“We know that people with disabilities experience a lot of inequities in both their health outcomes and the quality of the healthcare that they receive”, says Dr Denis Newman-Griffis, Lecturer in Data Science at the Information School and co-author of ‘A roadmap to reduce information inequities in disability with digital health and natural language processing’, a paper published in PLOS Digital Health.
This narrative paper is co-authored by Dr Max Hurwitz, Dr Gina McKernan, Dr Amy Houtrow and Dr Brad Dicianno, with whom Dr Newman-Griffis worked during their post-doctoral research in biomedical informatics at the Department of Physical Medicine Rehabilitation at the University of Pittsburgh. It looks at what the sources and causes of this disparity in care experienced by people with disabilities are. Previous research shows that similar disparities are prevalent across race, class, gender and geographical lines as well, but little work has been done specifically in the area of disability.
The paper was funded in part by the National Library of Medicine at the National Institutes of Health in the USA, through Dr Newman-Griffis’ postdoctoral fellowship training programme. Work began on the paper in late 2021, with final publication in October 2022.
“We often think of ‘disability’ as an attribute of a person; it’s a category that you are either in or you’re not, but that doesn’t reflect contemporary thinking about disabilities, and it certainly doesn’t reflect peoples’ experiences of living with disabilities”, explains Dr Newman-Griffis. “Disability is a much more complex phenomenon.”
The paper argues that this fundamental misunderstanding of peoples’ lived experiences is one of the major factors in the disparity in their care. Disability frequently does have a medical component, but as a disabled person it’s your experiences outside the clinical setting - participating in society, interacting with your environment, engaging with power structures - that are more relevant, and it’s these which get lost in the translation into the medical data which is then used to make decisions about your treatment. Care decisions based only on a small subset of all the information available about a person’s disability are bound to fall short.
“We went into this research knowing that there is valuable information that is being lost in medical records, and also knowing that there is data that is recorded but is hard to get to”, says Dr Newman-Griffis.. "We used our experiences and what we found in the literature to dig in to find out what the actual issues are.”
“We often think of ‘disability’ as an attribute of a person; it’s a category that you are either in or you’re not, but that doesn’t reflect contemporary thinking about disabilities"
Through an iterative process of literature review, thematic analysis, and the researchers’ own observations, the team identified three specific barriers which lead to loss of information in a patient’s medical records. The first is the ‘contextual measurement challenge’.
“How do we clearly define and measure aspects of peoples’ lived context?”, asks Dr Newman-Griffis, explaining this challenge. “That’s not a data problem per-se, but a measurement problem.”
The second barrier is the under-emphasis of the patient’s perspective.
“What is recorded in health records is what is used by health practitioners in making their decisions, and what is recorded in health records is almost entirely practitioners’ perspectives”, Dr Newman-Griffis explains. In other words, the patient’s own view of their situation is usually not making it into their record.
The third major barrier is a technical one. Health record systems are simply not set up to record functional disability experience. Defined data fields related to this kind of input are lacking, and there’s also a dearth of methods to analyse this data if it does exist.
With information being lost at each one of these three barriers, it follows that the final patient record is more of a rough sketch than a photograph; a limited representation of a small part of what patient experience actually looks like.
The solution, the researchers argue, is digital health technologies which can draw on multiple sources and types of information to fill in the gaps in the record. Wearable devices and biometric sensors are part of this, but also technologies to simply help translate patients’ narrative descriptions into their medical record.
“Supporting people with disabilities is a huge issue globally... [this] was a way I could try and make a difference with data science.”
Natural Language Processing, or NLP, is a computational technique employed in data science to extract pieces of information from written text and put them in a format more readily understandable for analysis and decision making.
“The easiest way to capture the kind of complex and multidimensional information that relates to disability experience is with human language”, says Dr Newman-Griffis. “We can explain things clearly in natural language which are very hard to capture in defined data fields in databases.”
NLP makes this information accessible and usable at scale, allowing the large issues of data loss and information access to be filtered into discrete causes and actions.
