The prevalence of multiple long-term conditions (MLTCs), defined as the coexistence of 2 or more chronic conditions in an individual, is rising globally due to aging populations and lifestyle factors [1,2]. MLTCs are associated with decreased quality of life, functional decline, and increased health care use, particularly as the number of coexisting conditions rises [1]. Managing MLTCs presents several challenges; for example, primary care practitioners must allocate limited resources to patients with complex needs, and prescribing guidelines focus on single conditions, often resulting in polypharmacy [3]. Polypharmacy, or the use of 5 or more medications daily [1], can lead to adverse events, inappropriate prescribing, and difficulties with adherence [4,5].

Shared decision-making (SDM) offers a way to address the complexity of MLTC management. SDM is a collaborative approach in which patients and clinicians make health care decisions together, incorporating patient preferences and the best available evidence [6,7]. Effective SDM relies on strong clinical communication skills, including rapport-building and structured consultations [6]. With MLTC, SDM is more complex, as clinicians must help patients prioritize concerns and make decisions within limited time and evidence [7].

Understanding how patients with MLTCs perceive and engage in SDM with a general practitioner (GP) is essential to improving care quality and outcomes. However, SDM is influenced by factors such as patient readiness, power dynamics, and limited organizational support [8]. A UK study found that both GPs and older patients with MLTCs expressed concerns about medicolegal vulnerability in SDM, arising from the risk of legal consequences if a negative outcome followed a shared decision [9].

Artificial intelligence (AI) offers the potential to support SDM and MLTC management. AI can perform tasks traditionally undertaken by health care professionals, including providing personalized information, decision aids, and predictive analytics to guide collaborative decision-making [10-12]. In GP consultations, AI tools could support medication decisions for patients with MLTCs by analyzing large datasets to generate tailored recommendations, optimizing treatment choices and minimizing the risks of polypharmacy. AI can consider individual patient characteristics and histories in ways beyond human capacity [13]. This is particularly valuable in the management of MLTCs, where adapting treatment plans to each patient becomes increasingly complex. As a result, AI could potentially be used to overcome the barriers of SDM for people with MLTCs by facilitating personalized discussions based on comprehensive data analysis and enhancing patient engagement through tailored information and decision aids.

For both patients and GPs, managing MLTCs can be complex, and although SDM is recommended, it is difficult to implement in practice due to polypharmacy, time pressures, and a lack of reliable evidence relevant to multimorbidity. AI could support personalized decision-making, but little is known about how patients perceive its use in GP consultations or how such tools should be designed and implemented to support patient-centered care. The Optimising Therapies, Disease Trajectories, and AI-Assisted Clinical Management for Patients Living with Complex Multimorbidity (OPTIMAL) project used 2 rounds of qualitative interviews to explore how AI could be used to analyze complex multimorbidity and prescribing data to support more personalized, safer decision-making for patients with polypharmacy [14]. This paper specifically explores the findings from a second round of interviews with patients who previously participated in interviews examining their perspectives on the potential benefits and risks associated with the use of AI in GP consultations [15]. In this study, the interviews focused on participants’ views relating to a prototype AI tool designed to support SDM. Participants reflected on how the tool could be used in practice and how it would support their own experiences of SDM and experiences of living with MLTCs. The following are the aims of the second round of interviews:

The data were collected through semistructured interviews conducted via telephone or video link (Zoom or Microsoft Teams) between January and February 2024. Patients were recruited from GP practices across the West Midlands, UK, with a total of 18 interviews conducted by 1 researcher (SF), who was a research fellow at the University of Birmingham at the time.

All participants had previously taken part in the first round of interviews exploring their perspectives on AI in health care consultations [15]. For the second round of interviews, a simulated patient vignette and a predictive algorithm prototype (AI tool) were developed by a multidisciplinary team. Vignettes may be used in social research to explore actions in context, clarify judgments, or provide a less personal way of discussing sensitive topics [16]. The simulated patient vignette was used to help participants explore and reflect on how the AI tool might function within a realistic GP consultation and helped convey complex concepts relating to AI-supported prescribing without requiring extensive explanation during the interview. The vignette was also developed and piloted with input from the participant advisory group (PAG).

The simulated patient vignette used in the study described a patient, “Janet,” with MLTCs to facilitate discussion (Multimedia Appendix 1). Alongside the vignette, the AI tool output provided participants with a visual representation of how the tool might function, using Janet’s case as an example (Multimedia Appendix 2). The AI tool was designed to support SDM during consultations between patients and GPs within a primary care setting by predicting risks for patients with MLTCs, considering their coexisting conditions and key metrics such as blood pressure and BMI. The tool could produce outputs comparing medication options by presenting the expected impact on the condition of interest alongside the predicted effects on other health conditions and key health metrics to support SDM.

