emphasise their compliance with European (ie, CE) or Chinese (ie, NMPA) certifications; however, these certifications do not completely guarantee the quality of the product nor an evaluation by a stringent regulation authority, such as the US Food and Drug Administration. Some certifications can be based on a self-declaration, as is the case of some CE certification levels.6EUCE marking.https://europa.eu/youreurope/business/product-requirements/labels-markings/ce-marking/index_en.htmDate accessed: May 26, 2022Google Scholar Moreover, it is not clear how issues related to software version updates should be regulated, which is particularly important for AI-CAD given the rapid pace of software development. For example, existing scientific literature on AI-CAD efficiency cannot be extrapolated to make inferences on the performance of new software versions. Software versions are particularly important because, if generalisability of accuracy and implementation data is downplayed, then clinical and programmatic decisions can be misguided. As more accurate versions are brought to market, health equity will be undermined if sites that purchased the previous versions have to use additional, typically scarce, resources to access updates. Regulations should be adapted to also address and include these health equity issues. Although WHO approved the use of AI-CAD for tuberculosis screening and opened up a global market, the absence of regulation with clear specifications maintains uncertainty. These specifications on both efficiency and technical aspects should be clarified to be discussed technically, programmatically, and politically by the stakeholders with complete transparency. Evaluative studies have consistently shown that AI-CAD solutions should be calibrated by use of local data from the setting where they will be deployed.7Tavaziva G Harris M Abidi SK et al.Chest x-ray analysis with deep learning-based software as a triage test for pulmonary tuberculosis: an individual patient data meta-analysis of diagnostic accuracy.Clin Infect Dis. 2022; 74: 1390-1400Crossref PubMed Scopus (17) Google Scholar However, interviews with users indicate that other factors influence their use, such as the logistical and financial availability of GeneXpert MTB/RIF assays that are used to confirm the presence of tuberculosis. The effectiveness and sustainability of AI-based tuberculosis screening thus depends, in part, on its integration with other technologies and, ultimately, with access to medicines and care. Health-system capacity for the people with detected tuberculosis (eg, number of GeneXpert MTB/RIF assays that are available) further influences algorithm use. Rather than finding the optimal calibration for driving down rates of tuberculosis, it might become more politically efficient to find a standard calibration that fits well with local political and programmatic realities. This adaptation to programmatic realities would lead to a missed opportunity to call out for structural changes, such as a dramatic increase in tuberculosis testing, funding, and capacity. Instead of transforming the diagnostic environment, therefore, AI-CAD could fit into the existing diagnostic landscape without being as disruptive and revolutionary as advertised. The conditions of calibration and use of AI-CAD in real life need to be discussed and clarified to achieve their greatest effectiveness. Indeed, the current approach to use of AI-CAD is based on selecting a threshold score that balances sensitivity with the interest to reduce the use of costly sputum-based molecular tests; however, in adopting this approach, programmes are explicitly accepting a proportion of people with missed tuberculosis identification. A patient-centred, and indeed, human rights-centred implementation would consider alternatives where AI-CAD enables longitudinal screenings, or the development of prediction scores, to maximize sensitivity. AI-CAD development companies support different business models (excluding models for chest x-ray hardware). On one end of the scale is a model based on a licence fee, which is paid depending on the number of reads or the number of months during which the AI-CAD is used. Charging per x-ray reading (often US$0·45–0·95) is rarely considered because developers usually support AI-CAD readings for pilot or research projects.8Qin ZZ Naheyan T Ruhwald M et al.A new resource on artificial intelligence powered computer automated detection software products for tuberculosis programmes and implementers.Tuberculosis. 2021; 127102049Crossref PubMed Scopus (22) Google Scholar Both pay-per-read and pay-per-month have the disadvantage that the more you diagnose (ie, the more you test), the more you pay, which goes against the grain of scaling up detection. Indeed, both pay-per-read and monthly subscription business models, in contrast to perpetual licensing models at the other end of the scale, could be obstacles to achieving the objectives promised by AI-CAD technologies. The Global Drug Facility catalogue presents technological solutions that propose a business model on the basis of only the perpetual licence (plus the price of maintenance and the box containing the algorithm that can be used online or offline).5Stop TB PartnershipGlobal Drug FacilityDiagnostics, medical devices & other health products catalog.https://www.stoptb.org/sites/default/files/gdfdiagnosticsmedicaldevotherhealthproductscatalog_0.pdfDate: September, 2022Date accessed: September 1, 2022Google Scholar This solution holds the advantage that the price is independent of the quantity of detection. Some procurement platforms, such as the Global Fund's Wambo, might propose alternatives, such as tier-based licensing, in which cost and services vary with volume. However, long-term maintenance, technical