AI in Health Still Leaves Most People with Disabilities Behind, Decade-Long Review Finds

Artificial intelligence is rewriting the rules of modern healthcare, from algorithms that flag tumors on scans to voice assistants that help patients manage chronic conditions at home. But a sweeping new analysis of a decade of AI research suggests that the technology’s benefits are flowing unevenly, and that the very people who stand to gain the most, the estimated 1.3 billion people worldwide living with some form of disability, are largely being left out of the picture. The finding comes from a scoping review published in PLOS Digital Health that examined how AI research in health has addressed the needs of people with disabilities between 2015 and 2025, and its numbers are sobering: of 93 studies that made the final cut, only about one in eight took concrete steps toward equity, and fewer than one in fourteen actually involved people with disabilities in designing or validating the tools built for them.

The review, led by Soutongnoma Safiata Kaboré and colleagues, set out to answer a deceptively simple question: to what extent does AI research in health actually contribute to inclusive technologies for people with disabilities? To find out, the team searched four major biomedical and scientific databases, MEDLINE, Embase, Web of Science, and CINAHL, for studies published over a ten-year window. They followed the five-stage scoping review methodology developed by Arksey and O’Malley and adhered to the PRISMA-ScR reporting guidelines, a framework designed to make systematic reviews transparent and reproducible. The initial search pulled up 4,233 references; after screening, eligibility checks, and inclusion decisions, 93 studies remained.

The geography of the included research tells its own story. Nearly three-quarters of the studies came from just two regions: North America accounted for 45.2 percent of the corpus and Europe for 26.9 percent. That concentration matters, the authors argue, because disability is not distributed evenly across the globe, and the technological priorities of high-income research ecosystems may not map onto the needs of communities where access to assistive devices, rehabilitation services, and even basic healthcare infrastructure is far more limited. An AI-powered prosthetic control system developed and tested in a North American lab, however sophisticated, says little about what is feasible or useful in settings where such devices never reach the clinic.

What the studies did cover was broad. The AI applications identified in the review targeted five major categories of disability: motor, sensory, cognitive, communication-related, and psychosocial. The dominant use cases clustered around three functions: prediction and diagnosis, assistive technologies, and personalized care. In practice, that means machine learning models trained to detect conditions earlier, smart systems designed to help people navigate their environments or communicate, and adaptive tools that tailor interventions to individual users. These are precisely the areas where AI’s pattern-recognition strengths align with the day-to-day challenges of living with a disability, which makes the field’s equity gaps all the more striking.

The review’s central finding lies in the gap between rhetoric and action. Roughly 60 percent of the included studies referenced equity in some form, whether in their framing, their limitations sections, or their discussions of future work. Yet when the researchers looked for concrete measures, adapting tools for vulnerable groups, ensuring diversity in training datasets, or accounting for the specific constraints faced by disabled users, only 12 studies, or 12.9 percent, had implemented anything of the sort. The rest treated equity as a talking point rather than an engineering requirement. In machine learning terms, this is the difference between acknowledging that a model may perform poorly on underrepresented populations and actually auditing its performance across those populations, rebalancing the training data, or redesigning the interface for users with different sensory or motor capabilities.

Even more revealing is the participation gap. Only six studies, 6.5 percent of the total, actively involved people with disabilities through co-design or user validation. Co-design, in the participatory design tradition, means that end users are not merely test subjects at the end of a development pipeline but collaborators who shape requirements, prototypes, and evaluation criteria from the start. User validation, a weaker but still meaningful form of involvement, at least puts the finished tool in front of the people it is meant to serve before deployment. That 93.5 percent of AI health studies aimed at disability did neither suggests that the field is still operating on an outdated model in which experts build for users rather than with them, a model that disability advocates have spent decades pushing back against.

The technical roots of this exclusion are well understood, even if the remedies remain underused. Algorithmic bias arises when training data fail to represent the diversity of the population a model will serve; a speech recognition system trained primarily on non-disabled speakers, for example, may struggle with the atypical speech patterns associated with some communication disabilities, effectively locking those users out of voice-based interfaces. Non-representative datasets compound the problem at every stage, from feature engineering to performance evaluation, because a model can only be as inclusive as the evidence it learns from. And when people with disabilities are absent from the design process, the resulting applications often encode assumptions, about dexterity, vision, cognition, or communication, that simply do not hold for the users they claim to target.

Why does this matter beyond the research community? People with disabilities make up 16 percent of the global population, and AI, deployed well, could enhance their autonomy, access, and quality of life in transformative ways. Predictive models could catch secondary complications before they escalate. Assistive technologies powered by computer vision or natural language processing could open up environments, physical and digital, that were previously inaccessible. Personalized care systems could adapt rehabilitation and support to individual needs rather than forcing users through one-size-fits-all protocols. The review affirms that this potential is real. But it also shows that potential is being squandered when the tools are built without the populations they are meant to serve, because exclusion at the design stage tends to reproduce exclusion at the point of use.

The authors frame the way forward in terms of socially responsive innovation. Realizing AI’s promise for people with disabilities, they conclude, will require moving beyond technical optimization, the pursuit of ever-better accuracy scores on benchmark datasets, toward inclusive design, participatory methods, and equity-driven frameworks aligned with responsible AI development and use. In concrete terms, that means treating sample diversity as a methodological requirement rather than an afterthought, building co-design into research budgets and timelines, and evaluating AI applications not only on how well they perform but on who benefits and who is burdened. It also means recognizing that equity considerations belong in the earliest stages of a project, when requirements are defined and datasets are assembled, because retrofitting inclusivity onto a finished algorithm is far harder than building it in from the beginning.

The review arrives at a moment when AI in healthcare is scaling faster than the safeguards around it. As health systems around the world begin integrating AI tools into diagnosis, care planning, and patient communication, the question of who those tools serve is no longer academic. A field in which fewer than 7 percent of studies involve the end users they target, and in which concrete equity measures appear in fewer than 13 percent, has a long road ahead. But the authors see the decade of research they surveyed not as a record of failure so much as a map of opportunity: the technical foundations are being laid, the methods for participatory and equitable design exist, and the remaining task is to connect the two. For the 1.3 billion people with disabilities worldwide, that connection cannot come soon enough.

Subject of Research: The contribution of artificial intelligence research to inclusive health applications for people with disabilities

Article Title: Contribution of artificial intelligence research to the development of inclusive applications for people with disabilities: A scoping review (2015–2025)

Article References: Kaboré, S. S., Bayani, A., Ewane, L. P. E., dos Anjos, D. S. O., Nikiéma, J. N., & Mac-Seing, M. (2026). Contribution of artificial intelligence research to the development of inclusive applications for people with disabilities: A scoping review (2015–2025). PLOS Digital Health, 5(10), e0001749. https://doi.org/10.1371/journal.pdig.0001749

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001749

Keywords: artificial intelligence, disability, digital health, health equity, assistive technology, algorithmic bias, inclusive design, participatory research, scoping review, machine learning, co-design, accessibility