Perceptions of healthcare professionals in rheumatology on the use of prediction models in clinical practice Authors: Sina Fadaei, Patty de Groot, Rob R G Kuin, Julia Spierings, Amin Herman, Paco M J Welsing, Lise M Verhoef Published: 2026 ABSTRACT AbstractObjectivesPrediction models are increasingly developed to aid decision-making in RA, but none are currently included in guidelines. The aim of this study was to explore rheumatology healthcare professionals’ perceptions of using prediction models in everyday clinical practice.MethodsAn exploratory qualitative study was conducted using semi-structured interviews with rheumatologists and nurse practitioners involved in RA care in the Netherlands. Participants were purposively sampled based on work experience, academic degree, profession and care setting. Interviews were transcribed and analysed thematically following Braun and Clarke’s framework.ResultsFifteen interviews uncovered for overarching themes. First, building support for prediction models depended on trust, awareness, understanding of model inputs and limitations, perceived clinical usefulness, evidence and cost-effectiveness. Second, making prediction models workable in routine care required integration into electronic health records, minimal additional workload, reliance on routinely collected data and actionable outputs. Third, healthcare professionals (HCPs) positioned prediction models within clinical judgement and patient care: models were considered valuable when supporting, rather than replacing, professional judgement and shared decision-making, while helping to communicate uncertainty and clarify difficult decisions. Fourth, HCPs anticipated consequences for care organization, medication use, professional autonomy and personalized care, including concerns about misuse by external parties, compulsion and unintended workload or loss of patient visibility.ConclusionRheumatology healthcare professionals value prediction models, but only if they are interpretable, clearly beneficial, integrated into workflows and protect professional autonomy. Addressing these concerns is essential for turning models into useful clinical decision tools.Keywords: prediction models, qualitative research, machine learning, clinical decision support, RA Healthcare professionals welcome prediction models when these are interpretable, useful and fit clinical reasoning. Healthcare professionals welcome prediction models when these are interpretable, useful and fit clinical reasoning. Successful implementation of prediction models depends on workflow integration and usability in addition to algorithmic sophistication. Successful implementation of prediction models depends on workflow integration and usability in addition to algorithmic sophistication. Concerns about misuse, autonomy and unintended effects were raised by healthcare professionals. Concerns about misuse, autonomy and unintended effects were raised by healthcare professionals. Introduction RA is a systemic autoimmune disease mainly affecting the joints, potentially causing long-term damage and disability. Although advances in treatment have improved the prognosis for many patients, significant challenges remain. Thirty to 40% of patients do not respond to first-line therapies, and 60–80% experience disease flares when tapering biological DMARDs (bDMARDs). Treatment often concerns a trial-and-error approach, which can lead to a prolonged period of suboptimal treatment and inefficient use of healthcare resources [1, 2]. Clinical decision aids based on predictive models have the potential to address some of these challenges in RA care. In recent years, numerous predictive algorithms with diverse applications have been developed and validated [3]. Examples include models that predict the response of patients to methotrexate or that predict flares when tapering biologics [4, 5]. Current research on artificial intelligence (AI) highlights the future potential of models in rheumatology, for example, for making treatment decisions through machine learning [6, 7]. Despite numerous advancements in the development of prediction models in rheumatology, to date, no prediction models have been recommended for use as a decision aid in guidelines [2]. Commonly cited barriers to the adoption of predictive models and clinical decision support (CDS) in clinical care are the lack of streamlining in workflows, lack of evidence that they improve current care, financial or infrastructural limitations, and physician preferences and attitudes [8–11]. In rheumatology, where shared decision making and long-term patient-doctor relationships are key, clinicians may hold specific views on the role of predictive models in daily practice. Therefore, a deeper understanding is necessary of the circumstances under which rheumatology healthcare professionals (HCPs) would be willing to use prediction models and which factors they consider important when deciding whether to do so [12, 13]. The aim of this study was to explore HCPs perceptions of using prediction models in everyday clinical practice. Methods This is an exploratory qualitative study. We conducted semi-structured interviews with HCPs involved in inflammatory arthritis care, exploring their perceptions, expectations and concerns regarding the use of prediction models in clinical practice. This study does not fall under the scope of the Dutch