ABSTRACT
Introduction
Research on persons with profound intellectual and multiple disabilities (PIMD) has been limited and faces unique challenges, including accessibility, methodological issues, and ethical considerations, especially when using AI-based technologies. Negative attitudes toward technology from gatekeepers exacerbate these challenges.
Methods
The text summarizes the experience of studies that explores AI’s application in recognizing emotional states in individuals with PIMD through complex and multimodal data collection, including video, facial expressions, and vocalizations. Methodological challenges include the need for extensive data and the heterogeneity of the target group. AI has been utilized to interpret behaviors of people with PIMD, albeit with limitations in data processing and AI performance.
Results
Ethical concerns arose regarding privacy and the comprehensiveness of data collection, impacting the participation and utility of AI in research with vulnerable populations.
Conclusions
Despite substantial impediments, AI research offers potential quality-of-life benefits for people with PIMD but requires an ethically reflective approach.
KEYWORDS:
INTRODUCTION
For a long time, persons with profound intellectual and multiple disabilities (PIMD) have been neglected from research (Munde & Zentel, Citation2020). Serious research efforts have been made only in the last 30 years, but they still face difficulties and are not carried out in the same number and intensity as research on people with milder forms of intellectual disabilities (ID) (de Haas et al., Citation2022).
With regard to available research, it is striking that empirical studies with the direct participation of people with ID and especially people with PIMD are rarely conducted and mostly third parties (i.e., parents or professional caregivers) are interviewed by proxy (Maes et al., Citation2021; Talman et al., Citation2019). There are multiple reasons for the difficulties in directly involving people with PIMD in research: (1) accessibility of the target group, (2) methodological issues and (3) research ethics issues. These difficulties are exacerbated in research on artificial intelligence (AI)-based technologies, as our findings have shown (Engelhardt et al., Citation2019; Hammann, Schwartze, et al., Citation2022; Hammann, Valič, et al., Citation2022; Schlomann et al., Citation2021; Zentel et al., Citation2019).
ACCESSIBILITY OF THE TARGET GROUP
People with PIMD are a vulnerable group (Munde & Zentel, Citation2020) who are dependent on close support and care (Matérne & Holmefur, Citation2022). In this respect, it is understandable that those responsible for their protection (i.e., parents or professional caregivers) are partially critical of complex and innovative research projects and associated interventions. However, the function of gatekeepers also involves a position of power that needs to be reflected upon. In the context of new media and AI, it is particularly difficult to separate one’s own attitudes and assessment of the expected benefits or risks from the objective evaluation that might be anticipated from people with disabilities. Our discussions with gatekeepers (i.e., institutions, relatives, legal guardians) during the course of our research were often dominated by negative views of technologies. Sentences such as “I don’t think AI is useful” or “I don’t like new technologies and AI in particular” were heard repeatedly and cited as reasons for rejecting research involvement. As a result, negative attitudes toward the media and AI-based technologies are transferred to people with PIMD without any certainty whether those individuals share such views. Prejudice and negative mind-set toward technologies are not uncommon for employees in disability care (Heitplatz et al., Citation2020). As an essential aspect of professionalism, it is crucial for employees in services for people with disabilities to reflect on their personal viewpoint and the boundaries of the caring role. The aim must be to identify potential benefits and harms for the person with PIMD, regardless of the carer’s personal views.
METHODOLOGICAL ISSUES
AI and associated research are data-intensive. The more data processed, the higher the likelihood that machine learning processes will be successful, and the AI can be better trained to achieve the best possible results. This is also referred to as Big Data (Kleppmann, Citation2017). For individuals with PIMD, it is difficult or even impossible to obtain large and especially comparable and transferable data. The significant heterogeneity of the target group (Munde & Zentel, Citation2020) means that data-based patterns or algorithms cannot be transferred from one person to another. Instead, databases must be “fed” with as much information as possible about a person (e.g., regarding preferences, behaviors, and their expressions through facial expressions, gestures, and speech/vocalizations). Moreover, especially in the context of people with PIMD, complex and multimodal data (videos of movements, pictures of facial expressions, audio files of vocalizations) need to be collected and correlated. In the project referenced here, we used these data and AI to automatically recognize the emotional states of individuals with PIMD, to provide caregivers with clues for interpreting displayed behaviors. The expressions of individuals with PIMD are often unconventional and difficult for unfamiliar caregivers to interpret (Hammann, Schwartze, et al., Citation2022).
