A research project reportedly involving more than 350 experts from 58 countries has seen the development of a series of recommendations with the aim of ensuring inclusivity in datasets used to train medical AI systems, hoped to allow “everyone in society to benefit from technologies which are safe and effective”.
The STANDING Together research project, published in The Lancet Digital Health, utilised a Delphi approach “supplemented by a public consultation and international interview study” to collect input from international experts on 29 consensus recommendations.
The first set of 18 recommendations on the documentation of health datasets seeks to offer guidance for “dataset curators” in supporting “transparency around data composition and limitations”.
Recommendations include ensuring a summary of a dataset “written in plain language” is included in its documentation to help users assess whether it meets their needs; making clear the reasons behind the dataset’s creation and its purpose(s); offering insight into the data’s origin, why it was selected, and “what individuals were told would happen to their data”; and sharing a summary of the groups present within the dataset, including “any known missing groups” and “reason(s) for their missingness”.
The second set of 11 recommendations on the use of health datasets aims to enable the identification and mitigation of algorithmic bias. These include providing sufficient information about datasets to allow traceability and auditability; identifying “contextualised groups of interest” in advance who may be at risk of “disparate performance or harm from AI health technology”; justifying that datasets have been “used appropriately to support the intended use population and intended use of AI health technology”; and reporting the performance of AI health technology for identified contextualised groups of interest to “enable comparison of performance for each group versus aggregate performance across the overall study population”.
Dr Xiao Liu, associate professor of AI and Digital Health Technologies at the University of Birmingham, acted as chief investigator of the study, and reflected on the potential for data to “magnify societal biases”, adding: “To create lasting change in health equity, we must focus on fixing the source, not just the reflection.”
To read the STANDING Together study in full, please click here.
Wider trend: AI regulation and managing bias in health AI
A HTN Now panel discussion from last year looked at whether the reality of AI will live up to the current hype, and how to manage bias in healthcare data. Expert panellists included Puja Myles, director at MHRA Clinical Practice Research Datalink; Shanker Vijayadeva, GP lead and digital transformation for the London region at NHS England; and Ricardo Baptista Leite, M.D., CEO at HealthAI, the global agency for responsible AI in health. The session explored topics including what is needed to manage bias; what “responsible AI” really looks like; how to ensure AI is inclusive and equitable; how AI can help support underserved populations; the deployment of AI in the NHS; and the potential to harness AI in supporting the shift from reactive to proactive care.
We asked our LinkedIn followers for their thoughts on the biggest concern for AI in healthcare: equitability, bias, transparency or regulation? 52 percent of over 100 voters highlighted regulation as their main concern, with 21 percent voting for bias. We also posted a follow-up question: what do you think is the biggest barrier to responsible AI in healthcare? Votes were closer than in the first poll, with the winning option – inadequate regulation – taking 38 percent whilst the next most popular choice – lack of standardised data – took 33 percent. In third place, 16 percent of voters said that issues with transparency are the biggest barrier to responsible AI in healthcare; whilst in fourth place, 13 percent of the vote went to bias as the biggest barriers.
In September, Health Level Seven International (HL7) published guidance on the artificial intelligence and machine learning data lifecycle, intended as an “informative document” to help developers in promoting the use of of standards to “improve the trust and quality of interoperable data used in AI models”. The paper provided a range of case studies to supplement the guidance exploring the impact AI can have in healthcare decision-making, based around a range of patient scenarios with different conditions.
NHSE also shared guidance on evaluating artificial intelligence projects and technologies with learnings from the ‘Artificial Intelligence in Health and Care Award’, which ran for four years until 2024 and supported the design, development and deployment of “promising” AI technologies; whilst NICE launched a new reporting standard designed to help improve the transparency and quality of cost-effectiveness studies of AI technologies, in a move it hopes will “help healthcare decision-makers understand the value of AI-enabled treatments” and offer patients “faster access to the most promising ones”.