Medical Research Involving Human Subjects Act (WMO). It therefore does not require formal approval from an accredited medical ethics committee in the Netherlands, but a compliance check was done by the hospital’s quality officer (file number 24U-1538). Participants HCPs currently involved in the pharmacological treatment of RA patients were eligible, which in the Netherlands meant rheumatologists, physician assistants and nurse practitioners in the field of rheumatology. Purposive sampling was applied based on age, gender, the type of hospital in which the HCP worked (academic or peripheral clinic), years of work experience and academic degree. HCPs were recruited from a variety of healthcare settings, including academic hospitals, general hospitals and a standalone rheumatology treatment centre, within the Netherlands. The professional networks of the authors were used to approach HCPs. HCPs were provided with an information form including background information (Supplementary file S1, available at Rheumatology Advances in Practice Online) and asked to participate via e-mail. Informed consent was obtained in writing before interviews were held. Recruitment was halted when data saturation was reached, for example, no new themes emerged from three subsequent interviews. Data saturation was reached after 15 interviews and Participant characteristics are shown in Table 1. Table 1Characteristics of participants.Age (mean, min–max)49 (33–67)Gender (female, %)9 (60%)Profession (rheumatologist,%)11 (73%)Academic degree (PhD, %)7 (47%)Years of experience (mean, min–max)12 (0.5–26)Hospital type (academic, %)5 (33%) Characteristics of participants. Interviews Data were collected through semi-structured interviews using an interview guide between October 2024 and July 2025. The guide consisted of four main questions and additional follow-up questions for specific topics if the HCP raised them. The interview guide was informed by theoretical domains framework (TDF) of behaviour change [14]. The TDF was used to support broad coverage of potentially relevant domains related to the future use and implementation of prediction models. The interviews were conducted flexibly rather than following a fixed sequence. Each interview consistently began with a broad, open-ended question about the participants’ initial thoughts on using prediction models in daily practice. Other general open questions were addressed throughout the interview in an unstructured manner but aimed to be touched upon across the session. The domain-based TDF questions were not used as a standardized checklist. Instead, they served as examples of possible follow-up prompts to explore topics spontaneously raised by participants. These prompts were used only when relevant and not necessarily in every interview. Before the study began, the interview guide was pilot tested in two interviews and adjusted accordingly. The complete interview guide in Dutch is provided in Supplementary file S2, available at Rheumatology Advances in Practice Online. Interviews lasted an average of 35 min. Out of 15 interviews, 13 were conducted by two researchers (S.F. and P.d.G.), and the other two by S.F. alone. S.F. is a male medical doctor and a PhD candidate in prediction models and clinical RA care with qualitative research experience. P.d.G. is a female PhD candidate with a background in industrial design engineering. She has experience conducting qualitative research in the domain of digital biomarkers for rheumatology. Most interviews were conducted either in person (N = 8) or via Microsoft Teams (N = 5). Two interviews were conducted over the phone at the interviewee’s preference. S.F. knew two of the interviewees beforehand as part of clinical duties. This relation was reflected upon during analysis to account for possible bias. P.d.G. had no prior ties with the interviewees. The interview began with a brief summary of this information and clarification of any ambiguities the interviewee might have had. The interview questions and techniques were continuously assessed throughout the process (S.F., P.d.G., L.V.). Interviews were transcribed using AI transcription with the Whisper large-v3 model. The generated transcripts were compared with the audio recordings, and errors were corrected (S.F.). The complete transcripts were sent back to the participants for their feedback. Data analysis We used the six phases of thematic analysis by Braun and Clarke [15]. All transcripts were coded independently by S.F. and P.d.G., after which the codes were compared, and a consensus was sought. Data analysis with Atlas.ti was initiated after four interviews. Throughout, codes were grouped into subthemes and preliminary themes and continuously adjusted to reflect the expanding dataset. The TDF contributed to developing the interview guide and supported data collection, but it was not employed as a deductive analytic framework. The analysis was primarily thematic and data-driven. During the process, two triangulation sessions with the whole study team were held to increase the dependability and reliability of the findings. An expert in qualitative research (L.M.V.) was involved in supervising data collection and analysis to strengthen the accuracy of the process. The Consolidated Criteria for Reporting Qualitative Studies (COREQ) checklist was used to report the results [16]. Results Our analysis identified four