Building an evidence standards framework for artificial intelligence-enabled digital health technologies

Abstract

Health technology assessment (HTA) programmes—as exemplified by the National Institute for Health and Care Excellence (NICE) HTA programme in the United Kingdom1Raftery J Powell J Health technology assessment in the UK.Lancet. 2013; 382: 1278-1285Summary Full Text Full Text PDF PubMed Scopus (29) Google Scholar—evaluate health technologies for their clinical effectiveness and cost-effectiveness after regulatory approval. The purpose of this evaluation system is to provide robust, evidence-based guidance through which key decision makers, principally at a health-system level, can understand the clinical and economic consequences of adopting a given technology.2National Institute for Health ResearchHealth technology assessment.https://www.nihr.ac.uk/explore-nihr/funding-programmes/health-technology-assessment.htmDate accessed: July 19, 2021Google Scholar Such evaluations necessitate a systems approach, drawing on expertise from multiple stakeholder groups to ensure adequate medical, economic, patient, organisational, social, and ethical coverage. A multi-stakeholder team aims to produce a set of specific evidence standards for AI health technologies to facilitate effective and equitable evaluation strategies. In this Comment, we explain the rationale for this project and call for collaboration from digital health experts. HTA programmes are tasked with a broad remit, with responsibility for assessing “medical devices, medicines, procedures, and systems developed to solve a health problem and improve quality of lives”.3O'Rourke B Oortwijn W Schuller T Announcing the new definition of health technology assessment.Value Health. 2020; 23: 824-825Summary Full Text Full Text PDF PubMed Scopus (10) Google Scholar To effectively manage this breadth, these programmes are supported by expert-derived frameworks, which ensure uniformity between discrete assessments and identify specific evidence requirements. The emergence of artificial intelligence (AI) as a medical device, a new group of complex health technologies known as AIaMD, has posed unique challenges to HTA evaluators owing to the lack of a similarly aligned classification system for these technologies. As a result, undue reliance has been placed on non-specific digital health technology (DHT) evaluation frameworks, which were initially driven by the need to triage and assess mobile health applications. These DHT classification systems tend to focus on generic issues, such as clinical risk and functionality, but could be strengthened to aid the assessment of AI-specific complexities, which include model adaptiveness, device autonomy, limited output explainability, and the consequences of human–AI interaction in clinical settings—factors that affect the risk that an AI-centred device might present. Although classification systems for AI technologies do exist within the literature, many of which are based on underlying computational methodology, these systems were not constructed for the purpose of HTA use. Key questions for a bespoke classification system include which classification strategy is most appropriate for the purposes of HTA, how best to classify technologies in a manner that aligns with global regulatory strategies, how an AI-specific HTA classification strategy with evidence requirements can address concerns pertaining to AI technology (as outlined in the previous paragraph), and whether key stakeholders should be considering new domains of HTA assessment (for example, the effect of AIaMD on sustainability considerations). Health technology evaluation frameworks, such as the NICE DHT evidence standards framework (ESF), have provided the initial inroads into accommodating AI interventions (albeit pertaining to fixed algorithms and explicitly excluding continuously updating algorithms)4Unsworth H Dillon B Collinson L et al.The NICE Evidence Standards Framework for digital health and care technologies—developing and maintaining an innovative evidence framework with global impact.Lancet Digit Health. 2021; (published online June 24.)https://doi.org/10.1177/20552076211018617Crossref Scopus (12) Google Scholar as part of evaluation strategies. As part of the expansion of the NICE ESF to better encompass the breadth of adaptive-algorithm AI technologies that are emerging, NICE has commissioned academic partners to develop a classification framework of AI technologies that is sufficiently granular to be useful for HTA. The aim is to define, for each technology category, standards for the levels and types of evidence needed to show clinical and economic value to patients and the UK health and care system. These standards include evidence of effectiveness relevant to the intended use(s) of the technology and evidence of the economic