Are disabled people shaping health-AI research?
A review of 93 studies found only six used co-design, usability testing or user validation with disabled people. The evidence maps a participation gap, not the performance of every disability-focused AI tool.
The short version
3 key points
- 1.The reviewers screened 3,820 unique records from four databases and included 93 studies published between January 2015 and May 2025.
- 2.Only six of the 93 studies (6.5%) used a participatory approach such as co-design, usability testing or user validation with disabled people.
- 3.Sixty-eight studies (73.1%) came from high-income economies; the authors also found small samples, heterogeneous data and limited external validation across the field.
The answer: disabled people were rarely active research partners
A decade of health-AI research aimed at disability has produced diagnostic models, rehabilitation tools and assistive systems, but disabled people were seldom involved in shaping or validating them. A peer-reviewed PLOS Digital Health scoping review included 93 studies published from January 2015 to May 2025. Only six studies—6.5% of the set—reported participatory work such as co-design, usability testing or user validation. Twelve studies, or 12.9%, put equity into practice through measures such as accessibility adaptations, attention to sample diversity or cultural considerations.
That does not mean the other 87 studies were technically poor, deliberately exclusionary or unusable. A scoping review maps what authors reported; it does not independently retest every model or device. Its central finding is narrower and consequential: the people a system is meant to support were rarely visible as co-creators, and explicit equity measures were much less common than general statements about inclusion. For patients, carers, clinicians and public-service buyers, that is a warning to ask who helped define the problem, whose data shaped the system and whether intended users tested it in the conditions where it will operate.[1]
How the review was conducted
The team worked with a Université de Montréal public-health librarian to search MEDLINE, Embase, Web of Science and CINAHL in English and French. The search covered January 2015 through May 2025 and combined terms for artificial intelligence, inclusion and disability. The reviewers followed the Arksey and O’Malley scoping-review method and PRISMA-ScR reporting guidance, using a population-concept-context framework: disabled people; AI, equity, diversity and inclusion; and health-related applications.
Four databases returned 4,233 records: 2,306 from Embase, 1,354 from MEDLINE, 300 from Web of Science and 273 from CINAHL. After 413 duplicates were removed, three reviewers screened 3,820 titles and abstracts, selected 201 full texts and retained 93 studies. Disagreements were resolved with additional reviewers. The team extracted study location, disability type, AI model, application and predefined equity indicators, including co-design, disabled-user involvement, sample diversity, accessibility adaptations, cultural responsiveness and attempts to identify or mitigate algorithmic bias.[1]
The evidence was concentrated in richer countries
Sixty-eight of the 93 studies (73.1%) came from high-income economies and 25 (26.9%) from low- and middle-income economies. The United States contributed 28 studies, while China contributed 13. By broad region, the paper reports North America at 45.2%, Europe at 26.9%, Asia at 28.0%, Oceania at 4.3% and Africa at 1.1%. Some studies spanned more than one country or region, so these regional percentages should not be treated as mutually exclusive shares that must sum to 100.
Geography matters because language, infrastructure, care pathways, disability policy and the availability of specialist support change what an AI system must do. A model trained around a well-resourced clinic may not transfer to a community service with intermittent connectivity or scarce rehabilitation staff. But the review did not test whether funding levels caused better inclusion, and it did not perform a formal time-trend analysis. Its map shows an imbalance in the published research base, not a ranking of countries’ disability rights or AI capability.[1]
What kinds of disability and applications appeared
Cognitive and neurodegenerative conditions were the largest broad category, appearing in 46 studies (49.5%). These included Alzheimer’s disease, dementia, mild cognitive impairment, intellectual disability, Down syndrome and dyslexia. Motor disabilities appeared in 28 studies (30.1%), sensory disabilities in nine (9.7%) and communication disabilities in seven (7.5%). The authors caution that these broad groupings combine conditions with different trajectories, accommodation needs and questions of consent, so a category-level result cannot be assumed to apply uniformly.
The applications ranged from clinical and behavioural prediction to automated classification, early screening, clustering, assistive technologies, decision support and language analysis. Examples included post-stroke recovery prediction, robotic orthoses, brain-computer interfaces, retinal screening, audiogram interpretation, autism-related behavioural analysis and speech support. This breadth explains why a single accuracy figure would be meaningless. It also shows why inclusion cannot be reduced to representative training data: physical access, communication format, consent, caregiver involvement and whether the tool supports autonomy can matter as much as model performance.[1]
Participation was the clearest gap
The six participatory examples included usability testing for a robotic rehabilitation device in Mexico, an accessibility-led tongue interface in Fiji, assessment of AI-generated health information in Canada, attention to communication and literacy in US preschool language-disorder work, analysis of bias affecting neurodivergent people, and user-centred validation of emotion-recognition tools for children with Down syndrome in Mexico. Two studies explicitly addressed equity in AI models: one examined biological bias in Alzheimer’s drug-discovery data and one focused on representative modelling for facial-emotion recognition.