overarching themes describing HCPs’ perceptions of using prediction models in rheumatology practice: (1) building support for prediction models, (2) making prediction models workable in routine care, (3) positioning prediction models within clinical judgement and patient care, and (4) anticipating consequences of prediction model use (Table 2). Across these themes, HCPs described prediction models as potentially valuable tools, but only under specific conditions related to trust, usability, clinical fit and broader effects on care. Table 2Overview of themes and subthemes.ThemeDescriptionSubthemeQuotesBuilding support for prediction modelsSupport for prediction models depended on whether HCP’s considered them trustworthy, understandable and meaningfully beneficial in practice. Support was also shaped by HCP background and the working environment.Initial (un)acceptance of models [Participant 3]: It’s as if there is a kind of natural resistance to prediction models, as if we don’t want to use them or don’t really believe in them. I think much more trust needs to be built. If they are genuinely good, why wouldn’t we want to use them? [Participant 1]: Yes, that’s word-of-mouth advertising. But that works best when you’re in a resident setting. And after that, you’re always working alone. And then you hear something every now and then. Well, here in this teaching hospital, thankfully, you do hear something every now and then… [Participant 6]: If a model indicates that you should stop an activity, or that you must do this or that, there will always be situations that deviate from it. And then it is important to know what such a model is based on. [Interviewer]: How well established does it [the prediction model] need to be before you would want to use it? [Participant 12]: I think somewhere between just one paper and the EULAR. EULAR might be very ambitious. And for that you need more than one trial. One paper is probably not enough either. So a number of successful papers would probably be good. Wait-and-see attitudeEnthusiasm for innovationInevitability of innovationPrior experienceUser characteristics shaping receptivenessPerception of achievabilityMisconceptions about prediction modelsAwareness of existing prediction modelsAvailability of prediction modelsColleagues’ experiences with prediction modelsWhat variables are taken into accountUnderstanding how variables contribute to model predictionsUnderstanding model limitationsPrediction models should solve a practical problemPrediction models should have solid evidence behind themPrediction models should add value to regular carePrediction models should be cost-effective [Participant 7]: Before doctors get used to that—we saw that with the implementation of protocols, for example—in the beginning, it wasn’t at all, but at some point it was, but before protocol adherence was so good that you could really say, 80%… works according to those protocols, you’re also quite a bit further along. I think it would be something like that for prediction models as well.[Participant 9]: If we still have to start filling things in for this, then it will never, never, never, never, ever be implemented … We already have to fill in so much … but in the long run it simply won’t work if doctors also have to take on this task. That would be a disaster … I have to say, I really hate it, because it often takes a lot of time. Time that is taken away from actually looking the patient in the eyes. Making prediction models workable in routine carePrediction models were considered workable only when they could be embedded in routine care without adding substantial additional burden.Logistical challenges of implementationConservative attitudes and adjustment problemsPreventing additional workloadLimited expected disruption to working proceduresExpected timeline of implementationIntegration into electronic health recordsFrequency of applicationReturn on time investmentUsable and actionable outputPractical data collection [Speaker 5]: So, I think it is the complexity of the disease profiles, the medications, and the patients themselves, each with their own comorbidities. That makes it very difficult, I think, to predict things with real precision. [Speaker 11]: Yeah, no, well, I think you shouldn’t entirely rely on a prediction model… you make decisions together with the patient… you have your own ideas, the patient has theirs… and a prediction model could help support those decisions. [Participant 11]: I see it this way; sometimes you have patients who are a bit hesitant to taper, for example, a biological…, and if you can then say,… you’ve been using it for so long, the disease has been in remission for so long, we know that for a certain percentage of people, things will actually go well. There’s only a very small chance of a flare-up of 10 or 20%. I think that then you might be able to convince those people to support a decision to, for example, taper Positioning prediction models within clinical judgement and patient carePrediction models were viewed as decision-support tools whose value depends on alignment with clinical reasoning and patient-centred careRole of the model in decision makingNuances and contextual factors in patient wellbeingTempering human bias and habitual decision-makingProviding clarity and reducing uncertaintyUsing your own judgementManaging disagreement and accountabilityLearning