Patient Perspectives and Expectations on the Use of Artificial Intelligence to Guide Treatment Decision‐Making in Juvenile Idiopathic Arthritis: A Qualitative Study Authors: Mei‐Sing Ong, Laura E. Schanberg, Melanie Kohlheim, Jenny Leal, Ashley Luo, Marc D. Natter, Yukiko Kimura Published: 2026 ABSTRACT ObjectiveArtificial intelligence (AI) is rapidly transforming clinical decision‐making. However, patient perspectives on these technologies remain understudied in pediatric care. We investigated patient and parent perspectives on AI use in juvenile idiopathic arthritis (JIA) care among a highly engaged cohort.MethodsWe conducted four focus groups with patients with JIA and parents of children with JIA to explore: (1) perceived utility of AI, (2) anticipated benefits and risks of its implementation, (3) information needs for fostering understanding and trust in AI tools, and (4) preferences for model calibration, including acceptable performance thresholds and trade‐offs between sensitivity and specificity. Transcripts were analyzed using reflexive thematic analysis.ResultsParticipants included 8 patients with JIA (3 children and 5 adults) and 10 parents (n = 18 total). Participants expressed enthusiasm for AI's potential to support personalized treatment, but acceptance was universally contingent on AI serving as decision support rather than replacing physician judgment. Primary concerns were physician over‐reliance undermining clinical reasoning, privacy vulnerability for pediatric populations, and potential algorithmic bias. Participants emphasized that successful implementation must preserve the patient–provider relationship and accommodate individual circumstances, including disease severity, patient preferences, and lifestyle factors. Acceptable sensitivity–specificity trade‐offs varied widely, shaped by each participant's disease experience. Some participants emphasized that trust required disclosing AI use, training data sources, model performance, model provenance and conflicts of interest, and cohort characteristics.ConclusionPatients expect AI technologies to augment rather than replace clinical judgment, operate transparently, and accommodate individual circumstances. These priorities should guide AI implementation to ensure trust and meaningful adoption. INTRODUCTION Artificial intelligence (AI) is rapidly transforming the landscape of medical care, offering promising tools to support clinical decision‐making. In pediatric rheumatology clinics, and other care settings where treatment decisions are often complicated by disease heterogeneity and the lack of evidence‐based guidelines, 1 AI has the potential to help guide personalized treatment strategies. However, the integration of AI into clinical practice raises important ethical concerns, particularly in pediatric settings. Concerns surrounding data privacy and informed consent take on greater significance, because children lack the legal capacity to provide consent and are dependent on caregivers and clinicians to safeguard their interests and advocate on their behalf. 2 Moreover, clinical decision‐making in pediatric care typically involves not only the patient, but also parents or legal guardians, whose diverse values and preferences can introduce additional complexity. 3 SIGNIFICANCE & INNOVATIONS This qualitative study is the first to examine the perspectives of patients with juvenile idiopathic arthritis and their parents on the use of artificial intelligence (AI) technology to facilitate treatment decision‐making.Patients and parents were highly enthusiastic about AI in treatment decision‐making when used as clinical decision support (not as a physician replacement), emphasizing the need for clear information about the model (eg, training data sources, performance, and potential conflicts of interest), as well as strong personalization to individual disease severity, cultural context, life stage, and personalized risk‐benefit preferences.Patients’ primary concern was physician over‐reliance on AI diminishing clinical judgment (automation bias), along with privacy risks, algorithmic bias, and potential harm to patient–provider relationships.These findings underscore that effective AI implementation requires calibrated communication that fosters appropriate trust without overwhelming patients, safeguards to protect clinical judgment against automation bias, and a patient‐centered design approach. SIGNIFICANCE & INNOVATIONS This qualitative study is the first to examine the perspectives of patients with juvenile idiopathic arthritis and their parents on the use of artificial intelligence (AI) technology to facilitate treatment decision‐making.Patients and parents were highly enthusiastic about AI in treatment decision‐making when used as clinical decision support (not as a physician replacement), emphasizing the need for clear information about the model (eg, training data sources, performance, and potential conflicts of interest), as well as strong personalization to individual disease severity, cultural context, life stage, and personalized risk‐benefit preferences.Patients’ primary concern was physician over‐reliance on AI diminishing clinical judgment (automation bias), along with privacy risks, algorithmic bias, and potential harm to patient–provider relationships.These findings underscore that effective AI implementation requires calibrated communication that fosters appropriate trust without overwhelming patients, safeguards to protect clinical judgment against automation bias, and a patient‐centered design approach. This qualitative study is the first to examine the