Those requirements have led to various challenges in research with this population group. First, countless hours of video material may have to be recorded, coded and interpreted in a joint process with close caregivers. The results are then integrated into an AI database. Only on this basis AI can be applied and recognize and name behaviors in the person. Therefore, a significant investment is required to initiate the machine learning process. A known drawback is that the performance of the trained machine learning models is at most average, without gaining entirely new insights. This raises the question of cost versus benefit. Furthermore, not all data can be used, as information about individuals with PIMD (e.g., how the person shows anger or joy) is variable and indeterminate even by relatives and professional caregivers. However, real world data that may also be ambiguous cannot be processed by an AI-based database. Ambivalences and contradictions that professional educators must deal with are beyond the performance spectrum of AI systems, at least at the current state of technology, thus limiting their utility in practice.
ETHICAL ASPECTS OF AI RESEARCH AND VULNERABLE POPULATIONS
The more comprehensive the data collected and the more life situations that need to be included in research, the greater the ethical conundrums to be resolved. Particularly in the case of people who, due to severe cognitive limitations, are not or only partially able to understand in what form and for what purpose they are involved in research activities and what their data are being used for (Hammann, Citation2022). On the one hand, this problem can and must be countered by interviewing proxies. On the other, a constant evaluation of the person’s emotional status during the research process is necessary in an effort to clarify whether the person is signaling agreement or disagreement with the research process. Finally, there are difficulties in fully understanding the individual statements of a person (Krämer & Zentel, Citation2020), as s/he communicates on a pre-symbolic or unusual way (Munde & Zentel, Citation2020).
Resolving such challenges but also ensuring that sufficient data are collected in order to make valid conclusions also means that many different situations have to be included which may compromise privacy of the individual. For example, if an AI assistant with a built-in camera is installed in the bedroom or a communal space at all times, it is challenging for the resident with PIMD to understand when recordings are being made. It is also necessary to deal with the fact that the recordings do not just affect the privacy of people with PMID but also of the staff who care for them and visitors who may be in the vicinity. These particular points are essential in determining prior to research commencing and are known to lead to refusal to participate in such research by provider organizations.
CONCLUSIONS
AI research with the direct involvement of people with PIMD faces substantial impediments. However, people with PIMD should not be denied the cultural added value inherent in AI, which has the potential to open up new opportunities for participation and can improve quality of life by unlocking potential for meaningful communication of the individual and his/her environment. The value of new research approaches for people with PMID is central to enhancing their life opportunities. Firstly, people with PIMD are some of the most vulnerable and underserved population groups and that increases the researcher’s responsibility to do no harm, to do good and apply justice which in this case means using all tools including AI in a safe way to create novel ways in which the PMID community can be included and understood And secondly, it seems to be unavoidable that increasing digitalization will ultimately lead to ubiquitous and invisible AI applications determining our everyday lives, including those of people with PIMD and their caregivers. Early studies open-up the opportunity to identify critical aspects of such innovation and to spot future problems that require solutions. Examining the limits of AI but also harnessing its power can open up new ways in which we can improve and maintain the physical and mental wellbeing of people with PMID.