to work with probabilitiesPatient suitability and selectionImpact on patient-doctor relationshipUse of prediction models in communication with patientsPerceptions of benefits for patientsAnticipating consequences of prediction model usePrediction models were expected to influence care delivery in ways that may improve efficiency and personalization, but also introduce unintended pressures and risks.Use for the wrong ends [Participant 2]: And finally, it came up in my annual reviews. It was always, yes, you filled in too few DAS assessments, or your average DAS is too high. Then I thought, we are moving from a score meant for trials to personally evaluating my work … So I wondered, will the same happen with a prediction model? Will an advice come, and I ignore it, or not always follow it. Will I have to justify that later? Negative influence of external partiesFear of compulsionUnintended consequencesPersonalized medicineEfficiency and organization of careUtilization of (expensive) medicineHCP: healthcare professional. Overview of themes and subthemes. [Participant 3]: It’s as if there is a kind of natural resistance to prediction models, as if we don’t want to use them or don’t really believe in them. I think much more trust needs to be built. If they are genuinely good, why wouldn’t we want to use them? [Participant 3]: It’s as if there is a kind of natural resistance to prediction models, as if we don’t want to use them or don’t really believe in them. I think much more trust needs to be built. If they are genuinely good, why wouldn’t we want to use them? [Participant 1]: Yes, that’s word-of-mouth advertising. But that works best when you’re in a resident setting. And after that, you’re always working alone. And then you hear something every now and then. Well, here in this teaching hospital, thankfully, you do hear something every now and then… [Participant 1]: Yes, that’s word-of-mouth advertising. But that works best when you’re in a resident setting. And after that, you’re always working alone. And then you hear something every now and then. Well, here in this teaching hospital, thankfully, you do hear something every now and then… [Participant 6]: If a model indicates that you should stop an activity, or that you must do this or that, there will always be situations that deviate from it. And then it is important to know what such a model is based on. [Participant 6]: If a model indicates that you should stop an activity, or that you must do this or that, there will always be situations that deviate from it. And then it is important to know what such a model is based on. [Interviewer]: How well established does it [the prediction model] need to be before you would want to use it? [Interviewer]: How well established does it [the prediction model] need to be before you would want to use it? [Participant 12]: I think somewhere between just one paper and the EULAR. EULAR might be very ambitious. And for that you need more than one trial. One paper is probably not enough either. So a number of successful papers would probably be good. [Participant 12]: I think somewhere between just one paper and the EULAR. EULAR might be very ambitious. And for that you need more than one trial. One paper is probably not enough either. So a number of successful papers would probably be good. [Participant 7]: Before doctors get used to that—we saw that with the implementation of protocols, for example—in the beginning, it wasn’t at all, but at some point it was, but before protocol adherence was so good that you could really say, 80%… works according to those protocols, you’re also quite a bit further along. I think it would be something like that for prediction models as well. [Participant 7]: Before doctors get used to that—we saw that with the implementation of protocols, for example—in the beginning, it wasn’t at all, but at some point it was, but before protocol adherence was so good that you could really say, 80%… works according to those protocols, you’re also quite a bit further along. I think it would be something like that for prediction models as well. [Participant 9]: If we still have to start filling things in for this, then it will never, never, never, never, ever be implemented … We already have to fill in so much … but in the long run it simply won’t work if doctors also have to take on this task. That would be a disaster … I have to say, I really hate it, because it often takes a lot of time. Time that is taken away from actually looking the patient in the eyes. [Participant 9]: If we still have to start filling things in for this, then it will never, never, never, never, ever be implemented … We already have to fill in so much … but in the long run it simply won’t work if doctors also have to take on this task. That would be a disaster … I have to say, I really hate it, because it often takes a lot of time. Time that is taken away from actually looking the patient in the eyes. [Speaker 5]: So, I think it is the complexity of the disease profiles, the medications, and the patients themselves, each with their own comorbidities. That makes it very difficult, I think, to predict things with real precision. [Speaker 5]: So, I think it is the complexity of the disease profiles, the medications, and the patients themselves, each with their own comorbidities. That makes it very difficult, I think, to predict things with real