This study presents findings from qualitative data collected within the European Union (EU)–funded project: “AI-PROGNOSIS: Artificial intelligence–based Parkinson’s disease risk assessment and prognosis” (Nr.101080581). To provide a perspective of people with PD throughout the project, an international patient panel was assembled in late 2023, which served as the participant group. The panel was recruited using a combination of convenience, purposive, and snowball sampling to reach multiple English-speaking people with PD from each of the 6 countries with an appropriate background. People with PD were selected based on their high PD knowledge levels, high levels of engagement in PD communities, and interest in technology, while at the same time maintaining diversity in age, gender, country, time since diagnosis, and educational or professional background. A few participants had prior (professional) knowledge of AI. An hour-long information meeting was held online for those interested prior to officially inviting participants, and those who missed were sent the recording.

We adopted participatory and co-design methodologies, which are important in the development of digital health tools, ensuring that the technology is useful and user-centered [13,18,19]. A researcher with PD was actively involved in all phases of the study: planning and design, data collection and analysis, and writing.

The 3 models under development within AI-PROGNOSIS, which were discussed with the panel, were [1] a risk score, predicting one’s risk of developing PD among individuals clinically identified as at risk [2]; a prognosis, predicting the progression of people with PD; and [3] medication response, predicting the response of people with PD to various medications and timings or doses.

The participants (patient panel) consisted of 14 people with PD (5 women and 9 men) from 6 countries: France (n=2), Germany (n=3), Portugal (n=3), Spain (n=1), Sweden (n=2), and the United Kingdom (n=3). The mean age at the time of data collection was 53.5 (SD 11.4; range 40‐75) years. The mean age at the time of PD diagnosis was 44.5 (SD 11.8; range 32‐64) years, and the mean number of years living with PD was 9 (SD 5; range 1‐17). All were of Caucasian ethnicity.

Semistructured interviews (January-March 2024) at an average of 36 minutes (range 13‐57 min) were conducted by JLL with 13 out of 14 panel members. One panel member chose not to interview but participated in an FG. The interview guide (Multimedia Appendix 1) covered four topics: (1) personal and PD background; (2) prior research participation (if any); (3) expectations, concerns, and hopes regarding this project; and (4) views on AI in health care or research in general and in AI-PROGNOSIS. No introduction to AI was provided prior to the interviews, in order to capture initial thoughts and expectations, as these were the first interviews of the project.

An introduction to the project’s AI tools and key development terms (eg, high-level features, user-needs, explainable insights) was given before each of the 2 FGs, however, to facilitate deeper and well-informed discussion (Multimedia Appendix 2). The FGs were attended by 9 or 14 panel members and were facilitated by SR and AC in February 2024. The FGs were held on 2 occasions to increase participation while keeping the size small enough for meaningful conversation, and participants signed up based on their schedule. Only the 2 most popular times were selected, meaning not all participated. Additionally, the 2 French participants joined the project after the FGs due to English language recruitment challenges. The FG objective was to elicit user needs for the project’s development and was held ahead of prototypes, so participants spoke hypothetically. For the sake of the discussion, participants were asked to respond with the assumption that the AI models would work perfectly. The content was similar between the two (the guide is provided in Multimedia Appendix 3), although the facilitators gave the second FG more time on questions unanswered by the first to ensure data on all topics of interest. FG 1 (1 hour 48 min) consisted of 2 female and 2 male panel members from 3 different countries, and FG 2 (1 hour 43 min) consisted of different participants: 3 female and 2 male panel members from 4 countries. The participants were divided with availability and diversity in mind. Each FG consisted of a primary and secondary facilitator, as well as 1 additional project member, who facilitated note-taking. All data were collected over Zoom in English, with occasional clarifications as necessary in participants’ native language to maximize participation and understanding.