perspectives of patients with juvenile idiopathic arthritis and their parents on the use of artificial intelligence (AI) technology to facilitate treatment decision‐making. This qualitative study is the first to examine the perspectives of patients with juvenile idiopathic arthritis and their parents on the use of artificial intelligence (AI) technology to facilitate treatment decision‐making. Patients and parents were highly enthusiastic about AI in treatment decision‐making when used as clinical decision support (not as a physician replacement), emphasizing the need for clear information about the model (eg, training data sources, performance, and potential conflicts of interest), as well as strong personalization to individual disease severity, cultural context, life stage, and personalized risk‐benefit preferences. Patients and parents were highly enthusiastic about AI in treatment decision‐making when used as clinical decision support (not as a physician replacement), emphasizing the need for clear information about the model (eg, training data sources, performance, and potential conflicts of interest), as well as strong personalization to individual disease severity, cultural context, life stage, and personalized risk‐benefit preferences. Patients’ primary concern was physician over‐reliance on AI diminishing clinical judgment (automation bias), along with privacy risks, algorithmic bias, and potential harm to patient–provider relationships. Patients’ primary concern was physician over‐reliance on AI diminishing clinical judgment (automation bias), along with privacy risks, algorithmic bias, and potential harm to patient–provider relationships. These findings underscore that effective AI implementation requires calibrated communication that fosters appropriate trust without overwhelming patients, safeguards to protect clinical judgment against automation bias, and a patient‐centered design approach. These findings underscore that effective AI implementation requires calibrated communication that fosters appropriate trust without overwhelming patients, safeguards to protect clinical judgment against automation bias, and a patient‐centered design approach. Understanding how patients and families perceive the role of AI in guiding clinical care is critical to ensuring that these technologies are implemented in ways that align with their values and expectations. Although a growing body of research has examined patient perspectives on AI use in health care, most studies have been conducted in adult populations. Moreover, existing literature primarily focuses on patients’ general attitudes toward AI. 4 , 5 Limited attention has been given to understanding patient expectations of how AI models should perform (eg, acceptable levels of accuracy) and the information patients need to foster trust in the technologies. Addressing these gaps is critical for the effective and ethical integration of AI in clinical practice. In this qualitative study, we examined the perspectives and expectations of patients with juvenile idiopathic arthritis (JIA) and their families on the use of AI for guiding treatment decision‐making. Our goals are to examine: (1) perceived utility of AI, (2) anticipated benefits and risks of its implementation, (3) information needs for fostering understanding and trust in AI tools, and (4) preferences for model calibration, including acceptable performance thresholds and trade‐offs between sensitivity and specificity. Although this study focuses on JIA, the insights gained will inform integration of AI in diverse pediatric health care settings. PATIENTS AND METHODS Study design and subjects We conducted four focus groups between February 2024 and January 2025. Patients with JIA and their caregivers were recruited through social media and the Childhood Arthritis and Rheumatology Research Alliance network. Study inclusion criteria included being conversant in English and being ≥14 years. Participants received $50 for their participation. Recruitment continued until thematic saturation was reached. The study was approved by the Harvard Pilgrim Health Care Institute Institutional Review Board (protocol 2087052‐3). Data collection Before the focus groups, participants completed a brief demographic questionnaire, including age, gender, race, ethnicity, JIA category, household income, and highest educational attainment. The questionnaire also included a question related to participant's familiarity with AI (“how familiar are you with AI technology?”). Focus groups were cofacilitated by three research team members, including a primary moderator and two patient partners (ie, individuals with lived experience of a related health condition who participate as active members of a research team). Both patient partners are parents of children with JIA. Each focus group began with a brief overview of AI, with an emphasis on its ability to learn from existing clinical data to predict outcomes in previously unseen data. Basic AI concepts were introduced, including the selection of input features for model development and the evaluation of model performance. Model evaluation was discussed with specific reference to metrics such as the area under the receiver operating characteristics curve (AUC), which quantifies the tool's ability to distinguish between two groups on a scale from 0 to 1, where 1.0 indicates perfect discrimination and 0.5 reflects performance no better than chance. Participants were then asked to reflect on the use of AI models to predict disease progression and guide treatment decisions. To ground the discussion, we presented an