precision. [Speaker 11]: Yeah, no, well, I think you shouldn’t entirely rely on a prediction model… you make decisions together with the patient… you have your own ideas, the patient has theirs… and a prediction model could help support those decisions. [Speaker 11]: Yeah, no, well, I think you shouldn’t entirely rely on a prediction model… you make decisions together with the patient… you have your own ideas, the patient has theirs… and a prediction model could help support those decisions. [Participant 11]: I see it this way; sometimes you have patients who are a bit hesitant to taper, for example, a biological…, and if you can then say,… you’ve been using it for so long, the disease has been in remission for so long, we know that for a certain percentage of people, things will actually go well. There’s only a very small chance of a flare-up of 10 or 20%. I think that then you might be able to convince those people to support a decision to, for example, taper [Participant 11]: I see it this way; sometimes you have patients who are a bit hesitant to taper, for example, a biological…, and if you can then say,… you’ve been using it for so long, the disease has been in remission for so long, we know that for a certain percentage of people, things will actually go well. There’s only a very small chance of a flare-up of 10 or 20%. I think that then you might be able to convince those people to support a decision to, for example, taper HCP: healthcare professional. Building support for prediction models Support for prediction models was shaped by HCPs’ initial attitudes, their awareness of available models, their ability to understand how predictions were generated, and the standards they applied regarding usefulness and evidence. HCPs showed a range of initial attitudes towards prediction models and innovation overall. Some enjoyed the process of innovation, including prediction models, and saw these tools as an inevitable part of medical progress. Others expressed hesitation or only partial acceptance. A common approach was to adopt a wait-and-see stance, with HCPs believing the models need to demonstrate their value in real-world situations and that only time will reveal if they meet expectations. Several HCPs based their views on previous experiences with earlier innovations, both positive and negative, suggesting that repeated encounters with failed tools have led some to become ‘innovation-weary’. Attitudes were also seen as being influenced by end-user characteristics such as personality, career stage, or openness to data-driven reasoning; younger HCPs were often regarded as more receptive, while more experienced colleagues relied more on clinical intuition and valued professional autonomy. Finally, opinions on the feasibility of prediction models diverged: some doubted their practicality in their field, while others were cautiously optimistic that advances in AI and data integration would eventually enable the effective application of prediction models. Beyond these general attitudes, HCPs highlighted that being well-informed is vital for shaping their perceptions of prediction models. Some HCPs clearly understood these models and their clinical applications, while others held misconceptions or expressed uncertainty, such as believing that current treatment protocols already serve as prediction models. A lack of knowledge about existing prediction models and how to access them was seen as a barrier to adopting these tools. Most information was gathered indirectly through colleagues, departmental discussions or professional meetings. Hearing peers’ experiences, whether positive or negative, influenced expectations and trust. HCPs recommended increasing awareness of effective models and providing better access to current, reliable information to promote broader acceptance and use. In addition, HCPs emphasized the necessity of understanding which variables are used in prediction models and how they affect the outputs. This knowledge is considered crucial for assessing whether model predictions match clinical experience and for building trust in their application. Most HCPs agreed that understanding the model’s inputs and logic helps foster confidence, but opinions varied on whether a full understanding of the computations is necessary. Some believed the mathematical details might be too complex for most clinicians, while others argued that physicians should at least be able to interpret and explain model outputs to patients. Additionally, HCPs pointed out that clinicians should understand how prediction models are developed to identify potential limitations, such as differences between the development and target populations. HCPs outlined several performance criteria that predictions should meet before they are adopted in clinical practice. Models must address a clear clinical issue and offer benefits beyond current standards of care, supported by strong evidence. HCPs further emphasized the importance of cost-effectiveness and a confidence margin that aligns with clinicians’ decision-making. While some HCPs believed that guideline endorsement (e.g. EULAR or national) was essential, others were open to earlier implementation based on a few robust scientific publications. Making prediction models workable in routine care HCPs perceived the workability of prediction models in routine care as influenced by both the wider implementation