Dr Newman-Griffis’ past experience during their PhD informed their involvement in this project. They had connected with a group at the US National Institute of Health who were working with the Social Security Administration to look at data-driven ways to improve the process for determining who should receive disability benefits. Dr Newman-Griffis’ PhD project was grounded in the idea that this kind of financial aid needs to be judged based on information about people, and identifying issues in how much of this information is available and how it can be standardised as well as analysed in order for fair decisions to be made.
“As a growing scientist, I saw in that a huge problem space with an enormous amount of unknowns”, they explain. “Supporting people with disabilities is a huge issue globally, and so being able to explore what we can do with data and design processes to help address that was something that was very appealing to me; it was a way I could try and make a difference with data science.”
This specific case highlighted wider issues with data and how disability is viewed and conceptualised, which led to future work, including ultimately this recent paper.
Dr Newman-Griffis’ experience of data science and NLP in relation to disability was a key element of the team’s skills, but each researcher brought a valuable and different insight to the table.
Dr Hurwitz’s background is also in researching health inequities related to disability, but specifically looking at them in relation to the unjust power structures which are often reflected in disability policy. Dr McKernan brought experience of methodologies employed to actually analyse disability data, whilst Drs Houtrow and Dicianno brought in their experience as clinicians, having experienced the data issues first hand whilst providing care. Dr Houtrow is also disabled, and works in advocacy, adding a further viewpoint. The mixture of practical, academic and personal perspectives was key to the paper’s narrative.
“My experience with this paper really reflects something that I’m trying to build into more of my work in this area, which is the importance and value of transdisciplinary perspectives that you get by building interdisciplinary research teams”, says Dr Newman-Griffis. “When you’re talking about data that relates to practise and personal experience, you need voices that represent all of those different experiences.”
“If we had not had any one of those voices, this would have been a much less complete argument and a much less just representation of the issues.”
Dr Newman-Griffis has two clear, intertwined strands planned for their future research. One is the intersection of data science and disability, which is where the PLOS Digital Health paper sits. Future work here will include further collaboration with clinicians, and hopefully interaction with the UK’s Department for Work and Pensions and advocacy groups, to look at the benefit schemes in this country.
The other strand is clearly linked, but comes at things from a different angle: how disability intersects with the actual processes of doing data science. Dr Newman-Griffis co-authored another paper, 'Definition drives design: Disability models and mechanisms of bias in AI technologies' - with colleagues working in critical data studies and critical disability studies - which came out in January, looking at the process of formulating an AI technology for data analysis.
“If you have some information relating to disability and you want to build an AI technology to analyse this information, that process requires you define what you mean by ‘disability’, and what information becomes relevant to that definition”, says Dr Newman-Griffis.
“This is not just a technology problem, but an issue of technology interacting with society and with humans.”
This new paper looks at AI technologies in terms of three common conceptual models of disability. Firstly: the medical model, which is deficit-focused and is often the default view of disability, seeing it as an attribute of a person. Secondly: the social model, which comes from sociology and views disability as a phenomenon related to how a person interacts with their environment and whether or not the environment supports their needs (this was the primary model used in the PLOS paper from last year). Finally, there’s the relational model, which is a more recent view of disability which has grown out of political action and advocacy. This model views disability as a political category and discusses it in relation to power structures, community and identity. Each of the models has different relevance and use in different settings.
“We did some analysis to see how you would design an AI process to address a given issue under each of the three models, and found that you get wildly different technology designs in each case, which have different potentials for harm or support”, explains Dr Newman-Griffis. This raises questions about how society is working with disability as a concept, and how it is - and should be - factored into how AI is developed, who is developing it, and how it is evaluated.
Dr Newman-Griffis adds that it’s important to remember that the three models used in this research are not the only three models to exist; there are infinite ways to describe disability, as it’s so personal.
“It really just comes down to ‘what is important to the person I’m talking to?’”, they conclude.
For Dr Newman-Griffis, the process of working on these two research papers has really brought into focus the potential implications of using data science in sensitive contexts, the power it can have, and the responsibility that data scientists have to work ethically.
“I’ve been developing an interest in responsible data science and responsible AI practices in general and these two papers have really furthered that”, they say. “This is not just a technology problem, but an issue of technology interacting with society and with humans.”
Comments