AI model that our team developed for predicting disease trajectories in JIA, including a review of the input features used for model development and the performance of the model (ie, AUC ~0.80). We then asked about participant comfort and concerns (“How would you feel about this technology being used for you or your loved one?”), perceived benefits and risks (“What do you think are the benefits and risks?”), expectations on model performance (“Is there an AUC threshold at which you will feel comfortable using the tool?”). To further understand participant expectations regarding AI model calibration, we asked the following question: “Which is more important to you – minimizing harm of aggressive treatment, or maximizing the chance of achieving inactive disease?” To assess the informational needs of participants when AI models are used to inform treatment recommendations, we asked the following questions, “How would you feel if you were to find out that your doctor used an AI tool to guide treatment recommendations without informing you?” and “What information should a provider disclose to a patient before using an AI tool for guiding treatment recommendations?” Finally, we explored patient perceptions regarding the issue of accountability in the context of AI use in health care: “Who should be responsible when AI gets it wrong?” The discussion concluded with a broader question aimed at addressing potential concerns: “What can we do to address your concerns about the use of AI models to inform care?” Following each focus group, a summary of the discussion was emailed to all participants to invite additional feedback (ie, member checking). Focus groups were recorded and transcribed. Transcripts were analyzed by two investigators through a process of independent and collaborative thematic analysis. 6 Team members first familiarized themselves with the data through independent review of transcripts. Initial codes were then generated to capture meaningful features of the data. The research team met to discuss and refine codes, developing a shared codebook through iterative consensus. Coded data were subsequently reviewed and synthesized to identify and refine candidate themes. Finally, themes were reviewed and collaboratively interpreted to generate a consensus summary of key findings. Upon completion of the analysis, the study team obtained feedback from patient partners involved in the focus groups as a final form of member checking to ensure validity of the findings. RESULTS The study included a total of 18 participants, 8 of whom were patients with JIA (3 children and 5 adults) and 10 were parents of a child with JIA (Table 1). Thematic saturation was reached after the fourth focus group. The major themes identified are detailed subsequently and summarized in Supplementary Table 1. Table 1Study participants (n = 18)* AttributenPatient with JIA or their caregiverPatient8Caregiver10Age, years14–17318–253≥2512GenderFemale15Male3Self‐reported raceWhite15Black2Asian1Self‐reported ethnicityHispanic1Non‐Hispanic17Highest level of educationSome high school (grades 9–11)3Graduated high school (grade 12 or GED)1Graduated college or technical school6Graduate school5Not reported3Field of occupationHealth care4Information technology1Does not involve health care or information technology13Familiarity with AINot familiar at all (I am not sure what AI is)0Slightly familiar (I am aware of examples of AI in daily life)17Moderately familiar (I understand what is required to develop AI applications)1Very familiar (I can explain to you the algorithms used for developing AI)0*AI, artificial intelligence; GED, general educational development; JIA, juvenile idiopathic arthritis. Study participants (n = 18)* AI, artificial intelligence; GED, general educational development; JIA, juvenile idiopathic arthritis. The perceived utility, benefits, and risks of AI Most participants expressed enthusiasm for AI‐guided treatment decision‐making, particularly those who recognized the inherent uncertainty in treatment choices and the consequences of delayed intervention in JIA. This group included patients with severe disease facing decisions about the timing of biologic initiation and those who had experienced complications from treatment delays. However, participants universally emphasized that AI models should serve as clinical decision support tools rather than physician replacements, “Having an additional tool will be helpful, but I don't want doctors to rely solely on AI.” Several participants expressed hesitancy about fully trusting AI models, stating that they would feel more comfortable with AI‐guided decisions if they understood how AI model makes its predictions, “AI does not explain how it comes up with the answer. If I understand why, I would be more likely to be persuaded.” Participants identified several potential benefits of AI tools, including earlier diagnosis and more proactive treatment through predictive modeling, advancement of medical knowledge, and facilitation of personalized medicine approaches. The most consistently cited risk was physician over‐reliance on AI tools, with overwhelming concerns that such dependence could erode clinicians’ critical thinking and clinical reasoning, “How do we ensure physicians are actually going to put their own thought into it?” AI tools may also inadvertently diminish the weight given to clinicians’ contextual knowledge of the patient and previous clinical experience—insights that no algorithm can fully capture. Privacy protection emerged as another prominent concern, particularly regarding children's inability to advocate