The interviews and FGs were analyzed abductively as 1 dataset, using a simplified version of Thompson’s 8-step abductive thematic analysis (Figure 2) [20]. This method of abductive reasoning, shown in Figure 3, involves iteratively moving between empirical data and theoretical framework to refine interpretations [21]. JLL familiarized herself with the data by transcribing the interviews and FGs and reading the transcripts. She then conducted inductive coding in Taguette, using “considerations for AI in managing PD” as an initial analytic lens. Codes were low-level descriptors of salient statements emerging from the text. The codes were then transferred into Miro, a collaborative online “whiteboard,” where codes were inductively grouped into preliminary themes and discussed between authors. An overarching ethical dimension emerged, prompting the abductive step of refining the research focus toward ethical considerations and the application of biomedical ethical principles [15] to deductively categorize the codes or preliminary themes (Multimedia Appendix 4). One author (CB), an expert in biomedical ethics, provided critical input to ensure alignment between themes and principles. JLL and SR finalized the analysis in a reflexive process, revisiting the transcripts, discussing, and revising accordingly. The data were not constrained by the framework, and a section highlighting outliers is also included. Due to the elements of trust in both truthfulness and confidentiality, and the smaller themes within those two principles, we combined all 3 for the presentation of our findings.

Participants were informed about the study aims and procedures, including data collection and processing, potential risks related to participation, and their rights under the General Data Protection Regulation (GDPR). The data were de‑identified prior to analysis and reporting. Direct identifiers (e.g., names) and potentially identifying information such as age were removed. When presenting quotations, only broad demographic characteristics were retained to minimize the risk of re‑identification. Access to the raw data was restricted to the research team, and all data were stored on secure, password‑protected systems in accordance with GDPR requirements. All participants provided informed consent prior to participation. Ethical approval was obtained from the Swedish Ethical Review Authority (reference number: 2023‑05949‑01). Participants received a modest gift card as compensation for their time.

Participants identified multiple ways in which AI could offer benefits to their health, particularly by improving data collection, optimizing clinical care, and enabling early detection and intervention.

While participants felt they could potentially benefit from predictive AI and personalized monitoring, they also expressed concerns that not everyone has equal access to the resources necessary to implement AI-generated recommendations.

Tools such as predictive AI, it was perceived, could further increase inequalities with regard to autonomy and beneficence among different individuals in PD care. Another question of fairness was related to the resources that AI requires and whether the added benefit of such a toolkit as this would outweigh the cost.

Privacy and confidentiality were key concerns among participants, who emphasized the need for control over their own data, including the ability to choose who can access their health information—whether care partners, HCPs, or third parties: “I live in Germany where people are, I would go so far as to say, obsessed about data protection” (female, person with PD 09 FG1). Participants stressed the need for transparency in data collection, storage, and sharing practices, ensuring that personal health data remain protected and used only with explicit consent.

Building trust in AI-driven PD tools requires both transparency and user control. Participants were hesitant about whether neurologists could accurately interpret their data without their input. They emphasized that people with PD should have access to their data first, before any report is sent to HCPs, ensuring transparency and patient involvement in interpretation. Participants questioned how accurate or “truthful” AI predictions could be, given the extreme complexity and variability of PD, although they were still interested in the potential insights AI could provide. “I would like to know as much [of the AI-predictions] as possible, but I also want to know the accuracy of those predictions” (male, person with PD 03, FG1). Another participant felt confident enough in her disease stability to be interested in an AI-generated prognosis:

To increase trust and perceived value, participants wanted clear explanations of how the AI models work, their limitations, and their accuracy rates so as to facilitate interpretation rather than misleading.

Beyond bioethical concerns, participants highlighted practical challenges, such as cultural differences across the EU, which could complicate AI implementation:

Participants expressed both potential benefits and ethical concerns of implementing AI in PD management. While optimistic that AI could enhance autonomy and beneficence by capturing one’s unique health patterns, motivating behavior change, and enabling shared decision-making, concerns arose over patient involvement, privacy, and the psychological impact of predictions. Risk prediction, PD prognosis, and medication response prediction were perceived differently in terms of value and ethical considerations, with risk prediction seen as the most ethically controversial, given that people with PD often have come to terms with their diagnosis in a way that those at risk would not have. Participants cautioned that AI could widen health care disparities and emphasized the need for transparency, user control, and the involvement of people with PD in development to ensure ethical implementation.

Consistent with others’ findings, participants felt that longitudinal monitoring and predictive AI have the potential to improve understanding and communication for both people with PD and HCPs, ultimately leading to more personalized and appropriate (self-)care decisions [14]. However, participants worried about inconsistencies in the interpretation of data between people with PD and HCPs, with one participant suggesting to see and approve AI predictions before sharing them with their HCP.