The case study used data from the AI tool to support a decision about the most appropriate medication for Janet’s depression, considering her current health, medications, and health conditions. The AI tool presented her health metrics, conditions, and medications, alongside visualizations of the potential impact and outcomes of different antidepressants.

The interviews were guided by a topic guide, focusing on participants’ experiences with SDM and their feedback on the AI tool’s usefulness in supporting health care decisions (Multimedia Appendix 3). Participants were asked about their views on the AI tool, personal experiences of SDM in GP consultations, and how the tool might support SDM.

For the first round of interviews, the study aimed to recruit up to 30 individuals with varied ages, genders, ethnicities, and numbers of long-term conditions. Recruitment continued until data saturation was reached, at which point additional data no longer generated new insights or themes [17]. Recruitment for the first round of interviews occurred between July and November 2023.

Eligible participants were recruited via the Clinical Practice Research Datalink (CPRD) Aurum database, an anonymized UK primary care dataset representative of the national population [18]. Using CPRD data on diagnoses, tests, prescriptions, and demographics, potential participants were identified and invited through the CPRD Patient Referral Service [19]. GP practices in the West Midlands were targeted to achieve a diverse sample, focusing on more deprived areas with higher proportions of patients from minoritized ethnic backgrounds and with MLTCs. CPRD contacted participating GP practices, which reviewed lists of eligible patients and invited them to take part.

Participants were eligible if they were aged 18 years or older who have not opted out of contributing data to the CPRD database. Participants were also required to have self-reported MLTCs, defined as the coexistence of 2 or more chronic conditions. Participants could also be a carer for someone with 2 or more conditions; however, no carers were recruited for this study. Exclusion criteria included having a terminal illness or lacking the capacity to provide informed consent. Those interested in taking part contacted the research team directly to ask questions, provide consent, and arrange an interview.

Additional recruitment was conducted via community-based methods, including flyers distributed through voluntary and community organizations such as patient support groups. Interested individuals were sent a participant information sheet and offered the opportunity to discuss the study with a member of the research team before deciding whether to participate.

Of those who participated in the first round of interviews, individuals were asked whether they consented to be contacted for a second interview. Twenty-eight participants gave consent and were subsequently contacted by email. The email invitation outlined the aims of the second interviews and what participation would involve. Of those who initially agreed to be contacted, 10 declined or were unavailable. In total, 18 participants completed a second interview.

For the second round of interviews, participants were sent the simulated patient vignette, “Janet,” and a PDF of the AI tool dashboard in advance, either by email or by post. They were asked to review these documents prior to the interview. At the start of each session, participants were asked if they had reviewed the materials and whether they had any questions. The interviewer (SF), who also conducted the first round of interviews, clarified any aspects of the vignette or AI tool for participants who found them challenging to understand. Interviews lasted up to 60 minutes. While the sample size for the follow-up interviews was guided by the number of participants who agreed to take part in the second interview, analysis indicated that no substantially new themes were emerging across interviews, suggesting sufficient thematic depth and consistency within the dataset [20].

People with lived experience of MLTCs were involved throughout the OPTIMAL study. A PAG of 9 members, including individuals with direct or carer experience of 4 or more long-term conditions, was established at the study’s inception. Recruitment aimed to ensure diversity in age, sex, ethnicity, and condition type. The PAG contributed to developing the topic guide, reviewing participant-facing documents, providing feedback on emerging analyses, and reviewing findings.

Interviews were audio-recorded and transcribed verbatim by a transcription company contracted by the University of Birmingham. Personal identifying information was removed from the transcripts prior to analysis. The transcripts were analyzed using the thematic analysis methods outlined by Braun and Clarke [21]. Following thorough familiarization with the transcripts, the researchers assigned codes to the data using an inductive approach. The codes were then organized by grouping related codes into overarching categories relating to SDM in GP consultations and the AI tool. Initial line-by-line coding generated a large number of descriptive codes capturing participants’ views on communication, trust, time pressures, and perceptions of AI. Codes were iteratively compared and refined, and related codes were grouped into higher-order categories. These categories were then developed into themes that captured patterns across the dataset.

The coding and analysis of the data were undertaken by 2 researchers (SF and CS). As the codes and themes were identified, these were discussed with the PAG and the multidisciplinary research team to identify further areas for exploration. Microsoft Word and Excel (Microsoft 365) were used to read the transcripts and support coding and theme development [22].

The COREQ (Consolidated Criteria for Reporting Qualitative Research) checklist was completed to support transparent reporting. The checklist is provided in Checklist 1, and the corresponding reflexivity statement is provided in Multimedia Appendix 4.