for their own data protection, “It's one thing for your data to get shared, but it's another thing for your child's, especially when they can't really advocate for themselves.” Additional concerns included the potential impact on shared decision‐making processes and patient–provider relationships and algorithmic bias, as one Black participant remarked, “Most studies comprise White patients. How do I know the model is relevant to someone like me?” Disclosures of AI use When asked whether physicians should disclose AI use in clinical practice, responses varied considerably. Five participants did not perceive disclosure to be necessary. This group recognized that AI is already widely integrated into health care delivery without routine patient notification. Six participants favored disclosure, emphasizing its importance for building trust: “I don't think it's hugely important to inform, but I think that they should… always explain why they're recommending a certain treatment. I think that's just part of the trust and the shared decision making.” Two participants expressed particularly strong feelings, stating that undisclosed AI use would permanently damage their trust in their provider, “If I found out they've been using a tool and not telling me, like do you even know what you're talking about?” The remaining seven participants were undecided about their preference. Information needed to foster trust in AI systems We explored participants’ informational needs to foster understanding and trust. Participants identified several key elements: training data sample size and source, model performance, model provenance (including potential conflicts of interest), and study cohort characteristics (as one participant asked, “How much does my son look like the patients that are in that model?”). Participants did not consider information about the specific variables used to train AI models as essential for building trust in these systems. In fact, such technical details may overwhelm patients with excessive information. Several participants suggested offering informational pamphlets for patients seeking additional details. Informational needs varied based on the strength of patient–provider relationships, with one participant noting, “if you're a longtime patient or caregiver, you've had time to build the relationship with that physician and you've built the trust. If you're a new patient, you might need more explanations.” Participants emphasized the importance of physician reassurance on two fronts: (1) that AI serves only as a supportive tool from which clinicians can deviate, and (2) that declining AI recommendations does not preclude achieving disease remission. To understand how to effectively communicate model performance, we introduced participants to the concept of AUC. Participants perceived AUC as a valuable metric for assessing treatment effectiveness. Knowledge of this performance measure enhanced their comfort and reassurance when considering AI‐informed recommendations. Participants generally associated higher AUC values with greater model reliability, which strengthened their trust in both the technology and the resulting treatment decisions. However, several participants reported difficulty comprehending model performance metrics. Participants emphasized that information presentation is equally important as content, noting that effective communication and thoughtful patient education are essential for building trust, enhancing understanding, and facilitating informed decision‐making. Acceptable model performance To determine acceptable model performance, we asked participants about minimum AUC thresholds upon which to base a treatment prediction. Most stated they would want an AUC of at least 0.7 to 0.8 before they would feel comfortable with AI‐guided recommendations. In contrast, several participants considered an AUC >0.5 acceptable, noting this would exceed chance prediction given treatment uncertainty. These participants’ acceptance of lower performance thresholds stemmed from their recognition that evidence for optimal JIA management remains limited and current practice often relies on trial‐and‐error approaches. Awareness that conventional treatment decisions also involve uncertainty made them more receptive to imperfect AI tools. As one participant noted, “When we were prescribed our first biologic, we were literally told we don't know if this is going to work or not… So, I would say if it's better than 50%, it's worth using as a tool.” Framing the AI tool as an aid, rather than a replacement, for clinical judgment increased patient comfort with its use, even at lower levels of accuracy. Model calibration AI model performance typically involves trade‐offs between sensitivity and specificity. To understand optimal model calibration for JIA treatment decision‐making (ie, whether to prioritize sensitivity or specificity), we asked participants whether maximizing remission probability or minimizing treatment‐related harms was more important. Five participants prioritized maximizing remission probability; these patients typically had experienced more severe disease courses and recognized the importance of early aggressive intervention. Two participants prioritized harm minimization, citing concerns about the potential side effects of biologic therapies, particularly on reproductive health; both were young adults with well‐controlled disease. The majority (n = 11), however, stated that the optimal balance depends on disease severity, patient preferences, life circumstances, and lifestyle factors. These participants emphasized that AI models should incorporate individual considerations, “There are different treatment burdens that I think that's kind of personal for each individual patient.” One participant highlighted the importance of considering cultural beliefs in shaping treatment preferences and the potential conflicts these beliefs can create between patients and their caregivers, “I think also taking into account culture… my parents really wanted me to try Asian medicine and stuff like that. That kind of steered me away from going on to the biologic as soon as possible, which looking back now, I 100% regret.” Accountability for AI‐related errors The use of AI in health care raises legal concerns about who should be held accountable when medical errors occur with AI assistance. 