Participants, in alignment with previous research [22], were interested in AI solutions that provide actionable insights, as this increases autonomy over one’s health. A review of technologies within PD cites multiple examples of how providing feedback from the digital monitoring elicited positive outcomes for the people with PD, compared to when the information was exclusively for HCPs [12]. This becomes more complex, however, with the addition of predictions. Participants were interested in AI-generated disease pathways and alternatives based on interventions, although not without skepticism. As others discuss, this may provide more false hope than benefit, given that there are currently no disease-modifying treatments, and even if such therapies become available, it may take many years to be individually tailored, given the heterogeneity of PD [22]. This leads to concerns such as people with PD feeling like a failure after complying with an intervention without any noticeable difference. In addition to symptom- and disease-altering recommendations, a motivation for predictive AI noted in our study and by others was that having a better idea of one’s future could allow them to prepare mentally and logistically (eg, retirement) [22]. Given the gravity of personal predictions and the uncertainty of model accuracy, transparency in terms of what defines a given prediction or recommended intervention and the accuracy of the models is crucial, to allow people with PD to interpret the information—with the help of their HCP. AI medication response prediction was viewed as the least ethically complicated and could allow participants together with their HCPs to optimize not only which medication to prescribe but also how to maximize its effects by identifying patterns.

Schaeffer et al [9] surveyed experts on whether to disclose PD risk and how (ie, as a percentage/value or using general terms); however, research on this topic from the patient perspective is limited. Our study indicated that graphs and trends were preferred over raw numbers, which were not seen as interpretable or actionable. Consistent with our findings, the experts surveyed cautioned that risk predictions could likely result in psychological distress, especially when there are questions about the accuracy of the prediction [9]. In alignment with previous findings, our participants stressed the importance of receiving AI insights from a trained HCP who can help interpret and guide them [8,22].

Participants alluded to the subject of overdiagnosis as potential outcomes of AI-assisted PD risk screening. Defined as “making people patients unnecessarily,” overdiagnosis involves a diagnosis that meets all diagnostic criteria (not misdiagnosis) but causes the patient more harm than benefit [23]. A type of overdiagnosis is overdetection, or the identification of abnormalities that resolve themselves, never progress into something harmful, or else progress slowly enough that they never cause harm during a person’s remaining lifetime [23]. This question regarding the value of early detection, in this case of prodromal PD, when symptoms may be minimal or unnoticed and the prognosis of the complex disease is uncertain, shows the relevance of the “right not to know.”

While participants stressed their “right to know” any analyses of their health data, they also highlighted the potential psychological damage and other unintentional consequences that could arise, and hence the “right not to know.” This ethical dilemma is discussed by Schaeffer et al [9] in the context of disclosing PD risk to patients with RBD, in which they state that “there is not only a right to know, there is also the ‘right not to know.’” Davies [24] outlines 2 main arguments for the “right not to know” as to respect autonomy and avoid harm; when a diagnosis can lead to distress and social stigma, and when there is no effective cure, the diagnosis may not be worth it. However, in a response to Schaeffer et al, Karagianis [25] points out that in order to consent to not knowing, one must first understand that risk (in this case, a patient with RBD knowing their heightened risk for neurodegenerative disease) and consent to being monitored but without knowing any resulting risk scores. Davies [24] highlights an additional complexity: because PD can be genetically inherited, learning one’s own genetic information may inadvertently disclose information about family members who have not consented to receive it. Whether to disclose such information can become a moral burden for the patient receiving the information. An interesting finding was that a participant wondered if AI could be used to predict how someone might react to an AI health prediction, essentially using the problem as the solution.

There is no question that AI is an increasingly important topic in health care. The 2024 Revision to the Declaration of Helsinki for ethical medical research [26] is a reflection of this trend and provides concrete focus areas that directly shape future research priorities for AI in medicine, including improved AI literacy and clarity regarding the risks of current and future AI applications. These international principles have direct implications for the use of AI in PD. For instance, the declaration’s emphasis on protecting participant data and establishing ethical data governance speaks directly to the primary concerns identified in AI health monitoring for PD, such as privacy, data ownership, and the potential for stigmatization or discrimination based on AI-generated health predictions. Future research must therefore prioritize creating frameworks that address who controls the vast amounts of personal health data collected by monitoring systems and how the data can be used without harming the individual’s social and economic well-being. Furthermore, the revised declaration highlights the need for transparency and accountability, which aligns with the FUTURE-AI framework’s principles of traceability and explainability. This is critical in PD, where disclosing information about future disease risk is already fraught with ethical complexity due to prognostic uncertainty and the lack of disease-modifying therapies. The FUTURE-AI Framework, developed by interdisciplinary experts from 50 countries to ensure trustworthy and ethical implementation of AI in health care, warns of many of the same ethical concerns, such as privacy, transparency, potential patient harm, and driving inequities. The framework provides 6 guiding principles: fairness, universality, traceability, usability, robustness, and explainability to guide best practice for the development and implementation of AI in health care [27]. This is especially important, as the increased adoption of AI could likely introduce new challenges and responsibilities for HCPs, in addition to avoiding patient harm.