support, and access to updated versions might not be offered within perpetual or tier-based models. Older and less performant versions might stay on the market, representing equity issues about who can or cannot afford to regularly update the software for their populations. There is a need for greater transparency with respect to the specific conditions of these economical contracts. The economic condition for AI-CAD technologies to achieve their objective is to respect open business models and move away from models that limit the quantitative use of the software. As scale-up of AI-CAD is underway, funders and advocacy groups should not miss the opportunity to push for, or develop, novel payment models driven by public health and equity goals over profit generation. The data economy is increasingly a central component of the valorisation and legitimacy of global health programmes. Development of AI-CAD necessitates access to vast amounts of medical imaging data. The geographical distribution of the global tuberculosis pandemic means that low-income and middle-income countries have great potential as generators of imaging data that can be used to fuel the advancement of AI-CAD for analysing chest x-rays for tuberculosis detection. It is therefore important to avoid practices of self-evident data capturing in connection with the implementation of digital diagnostic systems. Tangibly, it will be important to ensure that AI-CAD developers are not permitted to offer financial incentives to AI-CAD users in exchange for access to radiological images, even if the images are rigorously anonymised such that patient identification is impossible. Without such regulation, populations affected by tuberculosis in low-income and middle-income countries might be used to advance a technology that they are also, directly or indirectly, paying to use. As contracts between software developers and users are confidential, it is unknown whether any previous or existing contracts have included such exchange clauses. These are issues of data sovereignty9Hummel P Braun M Tretter M Dabrock P Data sovereignty: a review.Big Data Soc. 2021; 8 (2053951720982012)Crossref PubMed Scopus (83) Google Scholar and patient rights over their own data but are also concrete conditions for the success, or otherwise, of AI-CAD to contribute to the fight against tuberculosis. In summary, in addition to requiring a high level of diagnostic accuracy, for AI-CAD technologies to achieve their stated objectives particular technological, economic, and political conditions should be addressed. First, a consensus on precise specifications is necessary so that programmes know the quality and limits of what they are purchasing and have a clear understanding of how version updates will be regulated and distributed. Second, the commercial logic of price-per-reading and locked and non-compatible systems represent major obstacles to use. Third, it is important to avoid capture and appropriation of data acquired and produced by AI-CAD. At a time of growing efforts to decolonise global health,10The Lancet Global HealthGlobal health 2021: who tells the story?.Lancet Glob Health. 2021; 9: e99Summary Full Text Full Text PDF PubMed Scopus (29) Google Scholar to view health as a right and a necessity for those in need, we should make sure that technological solutions, such as AI-CAD, do not embody political power relations that reproduce past inequities. FAK declares operating grants and salary support from publicly funded research agencies to study computer-aided detection of tuberculosis on chest x-rays (the Canadian Institutes of Health Research, Fonds de Recherche Québec Santé, and Observatoire international sur les impacts sociétaux de l’IA et du numérique [from the Fonds de Recherche de Québec]). FAK reports participating in a technical consultation on WHO prequalification requirements for computer aided detection software for tuberculosis. FAK reports that the following developers of computer-aided detection software provided his research group with either free or reduced pricing access to their software for evaluative research, and that the groups did not have any role in the study design, analysis, result interpretation, or decision to publish previous research and the submitted work: Delft (Netherlands, makers of CAD4TB), qure.ai (India, makers of qXR), and Lunit (South Korea, makers of LUNIT INSIGHT). P-MD declares operating grants from publicly funded research agencies to study computer-aided detection of tuberculosis on chest x-rays (the Canadian Institutes of Health Research, Fonds de Recherche Québec Santé, and Observatoire international sur les impacts sociétaux de l’IA et du numérique [from the Fonds de Recherche de Québec]). JO declares operating grants and salary support from publicly funded research agencies to study computer-aided detection of tuberculosis on chest x-rays (Observatoire international sur les impacts sociétaux de l’IA et du numérique [from the Fonds de Recherche de Québec]). SK is the chair of the Steering Committee of the Zero TB Initiative, an alliance of communities committed to eliminating tuberculosis. AI-CAD for tuberculosis and other global high-burden diseasesWe applaud Pierre-Marie David and colleagues1 for their Comment in The Lancet Digital Health on the conditions required for artificial intelligence (AI)-based computer-aided detection (CAD) tools to attain their global health potential. David and colleagues underscore the fundamental challenges that must be addressed to allow AI-CAD, and more broadly, AI diagnostics, to achieve the greatest effect. We believe these challenges call for partnerships across the digital health ecosystem to accelerate the ability to realise this potential. Full-Text PDF Open Access

Ähnliche Arbeiten