Additional information
Funding
References
- de Haas, C., Grace, J., Hope, J., & Nind, M. (2022). Doing research inclusively: Understanding what it means to do research with and alongside people with profound intellectual disabilities. Social Sciences, 11(4), 159. https://doi.org/10.3390/socsci11040159
- Engelhardt, M., Kosiedowski, M., Krämer, T., & Tyrakowska, J. (2019). Should artificial intelligence be used to empower people with profound intellectual disabilities? Proceedings of the STS Conference Graz 2019, 84–104. https://doi.org/10.3217/978-3-85125-668-0-06
- Hammann, T. (2022). Forschungsethische Diskussion im Kontext komplexer Behinderung. Zeitschrift für Heilpädagogik, 73(8), 382–393.
- Hammann, T., Schwartze, M. M., Zentel, P., Schlomann, A., Even, C., Wahl, H. ‑., & Rietz, C. (2022). The challenge of emotions — An experimental approach to assess the emotional competence of people with intellectual disabilities. Disabilities, 2(4), 611–625. https://doi.org/10.3390/disabilities2040044
- Hammann, T., Valič, J., Slapničar, G., & Luštrek, M. (2022). The missing piece. Physiological data as a factor for identifying emotions of people with profound intellectual and multiple disabilities. International Journal of Developmental Disabilities, 1–17. https://doi.org/10.1080/20473869.2022.2154928
- Heitplatz, V. N., Bühler, C., & Hastall, M. R. (2020). I can´t do it, they say! Perceived stigmatization experiences of people with intellectual disabilities when using and accessing the Internet. In: M. Antona & C. Stephanidis (Eds.), Proceedings of the 22nd International Conference on Universal Access in Human Computer Interaction (HCI 2020), Applications and Practice (pp. 390–408). Basel: Springer. https://doi.org/10.1007/978-3-030-49108-6_28
- Kleppmann, M. (2017). Designing data-intensive applications: The big ideas behind reliable, scalable, and maintainable systems. O’Reilly.
- Krämer, T., & Zentel, P. (2020). Expression of emotions of people with profound intellectual and multiple disabilities. A single-case design including physiological data. Psychoeducational Assessment, Intervention and Rehabilitation, 2(1), 15–29. https://doi.org/10.30436/PAIR20-02
- Maes, B., Nijs, S., Vandesande, S., Van Keer, I., Arthur-Kelly, M., Dind, J., Goldbart, J., Petitpierre, G., & van der Putten, A. A. J. (2021). Looking back, looking forward: Methodological challenges and future directions in research on persons with profound intellectual and multiple disabilities. Journal of Applied Research in Intellectual Disabilities, 34(1), 250–262. https://doi.org/10.1111/jar.12803
- Matérne, M., & Holmefur, M. (2022). Residential care staff are the key to quality of health care for adults with profound intellectual and multiple disabilities in Sweden. BMC Health Service Resources, 22(1), 228. https://doi.org/10.1186/s12913-022-07641-y
- Munde, V., & Zentel, P. (2020). Evidence-based practices for teaching learners with multiple disabilities. In Oxford research encyclopedia of education. Oxford University Press.
- Schlomann, A., Wahl, H. ‑., Zentel, P., Heyl, V., Knapp, L., Opfermann, C., Krämer, T., & Rietz, C. (2021). Potential and pitfalls of digital voice assistants in older adults with and without intellectual disabilities: Relevance of participatory design elements and ecologically valid field studies. Frontiers in Psychology, 12, 1–5. https://doi.org/10.3389/fpsyg.2021.684012
- Talman, L., Wilder, J., Stier, J., & Gustafsson, C. (2019). Staff members and managers’ views of the conditions for the participation of adults with profound intellectual and multiple disabilities. Journal of Applied Research in Intellectual Disabilities, 32(1), 143–151. https://doi.org/10.1111/jar.12516
- Zentel, P., Sansour, T., Engelhardt, M., Krämer, T., & Marzini, M. (2019). Mensch und/oder Maschine? Der Einsatz von Künstlicher Intelligenz in der Arbeit mit Menschen mit schwerer und mehrfacher Behinderung. Schweizerische Zeitschrift für Heilpädagogik, 25(11–12), 35–42.