effect. The framework can be used for decision making by those who develop, finance, and deploy AI technologies, including innovators, technology developers, and commercial organisations, in addition to commissioners, research funders, and other investors who are considering funding the development of data-driven technologies that incorporate AI. A key strength of this programme is the close engagement of NICE with global regulators and the direct contribution of the Medicines and Healthcare Products Regulatory Agency (MHRA). Since 2016, regulators5Benjamens S Dhunnoo P Meskó B The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database.npj Digit Med. 2020; 3: 118Crossref PubMed Scopus (181) Google Scholar—such as the US Food and Drug Administration (FDA), EU competent authorities, and the MHRA—have seen an increasing number of AI-based health technologies brought to market for diagnostic,6US Food and Drug AdministrationFDA permits marketing of clinical decision support software for alerting providers of a potential stroke in patients.https://www.fda.gov/news-events/press-announcements/fda-permits-marketing-clinical-decision-support-software-alerting-providers-potential-strokeDate: Feb 13, 2018Date accessed: July 26, 2021Google Scholar prognostic,7The LOOP BlogFDA approves the guardian connect system, the world's first “smart CGM” to help people outsmart their diabetes.https://www.medtronicdiabetes.com/loop-blog/fda-approves-the-guardian-connect-system-the-worlds-first-smart-cgm-to-help-people-outsmart-their-diabetes/Date: March 12, 2018Date accessed: July 26, 2021Google Scholar and therapeutic8US Food and Drug AdministrationFDA clears mobile medical app to help those with opioid use disorder stay in recovery programs.https://www.fda.gov/news-events/press-announcements/fda-clears-mobile-medical-app-help-those-opioid-use-disorder-stay-recovery-programsDate: Dec 10, 2018Date accessed: August 9, 2021Google Scholar purposes. In October 2021, guidance jointly produced by the MHRA, FDA, and Health Canada was published to set out additional considerations (Good Machine Learning Practice principles) for AIaMD.9Medicines and Healthcare Products Regulatory AgencyGood machine learning practice for medical device development: guiding principles.https://www.gov.uk/government/publications/good-machine-learning-practice-for-medical-device-development-guiding-principlesDate: Oct 27, 2021Date accessed: November 5, 2021Google Scholar Close working relationships with regulators provide opportunities for this ESF to harmonise evidential requirements that are being considered by both regulator and HTA, thereby increasing efficiencies for all stakeholders. AI health technologies are a global opportunity. HTA evaluators and regulators worldwide are tackling the same challenges. As part of the development of the NICE ESF we aim to produce a classification system for AI health technologies that is based on common HTA principles and can be used for HTA evaluations worldwide. We invite international stakeholders—including clinicians, AI academics, industry representatives, policymakers, regulators, funders, bioethicists, legal experts, and patient representatives—to contribute to this open and transparent development process, as we work together to provide a consensus-driven framework for the effective and efficient evaluation of AI health technologies for the benefit of patients and health systems. AW acknowledges support from a Turing AI Fellowship under grant EP/V025379/1, The Alan Turing Institute, the Leverhulme Trust via the Leverhulme Centre for Future Intelligence (CFI), and Engineering and Physical Sciences Research Council (EPSRC) grants EP/V056522/1 and EP/V056883/1; is an advisor to retrain.ai (received options); and is a board member and senior advisory board member of GNS Healthcare (Boston) and a board member of its parent company GNS (Gene Network Sciences; receives options), and an advisory board member (paid) of the UK Government Centre for Data Ethics and Innovation. AD is the chairman of the Preemptive Medicine and Health Security Initiative of Flagship Pioneering. CH has consulted for Norvartis and the Novo Nordisk Foundation; is supported by grants from the EPSRC, the Medical Research Council, and the Department of Health and Social Care of the UK Government; and is on the international scientific advisory board of the UK Biobank. XL acknowledges support from the Medicines and Healthcare Products Regulatory Agency, the Wellcome Trust, the National Institute of Health Research, NHSX, and the Health Foundation. HA is the chief scientific officer of the Preemptive Medicine and Health Security Initiative of Flagship Pioneering. All other authors declare no competing interests. Funding and infrastructure support for this research is provided by the NHS AI Lab, NHSX, London, UK.

Topics & Concepts