Electronically completed written consent or audio informed consent was collected from participants. Ethical approval for the study was obtained from the National Health Service Research Ethics Committees (reference: 22/SC/0210). Ethical approval to use the anonymized data from the CPRD dataset to select eligible participants was obtained from the CPRD Expert Review Committee and the Central Advisory Committee (reference: 21_000683). All interviews were transcribed, anonymized, and stored securely in accordance with UK data protection regulations and the University of Birmingham security policies. Participants received a £15 (equivalent to US $20.12 as of June 17, 2026) voucher as a thank you for their time.

A total of 18 interviews were conducted. Figure 1 presents a flow diagram of recruitment for the interviews. Among the participants who completed the second interview, 9 were female and 9 were male. The majority were White British (n=14, 77.8%), with 3 identifying as Asian or British Asian and 1 as mixed ethnicity. Age distribution was as follows: 6 participants aged 70 to 79 years, 3 aged older than 80 years, 5 aged 60 to 69 years, 3 aged 50 to 59 years, and 1 aged 40 to 49 years (Table 1). In the quotations presented, each participant is identified by their ID number, sex (female [F] or male [M]), and age.

Two overarching categories were used to organize the themes and subthemes: SDM in GP consultations and the AI tool for SDM in GP consultations. Within the SDM category, 2 themes were identified: the role of communication and collaboration and system-level barriers associated with SDM. Within the AI tool category, themes included practical considerations for AI tool design and implementation, implications for clinical practice and decision-making, and perceived risks and limitations of the AI tool.

The interviews identified several circumstances within GP consultations that support or hinder SDM for patients with MLTCs.

For the second part of the interview, participants were given the vignette and the AI tool output. They reflected on its usability, suggested improvements, and considered its role in SDM.

This study explored how an AI tool could be used in GP consultations to support SDM for people living with MLTCs. Participants engaged with the AI concept while reflecting on the complexities of managing their MLTCs, unfamiliar terminology, and receiving conflicting health information. Although many participants valued involvement in decisions about their care, experiences were shaped by continuity with the same GP, consultation time constraints, and communication quality.

These findings echo broader evidence that SDM depends not only on patient preference but also on systemic conditions that enable meaningful dialog [6,23]. While patients are generally supportive of SDM, their engagement is influenced by factors such as knowledge, continuity of care, and time [8]. This highlights the systemic, cultural, and interpersonal barriers that prevent patients’ full participation in the decision-making process [24].

Although grounded in their own experiences of managing MLTCs, participants often reflected on how the AI tool might affect others, particularly those with lower health literacy, language barriers, or accessibility needs. Participants questioned whether AI-driven recommendations might alleviate or exacerbate these existing challenges. Concerns were also raised around whether AI-driven recommendations would make health care less personalized, and some questioned the trustworthiness of the data used in the AI tool. Participants were concerned that AI tools could reconfigure the dynamics of the decision-making process away from their own priorities and toward external pressures, such as cost-saving measures.

This study is 1 of the first to explore how an AI tool can support SDM in GP consultations for people with MLTCs in the United Kingdom. There is a particular lack of evidence on AI-supported SDM within the structure of UK primary care, where consultation length, continuity of care, and challenges associated with MLTCs differ from other health systems. This study therefore offers novel, context-specific evidence. Understanding patients’ perspectives can inform the development of AI tools that are acceptable and appropriate, particularly for those with high treatment burdens [7]. Such patients typically attend more frequent appointments and are prescribed multiple medications compared to those without MLTCs, which can impact adherence and engagement with care [7,25].

While SDM is well explored, the use of AI to support it in MLTCs remains underdeveloped. Current guidelines are based largely on single-disease models, as clinical trials often exclude people with multimorbidity, increasing the risk of polypharmacy and complications [7,26]. While existing studies provide general support for the use of AI in primary care settings [27,28], there is limited evidence on how AI can be applied specifically to facilitate SDM within GP consultations for patients with MLTCs.

In this study, the AI tool was received positively by participants, although some worried it could complicate consultations, which already have limited time. In addition, some participants feared the tool might be applied in a “one-size-fits-all” manner, removing personalization from consultations. For participants, this depersonalization meant that the AI tool would not be able to consider the patient’s broader social, physical, or psychological context in the way that doctors can. These concerns reflect worries about how AI may standardize care for individuals, without considering how MLTCs and wider individual circumstances shape care needs.