7 When asked about accountability for AI‐related errors, perspectives varied considerably. Two participants assigned responsibility to physicians, reasoning that clinicians ultimately make the final treatment decisions. One participant attributed responsibility to AI developers and the institutions implementing the technology. The majority, however, rejected the notion of assigning blame, particularly when AI serves only as a decision support tool. These participants noted that physicians also make errors and that treatment response uncertainty is inherent to clinical practice, making blame attribution inappropriate. A number of participants highlighted that transparency in the use of AI could serve to mitigate issues related to accountability for AI‐related errors, “I think that's one of the reasons I feel like if a physician is going to use AI to help, they need to be open about it. Let patients be informed, because liability in this country is such a sticky issue… Liability comes down to not informing.” DISCUSSION We conducted the first qualitative study to examine the perspectives of patients with JIA and their parents on the use of AI technology to facilitate treatment decision‐making. Our study shows that patients with JIA and their parents hold nuanced perspectives on AI integration into treatment decision‐making: although enthusiastic about AI's potential to support personalized treatment, their acceptance was universally contingent on AI serving as a clinical decision support tool rather than replacing physician judgment. This sentiment is consistent with published studies in other settings. 8 , 9 , 10 For some participants, acceptance also depends on transparency (ie, disclosure of AI use to patients) and interpretability (ie, the ability to understand the rationale behind a model's prediction). To trust AI recommendations, participants sought specific information about training data sources and size, model performance, model provenance and potential conflicts of interest, and cohort representativeness. Participant interpretation of acceptable model performance was also shaped by their lived experiences with JIA. Those who had navigated severe disease courses and understood the trial‐and‐error nature of current treatment approaches demonstrated greater tolerance for imperfect AI predictions, recognizing that even modest improvements over chance would be valuable given the inherent uncertainty in treatment outcomes. Beyond technical performance of the AI prediction models, participants emphasized that successful AI implementation must preserve the patient–provider relationship and accommodate individual circumstances. The majority rejected one‐size‐fits‐all approaches, emphasizing that models should account for disease severity, patient preferences (including needle phobia and cultural beliefs), life stage, and lifestyle factors. In discussions of sensitivity–specificity trade‐offs, priorities diverged based on disease experiences and life circumstances. For example, patients who developed permanent joint damage due to delayed treatment expressed greater tolerance for the potential side effects of early aggressive treatment, having experienced firsthand the consequences of undertreatment. Conversely, patients of reproductive age voiced concerns about how biologics might affect their fertility or pregnancy outcomes. These varied perspectives suggest that AI tools must accommodate personalized risk tolerances, allowing clinicians and patients to adjust sensitivity–specificity trade‐offs based on individual values and circumstances. Participants also identified several critical concerns. The primary concern was physician over‐reliance on AI diminishing clinical judgment. Privacy concerns were also frequently cited, especially for pediatric populations unable to self‐advocate. Although training data may be de‐identified, privacy risks persist because combinations of clinical details can remain uniquely identifying—a risk heightened in rare diseases such as JIA, in which distinctive patterns of diagnoses, timelines, or treatments can make individuals recognizable. Our findings have important implications for multiple stakeholders in the implementation of AI tools to support treatment decision‐making. There is a need for policymakers to consider whether disclosures and informed consent are necessary when such tools are used in clinical care. Published research on health care AI consistently identifies consent as a key concern for both patients and clinicians 11 , 12 , 13 , 14 ; however the consequences of withholding AI disclosure on patient–provider trust remain largely unexamined. The complexity of this challenge is illustrated by a recent study examining AI‐assisted ambient documentation, 11 in which 82% patients consented when provided with basic information about the technology