Additionally, it is imperative that future research explores how these ethical considerations are influenced by diverse global contexts, including different health care systems, cultural attitudes toward technology, and varying levels of health care access. The ethical frameworks guiding AI development, such as the Declaration and the FUTURE-AI framework, are intended to be international. However, their practical application will inevitably vary.

Furthermore, AI-assisted decision-making raises questions about accountability; while human error is an expected risk in clinical practice, it is unclear if AI-driven errors are expected or acceptable. For example, it remains unclear to what extent HCPs should be held responsible for mistakes arising from AI-driven recommendations. Despite these challenges, AI could offer valuable insights into how people with PD are doing between the infrequent and brief clinical visits and foster self-tracking and self-care for a more collaborative approach to care.

There are strengths and limitations to consider when interpreting the results. The panel’s reflections were based on hypotheticals, and ahead of a prototype, making them more general. People with PD discussed the risk prediction tools in the context of their diagnosis journey, which may impact their perspective, but this also mitigates the ethical concerns of asking and worrying people actually at risk. Another challenging but interesting aspect of this study is that it focused on estimating PD risk and prognosis as well as medication response for people with PD. This made it at times difficult to separate which tools and predictions participants referred to.

We did not explicitly ask about ethics in either form of data collection, so it is possible that the participants could have additional thoughts on ethical considerations. The topic, instead, emerged naturally among the participants, which is also a strength, as it was unsolicited, highlighting the importance of ethics in such a context. Another consideration is that the data included questions centered around the AI-PROGNOSIS project specifically, as well as about AI in health care and research in general. This was considered in the analysis process, and a project example was in many cases useful to structure and focus the conversation. In some cases, participants spoke of all AI as one, making it unclear as to whether they were referencing predictive or perhaps generative AI, which may reflect the “black box” behind AI and indicate a need for user education when implementing such tools.

It is a strength that our participants represent multiple European countries; however, there are limitations to consider regarding the transferability of our findings. Paccoud et al [28] found that user preferences varied across EU countries and clinical and sociodemographic factors, though our participants did not mention many cultural factors. Our participants, being on an expert panel, likely differ from the average people with PD in many ways, given that they are well educated, quite involved in their own PD management, as well as PD awareness or advocacy. Many were also quite tech-savvy and had self-tracked in some manner prior, suggesting that this group might be more open toward AI but might also have more specific and well-thought-out feedback. Additionally, several of our participants have young-onset PD, and the mean age at the time of this study was 53.5 (SD 11.4), younger than the average people with PD. Older or cognitively impaired people with PD may encounter more barriers to navigating and interpreting the AI feedback and may not engage with the technology as much as our participants wish to. A recent survey on willingness to use digital medical devices among people with PD in Europe found that willingness decreased with age [28]. Another study found a strong significant correlation between cognitive ability and perceived ability to use everyday technology among people with PD [29]. At the same time, it may be considered a strength that our study provides insights into the ethical considerations of disclosing risk and progression to middle-aged individuals without cognitive impairment, some of whom are still working, for whom a prediction of disease progression may represent a more pronounced shock or harsher news.

Finally, given that our participants were Caucasian and from high-income countries, their experiences and perspectives may be very different from those from minority groups and of low- and middle-income countries, as one of our participants pointed out.

Our findings highlight the need to balance hope and harm in the application of AI in PD management. We outline potential benefits for people with and without PD, HCPs, and research, as well as important concerns from the perspective of people with PD. Involving people with PD in development going forward is crucial for aligning technological advancements with patient needs and ensuring the successful and ethical adoption of AI in PD care.