context and the specific practical design features of the models. HCPs anticipated several challenges in integrating prediction models into routine clinical practice. They expected adjustment difficulties, noting that physicians tend to be conservative and slow to adopt new tools. Logistical and financial barriers were also highlighted, such as the need for unified IT systems and sufficient resources for integration. It was observed that deploying digital innovations in hospitals generally takes years and advances more slowly than expected. A concern was that prediction models might be seen as ‘something extra’, requiring the removal of other tasks before they could be realistically implemented. On the other hand, other HCPs believed that prediction models would not significantly change existing workflows. Overall, HCPs did not expect rapid implementation; instead, they anticipated gradual adoption over several years as systems, infrastructure, and attitudes evolve. Beyond these broader implementation challenges, HCPs emphasized that the practical design of prediction models would be crucial to their actual use in daily care. HCPs highlighted that prediction models should be fully integrated into electronic health records to ensure practicality and routine use in daily care. When a model is embedded within the patient file, automatically utilizing existing clinical data and showing results without manual input, it naturally becomes part of clinical reasoning rather than an extra task. On the other hand, models requiring separate logins, duplicate data entry, or additional forms were seen as impractical and unlikely to be adopted. The frequency of use is also crucial: models that need frequent input (e.g. at every visit) were considered unrealistic, while those used ‘as needed’ or at key decision points were seen as more feasible. HCPs emphasized that the time saved must be evident. If the model streamlines consultations or helps patients prepare, it adds value. However, an increase in administrative tasks would discourage use. HCPs implicitly viewed models through a return-on-investment lens. Complex models or gathering more data were acceptable if they yielded clear long-term benefits, such as fewer consultations or better use of treatment. The acceptable level of model complexity depended on context: in emergencies, simplicity and speed were critical; for non-urgent issues like tapering, more detailed calculations were permissible if they provided clear value. The output needs to be usable and actionable; some HCPs preferred raw probabilities, while others wanted specific advice from the model. The advice needs to be practical in a clinical setting. Finally, data collection for predictions should primarily rely on information already gathered in routine care or automatically obtained from patients, since models that require complex or expensive inputs are unlikely to offer a good return on investment. Positioning prediction models within clinical judgement and patient care HCPs positioned prediction models as supportive tools within care, whose value depended on how well they aligned with clinical judgement, addressed uncertainty and accounted for the relational aspects of patient care. HCPs highlighted that prediction models should support, not replace, clinical judgement. They stressed the importance of ‘relying on your own judgement’, maintaining independent reasoning. Many believed models cannot fully capture individual patient nuances, such as emotional states, pain unrelated to inflammation, comorbidities, or treatment preferences. Several pointed out that models are less capable than the human brain at integrating the complexity of real-life clinical presentations. Conversely, some valued models for their ability to reduce human bias and emotional influence, providing a more objective reference in uncertain or patient-expectation-driven situations. Models can help clarify cases and alleviate doubt, especially in complex or borderline situations. HCPs agreed that models should assist, not replace, shared decision-making, with the final decision made by both HCP and patient. Some noted that when HCPs disagree with a model’s suggestion, they might be expected to justify their reasoning, which could increase administrative workload and reduce autonomy. Lastly, they discussed the challenge of working with probabilistic information, recognizing that developing this skill requires time and experience. HCPs believed that prediction models would not fundamentally change their relationships with patients, emphasizing that personal interaction would remain central. Some anticipated a shift towards a more coaching role for certain patients, while others might still require close guidance. They noted that prediction models may not work well for everyone: older patients or those with limited health literacy could find technology-based tools challenging. The models were seen as potentially useful for communication, especially for explaining treatment options or supporting shared decision-making when results aligned with clinical judgement. However, communicating probabilities was viewed as difficult because many patients consider percentages abstract; visual aids were suggested to address this. Some HCPs also foresaw benefits for patients, such as improved clarity, reassurance and acceptance of treatment choices when uncertainty