These concerns are echoed in a study of primary care physicians in Germany, which found that AI frequently gave irrelevant recommendations and generated patient alerts that were not clinically appropriate [29]. This reinforces the importance of clinical oversight of AI tools. Participants in this study felt that AI could not completely replace GP decision-making and emphasized the importance of doctors continuing to use their own knowledge, experience, and judgment. Participants also noted that AI outputs could influence clinician decision-making, potentially shifting power dynamics in consultations if doctors feel pressured to follow algorithmic guidance, highlighting the need to balance AI support with clinician and patient autonomy.

These findings align with research from other studies, emphasizing the importance of physician oversight in AI-supported care [11]. Similarly, a recent study with health care professionals found that while AI tools hold promise for the management of clinical complexity, they must also provide clear rationale for recommendations, preserve clinician autonomy, and integrate within existing workflows [30]. Overall, AI offers promising opportunities to support SDM, but challenges remain regarding usability, patient understanding, and preserving human elements in care. Balancing technological innovation with patient-centered values is important to ensure tools enhance, rather than replace, clinical relationships.

This qualitative study has some strengths and limitations. A comprehensive recruitment approach was adopted, incorporating invitations via the CPRD database and engagement with community, patient, and third-sector organizations. This approach supported the inclusion of participants with diverse experiences and characteristics, enhancing the transferability of the findings [31]. Guided by a multidisciplinary research team, the study applied robust methods for data collection and analysis. Semistructured interviews enabled in-depth exploration of patients’ experiences with MLTCs, SDM, and AI, providing insight into social, cultural, and personal factors shaping views. Furthermore, the use of the COREQ checklist provides a clear and transparent account of its analytic approach, enabling readers to understand how interpretations were developed.

The lower-than-expected response rate may have been influenced by the time commitment required for participation, involving two 60-minute interviews, highlighting the need to consider participant burden in future studies. Additionally, as noted elsewhere, individuals without direct experience of AI may find it difficult to conceptualize or engage with such topics, particularly when examples of AI applications are not clearly defined [32]. Participants who decided to take part in the study and attended both the first and second interviews may have had a particular interest in the topic or been disproportionately more digitally literate or more physically well, meaning the sample might not fully represent the broader population of patients with MLTCs.

In addition, while the sample included some diversity, individuals from ethnic minority backgrounds were underrepresented. As a result, certain perspectives may not be fully captured in this analysis, which is particularly relevant given the potential for bias in AI systems [33].

In this study, 1 of the central limitations is related to the fact that the study used a static AI prototype, presented as a PDF, which served only as a visual guide and was not interactive. This means responses reflected expectations rather than real-world use. The early prototype offers some benefits, including obtaining direct feedback about preferences and requirements at an early stage, supporting responsible AI development. However, the findings are speculative rather than based on genuine user experience. Therefore, the design-related conclusions drawn from this study need to be tested, and further research is required to determine whether a real-world AI tool for SDM and MLTC would produce the same findings as those presented here.

Despite its limitations, this study offers recommendations for the design of AI tools for use during GP consultations. These recommendations are separated into 3 design areas. The first relates to AI tool design. AI tools that produce outputs for patients need to present a mixture of text and visual information in a clear and simple manner. The design also needs to consider accessibility issues that can affect readability, such as color choices for color blindness and how outputs can be made accessible to people with visual impairments.

The second area relates to how the AI tool is communicated. The terminology used needs to be in lay language, and health care professionals using the tool to support SDM need to be able to explain it adequately to patients. Equality, diversity, and inclusion should also be considered in the design, including involving patient and public groups and providing scope to produce outputs in languages other than English. Future research should include diverse populations, including minority ethnic groups, people with language barriers, and those with sensory or learning disabilities, to ensure AI tools are inclusive and do not risk exacerbating existing inequalities [34].

The third area relates to how AI tools can be used in practice. Doctors need adequate time to explain the tool, and tools should be designed to enhance SDM rather than replace clinical decision-making. Patients in this study questioned data sources and transparency, so it is important that data provenance is communicated clearly. Furthermore, patients require support to understand AI outputs, and GPs need sufficient time and training to integrate these tools without compromising personalized care.

This study demonstrates how AI could be used to support SDM in primary care consultations for patients living with multimorbidity. Participants largely supported the idea of using an AI tool to improve communication and collaboration with their GPs, recognizing its potential to present complex information in a more accessible way and to prompt more meaningful dialogue during consultations. However, they also highlighted the importance of clear explanations and the need for the tool to complement rather than replace human clinical judgment. Future research should focus on evaluating how such tools are implemented and used in real-world primary care settings, including their impact on consultation dynamics, patient understanding, and clinician workload over time. Further work is also needed to explore longer-term patient experiences and to assess acceptability and effectiveness across more diverse patient groups to inform equitable implementation.