These counts are best read as minimum documented practice. A study may have consulted users without reporting it clearly, while the review excluded grey literature and conference work where early co-design is sometimes described. Even so, missing reporting is itself a governance problem. Commissioners and ethics reviewers cannot assess participation that is invisible. A credible claim of inclusive design should say which users took part, at what stage, how access needs were met, what changed because of their input and whether participants had power over the definition of success.[1]
What this means for patients, carers and professionals
For disabled people and carers, the practical risk is not only a biased prediction. It is a product that solves the wrong problem, requires inaccessible interactions, burdens a caregiver, or works in a demonstration but not in daily life. Clinicians and rehabilitation teams should ask whether validation included people with the relevant functional and communication needs, not merely whether the diagnostic label appeared in a dataset. Performance should be reported across meaningful subgroups, with failures and non-use recorded rather than treating completion by a selected sample as proof of accessibility.
For public-service professionals, procurement can make participation testable. Contracts can require evidence of co-design, accessible documentation, compatibility with assistive technologies, routes for correction and withdrawal, and monitoring for unequal error or abandonment rates. An independent accessibility assessment should include real tasks and real users. The review supports these questions, but it does not identify a universal threshold for acceptable representation or prove that participatory design automatically produces a safe, effective tool.[1]
Limits of the review
The search covered four bibliographic databases and only English- or French-language full texts. It excluded grey literature, theses, preprints, conference proceedings and many community-led projects. Those choices improve consistency but may undercount innovation in lower-resource settings and participation reported outside journals. The eligibility criteria also allowed original quantitative and qualitative studies as well as peer-reviewed commentaries, creating a heterogeneous evidence set that a scoping review can describe but not combine into an effect estimate.
Most included work took place in clinical settings, so the review says less about community-based disability support. The authors did not formally assess change over time and did not claim a causal connection between investment and inclusion. Their broad disability categories can blur important differences. Finally, the six-study participation count relies on publication reporting. The result is therefore a map of visible practice and gaps, not a clinical effectiveness review and not evidence that a particular product is unsafe.[1]
Funding, independence and what would change the assessment
The review was funded by the Université de Montréal Vice-Rectorate for Planning and Strategic Communication through an equity, diversity and inclusion call linked to the Montréal Declaration for Responsible AI. The paper says the funder had no role in design, data collection, analysis, publication or manuscript preparation, and the authors declared no competing interests. All review data are reported in the article or supporting information.
Confidence would rise with multilingual searches, systematic inclusion of grey literature and community-led work, and a quality appraisal separating exploratory prototypes from externally validated systems. More importantly, future studies should preregister participation plans, report recruitment and accommodations, and measure whether co-design changes safety, use, autonomy and quality of life. Independent evaluations across income settings should test whether systems retain performance and accessibility in routine care. Until then, the strongest conclusion is about the research process: disability-focused AI has often studied disabled people without giving them a documented role in shaping the technology.[1]
What this means for people
- Disabled people may face tools that are technically promising but poorly matched to their access, communication and daily-life needs when they are not involved in design and validation.
- Clinicians and rehabilitation teams need evidence from users with the relevant functional needs, not only aggregate model accuracy.
- Public-service commissioners can require documented co-design, accessible testing, correction routes and subgroup monitoring before deployment.
Global context
The 93 studies were concentrated in high-income economies, with very limited African representation. The review included English and French publications only, so it may miss important community and research activity in other languages and outside indexed journals.
What the evidence does not yet show
- Scoping review: it mapped published research but did not pool effectiveness or independently retest tools
- English- and French-language indexed literature only, excluding grey literature, preprints and conference proceedings
- Broad disability categories combine conditions with different access, consent and clinical needs
- Participation was counted from published reporting and may be underreported
- No formal time-trend analysis and no causal test of why some settings or studies were more inclusive
What to watch next
- Preregistered co-design and disabled-user validation in disability-focused AI studies
- Multilingual and community-led evidence from low- and middle-income settings
- Accessible real-world evaluations that measure autonomy, abandonment, harms and caregiver workload
- Subgroup performance reporting and correction routes in procurement standards
- Independent replication of systems now supported mainly by small or internally validated studies
Living evidence record
Impact record IAI-064F3XM
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what PLOS Digital Health published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether health-AI research for disabled people turns equity language into representative data, accessible design and participation
Topics: Disability · Inclusive design · Algorithmic bias · Clinical AI · Assistive technology · Research participation
Regions: Global · North America · Europe · Asia · Africa · Oceania
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
AI basics · Background explainer
Published 2018-11-28 · 2 min 31 sec
What is artificial intelligence? | The Royal Society
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
AI Risks & Safety
Can hidden image text mislead dental AI?
A peer-reviewed German stress test found that adversarial text placed inside 270 dental radiographs could flip four vision-language models from an abnormal to a normal finding. OCR sanitisation sharply reduced the measured attacks, but the experiment used a permissive prompt, a pathology-heavy benchmark and no live clinical system.
7 min · 3 sources
AI Risks & Safety
Can an ICU AI model survive a hospital change?
Three mortality models lost reliable decision behaviour when moved between MIMIC and eICU. One random forest made no positive predictions after transfer, showing why a respectable AUC cannot substitute for local calibration and threshold testing.
8 min · 1 source
AI Risks & Safety
Can small AI models replicate across ordinary devices?
Anthropic's controlled evaluation shows that advanced models can assemble inference on varied hardware and that GLM-5.2 completed a simplified eight-node replication chain. It did not demonstrate an autonomous worm escaping onto the internet: access was deliberately easy, the agent model ran elsewhere, and the near-term timeline is a projection.
7 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.