could be clearly presented. Anticipating consequences of prediction model use HCPs expected that prediction models might influence more than just treatment choices, potentially impacting how care is organized, medication use, professional autonomy and the overall nature of healthcare. HCPs voiced concerns that prediction models might be misused. They worried that scores meant to support clinical decisions could instead be repurposed as performance metrics for evaluating HCPs or institutions. For instance, it was mentioned that the DAS28 was occasionally misused in this way. External entities such as insurers or hospitals might apply pressure, particularly if the prediction results affect reimbursement or operational efficiency targets. Transparency about who develops models and their intended use was seen as crucial to maintaining trust and identifying conflicts of interest. Additionally, some expressed concern that the mere presence of a model could subtly coerce behaviour, making deviations seem risky or administratively difficult, even without formal mandates. HCPs believed that prediction models could enhance the efficiency of care by improving the planning of outpatient visits, reducing unnecessary consultations and automating routine tasks. This could help manage rising workloads and free up time for more complex cases. Models were also considered valuable for medication management, aiding taper decisions, making treatment choices and preventing unnecessary exposure to costly or ineffective treatments. However, HCPs also cautioned about unintended consequences: increased data collection and monitoring might add to, rather than reduce, workload, and fewer in-person visits could diminish patient visibility. Opinions on personalized medicine differed: some viewed prediction models as a move towards genuinely individualized treatment, whereas others feared that relying on them for every patient could result in more rigid treatment approaches and less personalized care. Discussion This study examined how rheumatology HCPs perceive the use of prediction models in their routine work. Their level of acceptance depends on their attitudes towards these models and innovation, their awareness of available tools, and how well they can interpret predictions within clinical reasoning. HCPs highlighted key performance and practical design factors they perceived as important for the use of prediction models, along with the anticipated effects on patient–doctor interactions and overall care quality. They expressed concerns about potential misuse and unintended consequences, emphasizing that HCPs viewed alignment with professional judgement and real-world clinical practice as important for the future integration of predictive models. Overall, HCPs perceived prediction models as potentially valuable tools, as long as they support, not override, professional judgement and add clear value to care. HCPs held differing opinions on prediction models regarding their potential disruption to workflow, their influence on personalized medicine, and the level of evidence regarding added value needed for their use. Our results align with established findings from studies on CDS systems, where workflow integration, transparency and demonstrated usefulness consistently serve as key factors in adoption [9, 17, 18]. The emphasis on interpretability and alignment with clinicians’ reasoning processes aligns with previous research demonstrating that confidence in predictive tools depends on how well predictions match established clinical reasoning and can be explained to patients [19–22]. Concerns about erosion of autonomy and the potential use of predictions for monitoring clinician behaviour have likewise been described in the broader clinical AI literature [23]. The present study offers a rheumatology-specific perspective on the use of prediction models in routine care. Characteristics of RA care, such as long-term follow-up and shared decision-making, may influence how HCPs evaluate new decision-support tools. Some HCPs expressed concerns about maintaining personalized care and clinical judgement, which may carry stronger weight in specialities where continuity and relational aspects are central. These findings suggest that, in rheumatology, prediction models may be perceived more favourably when they support the interpretive and relational elements of routine care. Earlier, our group conducted a qualitative study on the factors influencing patients’ acceptance of prediction models [24]. Several views were shared among patients and HCPs: both saw prediction models as support for clinical judgement rather than as a replacement. Patients expected HCPs to clearly explain the risks and benefits of using a prediction model. This aligns with our finding that HCPs need to understand how a model generates predictions. In contrast, patients placed more importance on maintaining the patient-doctor relationship and expressed concerns about the effects of less frequent hospital visits. While some HCPs in the current study mentioned this drawback, there was also emphasis on the time and cost-saving benefits, possibly through fewer consultations, guided by prediction models. When developing and implementing prediction models, both perspectives and interests should be considered. Strengths and weaknesses This study’s strengths include involving both rheumatologists