and only 55% consented when given details about AI features, data storage, and corporate involvement were disclosed. Similar studies are needed to examine how varying levels of disclosure about treatment prediction models influence patient consent and perceptions, alongside research on communication strategies that convey model performance to patients with varying health literacy and on patient education materials that build AI literacy. Importantly, studies are needed to understand how AI use affects patient–provider relationships and shared decision‐making making. A study showed that women's perception of their physician's competence and integrity were significantly lower when their physicians used a clinical decision support system compared to when they did not. 15 AI‐assisted ambient documentation has also been reported to undermine provider–patient relationships, with patients withholding sensitive information due to privacy concerns. 11 Furthermore, the opaque nature of AI tools may present challenges for shared decision‐making. Without clear explanations of how predictions are generated, patients may struggle to weigh AI recommendations against their own preferences, whereas clinicians may question whether they can ethically endorse recommendations they do not fully understand. 16 The result could be “decision paralysis,” whereby uncertainty about AI's reasoning process leaves both clinicians and patients hesitant to act, potentially delaying important treatment decisions. 17 There is also a need for research examining whether and how AI tools affect physician clinical reasoning. Earlier studies have found evidence of automation bias, whereby clinicians defer to AI‐generated outputs even when these outputs conflict with their clinical judgment, 18 , 19 posing risks to patient safety. Developing strategies to mitigate automation bias is therefore critical. Potential approaches include training clinicians to critically evaluate AI recommendations, designing AI systems that encourage critical thinking rather than passive acceptance, and implementing safeguards that require explicit clinical justification when AI outputs are accepted despite initial clinical doubt. 20 , 21 For AI developers, there is a critical need to develop explainable AI models that can provide justifications for predictions, as patients are more likely to trust and accept medical decisions when they understand the reasoning behind them. Additionally, AI models must also incorporate patient preferences, lifestyle factors, and cultural beliefs, alongside traditional clinical variables, to support patient‐centered decision‐making. McCradden et al 19 proposed a framework for ethical clinical decision‐making using machine learning that addresses this challenge. The framework emphasizes that model outputs must be contextualized with direct knowledge of the individual patient, positioning the patient as the “anchor” for verifying whether a prediction is reasonable for their specific circumstances. 19 Without accounting for patient‐specific factors, AI‐driven recommendations risk undermining patient autonomy and producing clinically inappropriate guidance despite technical accuracy. 22 , 23 Our study has several important limitations. First, our sample consisted of highly engaged patients and caregivers, the majority of whom were college educated. This limits the generalizability of our findings, as the views expressed may not fully represent the broader population of patients and families, particularly those with lower educational attainment or limited access to digital resources. Second, all but one participant had been diagnosed three or more years before the study. The perspectives captured here may not reflect the perspectives of individuals navigating early treatment. Future research should therefore prioritize dedicated qualitative studies of these populations, using designs tailored to the specific informational and decision‐making contexts of newly diagnosed patients and families from diverse educational, socioeconomic, and cultural backgrounds. Despite these limitations, our study offers several notable strengths. The inclusion of adolescents and young adults with a rare disease like JIA represents a significant contribution, as few studies have meaningfully engaged this population in discussions about AI‐driven clinical tools. Additionally, we explored critical yet understudied issues, including patients’ informational needs regarding AI and their expectations of AI performance, both of which are essential to inform real‐world implementation and ensure that AI tools are designed with patient values and preferences in mind. Addressing the research gaps identified in our study will be essential for ensuring that AI integration enhances rather than undermines the care of children with chronic diseases. AUTHOR CONTRIBUTIONS All authors contributed to at least one of the following manuscript preparation roles: conceptualization AND/OR methodology, software, investigation, formal analysis, data curation, visualization, and validation AND drafting or reviewing/editing the final draft. As corresponding author, Dr Ong confirms that all authors have provided the final approval of the version to be published and takes responsibility for the affirmations regarding article submission (eg, not under consideration by another journal), the integrity of the data presented, and the statements regarding compliance with institutional review board/Declaration of Helsinki requirements. Supporting information Disclosure Form: Supplementary Table 1. Major themes and representative quotes.