and nurse practitioners with varying experience levels, attitudes towards innovations, and hospital settings, which enhances the richness of the findings. The use of iterative analysis and investigator triangulation improves credibility and depth of interpretation. Nonetheless, several limitations are noteworthy. HCPs’ were recruited from a limited number of Dutch institutions, possibly affecting the applicability of their views to other healthcare settings or countries. As with all qualitative research, the findings represent perceptions rather than actual behaviours, and responses during interviews might have been influenced by social desirability. Additionally, because prediction models are still developing in rheumatology, HCPs’ experiences were often anticipatory rather than grounded in real-world experience. Considerations for future development evaluation The findings of this study suggest several considerations for the future development, evaluation and implementation of prediction models in rheumatology. HCPs perceived that support for such tools may depend on whether they address a meaningful clinical need, fit within existing workflows and produce outputs that are understandable and actionable in daily care. Early involvement of HCPs and patients may help ensure that prediction models are developed for relevant problems and used in ways that are compatible with routine decision-making. A recurring issue in the interviews was the practical fit of prediction models in routine care. Models were viewed less favourably when they required additional administrative steps, duplicate data entry or substantial extra time to use. By contrast, integration with electronic health records and reliance on routinely collected data were seen as more supportive of future use. HCPs also emphasized the importance of interpretability, noting that clinicians may be more comfortable using prediction models when they understand what informs the output and how it relates to their own reasoning. In addition, they highlighted the challenge of communicating probabilistic information to patients, suggesting that how uncertainty is explained may be an important part of future implementation. The interviews also underscored the importance of evidence on added value beyond model development and validation. Although many prediction models have been developed in rheumatology, relatively few have been prospectively evaluated for their effect on clinical decisions or patient outcomes compared with usual care [25, 26]. In that context, clinical impact studies appear particularly relevant, while decision-analytic and modelling studies may offer early insight into potential benefits and cost-effectiveness [27]. In rheumatology, early examples include the flare-prediction research supporting the PATIO trial, in which a clinical prediction model is tested prospectively against standard care [5, 28]. Finally, HCPs’ concerns about professional autonomy and potential misuse suggest that future implementation efforts may benefit from greater transparency regarding a model’s purpose, limitations and intended use. Conclusion This study identified four overarching themes describing rheumatology healthcare professionals’ perceptions of the future use of prediction models in clinical practice: building support for prediction models, making them workable in routine care, positioning them within clinical judgement and patient care, and anticipating the consequences of their use. Across these themes, prediction models were generally viewed as potentially valuable tools when they were perceived as understandable, clinically relevant, practical and compatible with professional judgement and patient-centred care. At the same time, HCPs expressed concerns about possible misuse and unintended consequences. Together, these findings provide insights into healthcare professionals’ perceptions, expectations, and concerns regarding the use of prediction models in rheumatology practice. Supplementary Material Contributor Information Sina Fadaei, Department of Rheumatology & Clinical Immunology, University Medical Center Utrecht, Utrecht, The Netherlands. Patty de Groot, Department of Rheumatology, Erasmus University Medical Center, Rotterdam, The Netherlands. Rob R G Kuin, Department of Rheumatology & Clinical Immunology, University Medical Center Utrecht, Utrecht, The Netherlands. Julia Spierings, Department of Rheumatology & Clinical Immunology, University Medical Center Utrecht, Utrecht, The Netherlands. Amin Herman, Department of Rheumatology, St. Antonius Hospital Nieuwegein, Nieuwegein, The Netherlands. Paco M J Welsing, Department of Rheumatology & Clinical Immunology, University Medical Center Utrecht, Utrecht, The Netherlands. Lise M Verhoef, Department of Research, Sint Maartenskliniek, Ubbergen, The Netherlands. Supplementary data Supplementary data are available at Rheumatology online. Data availability Due to the sensitive nature of the research, supporting data are not available. Funding No specific funding was received from any bodies in the public, commercial or not-for-profit sectors to carry out the work described in this article. Disclosure statement: The authors have declared no conflicts of interest. References Associated Data Supplementary Materials Data Availability Statement Due to the sensitive nature of the research, supporting data are not available.