Abstract
Objective
The rapid development and widespread use of artificial intelligence (AI) in healthcare are reshaping medical services. However, influenced by the “double-edged sword” effect of AI, technical limitations and human-AI interaction uncertainties may trigger multidimensional patient safety risks. This study aims to analyze healthcare professionals’ risk perception of medical AI and to construct a relevant theoretical model, thereby providing scientific evidence and practical pathways to promote safe and efficient human-AI collaboration in clinical settings.
Methods
This study adopted a grounded theory approach, conducting semi-structured interviews with 18 healthcare professionals (e.g., physicians, nurses, administrators) from three tertiary hospitals in China between April and May 2025. Data were analyzed using NVivo 12.0, following open, axial, and selective coding processes to identify core categories.
Results
Medical AI elicits dual behavioral outcomes based on healthcare professionals’ benefit-risk perceptions. These perceptions are shaped by individual, technological, information dissemination, and organizational factors. Risk perceptions are structured across six dimensions: safety and privacy, technical efficacy, ethical and social, legal and liability, capacity development, and resource consumption. Among these, technical efficacy risks are directly related to patient safety and received the greatest attention (15/18, 83.33%).
Conclusions
Healthcare professionals’ risk perceptions of medical AI are dynamically constructed through clinical practice, organizational contexts, and technological evolution. The findings reveal a dynamic equilibrium between technological innovation and patient safety. Targeted optimization strategies should thus be implemented from the perspectives of technology developers, healthcare institutions, and policymakers to achieve balanced development between technological empowerment and risk control.
Keywords
Introduction
Artificial Intelligence (AI), defined as computer systems capable of performing tasks or making decisions that typically require human intelligence, 1 has achieved remarkable advancements in recent years and is reshaping the landscape of global healthcare services at an unprecedented pace.2,3 From the strategic deployment of “in-depth integration of new-generation information technology and healthcare” in the Healthy China 2030 Plan Outline 4 to the Guidelines on the Ethics and Governance of Artificial Intelligence for Health 5 issued by the World Health Organization (WHO), countries around the world have regarded medical AI as a core driver for optimizing resource allocation and enhancing diagnostic and treatment equity. Evidence has shown that medical AI offers significant advantages in improving diagnostic efficiency, optimizing treatment plans, and reducing preventable medical errors. 6 Currently, driven by multiple factors including policy support, technological innovation, and clinical needs, the clinical penetration rate of medical AI technology has significantly increased, covering key links such as imaging diagnosis, drug research and development, clinical decision-making, and health management. 2
Despite these benefits, medical AI also presents multidimensional patient safety risks, which constitute the core dilemma of its double-edged sword effect in healthcare. Recent reports from organizations such as ECRI and the Institute for Safe Medication Practices (ISMP) have identified insufficient governance of medical AI as a major emerging patient safety issue.7,8 AI-related patient safety risks refer to potential harms caused by inaccurate system outputs, including misdiagnosis, missed diagnosis, or inappropriate treatment recommendations that may cause direct or indirect harm—even life-threatening consequences.9,10 These risks stem not only from the inherent characteristics of AI technologies—such as hallucinations and algorithmic bias 11 —but are also closely intertwined with the uncertainty and complexity of human-AI interaction. In particular, AI-based clinical decision support systems (CDSS) may foster technological dependence and contribute to alert fatigue due to false positives, increasing the likelihood that genuine risks are overlooked.12,13 Public attitudes also reflect these concerns, with surveys indicating that 60% of U.S. adults feel discomfort with healthcare professionals relying on AI, while 75% worry that such technologies may be adopted prematurely. 14 In addition, ethical concerns and algorithmic biases related to race, socioeconomic status, and gender may further exacerbate existing healthcare disparities.15,16
Risk perception refers to the process through which individuals gather, interpret, and form subjective judgments about risk-related information.17,18 Early conceptualizations proposed by Bauer 17 emphasized the uncertainty of outcomes and severity of potential consequences from erroneous decisions. Subsequent research has expanded this perspective. The psychometric paradigm of risk perception developed by Slovic 19 suggests that individuals assess risks through cognitive and affective dimensions such as perceived controllability, dread, and familiarity. Similarly, the Extended Parallel Process Model (EPPM) proposes that behavioral responses to risk are shaped by the interaction between perceived threat and perceived efficacy, potentially resulting in protective engagement or risk avoidance. 20 These perspectives highlight that risk perception is not merely a cognitive evaluation of technological hazards but a complex psychological and social process influencing attitudes and behavioral decisions.
Previous studies have reported that individuals’ risk perception of medical AI may be influenced by education level, 21 AI literacy, 22 experience with AI applications, 23 task-technology fit, 24 and other factors. In the field of medical AI, many studies draw on established theoretical frameworks, including the Technology Acceptance Model (TAM) 25 and the Unified Theory of Acceptance and Use of Technology (UTAUT), 26 to examine the relationship between risk perception and technology adoption. Empirical studies have demonstrated that higher risk perception may reduce healthcare professionals’ satisfaction with and intention to use medical AI, and adversely influence their performance expectancy.23,24,27,28 Although these studies provide valuable insights into risk perception, they rely primarily on survey-based approaches, which cannot fully capture how healthcare professionals weigh and evaluate AI-related risks in real clinical contexts. In addition, measurement tools specifically designed to assess medical AI risk perception remain limited. Although several qualitative studies have begun to explore the facilitators and barriers to medical AI adoption, they often remain at a descriptive level and fail to delve sufficiently into the formation mechanisms and dynamic development of risk perception.29,30 Meanwhile, research on explainable AI (XAI) and AI trust suggests that limited explainability and algorithmic opacity may undermine healthcare professionals’ trust and hinder effective human-AI collaboration in clinical decision-making.31–33 Yet empirical evidence explaining how these perceptions are constructed within clinical settings remains limited.
Given these gaps, qualitative research methods are particularly valuable for capturing the contextualized experiences and cognitive processes underlying healthcare professionals’ risk perceptions of medical AI. Therefore, this study adopts a grounded theory approach to explore how healthcare professionals perceive and interpret AI-related risks based on their clinical experiences, allowing theoretical insights to emerge inductively from empirical data. By identifying the antecedents, dimensions, and behavioral outcomes of risk perception, this study aims to develop a theoretical model that contributes to understanding human-AI collaboration in healthcare and provides insights for promoting the safe and responsible integration of AI technologies in clinical practice.
Methods
Study design
A grounded theory methodology aims to develop an explanatory theory of experience grounded in the data. 34 This study adopted the procedural grounded theory methodology proposed by Corbin and Strauss 35 to identify the experiences and processes of healthcare professionals’ risk perception toward medical AI. Heavily influenced by pragmatism and symbolic interactionism, the procedural grounded theory emphasizes that individuals perceive phenomena and events, attribute meanings to them, and construct a meaningful reality through interactions in social contexts.35,36 This aligns closely with the objective of this study, which is grounded in the context of human-AI interaction in healthcare and seeks to explore the internal mechanisms of AI risk perception among healthcare professionals. Additionally, the procedural grounded theory features clear and structured coding steps, making it highly operational and enabling researchers to systematically generate theories from the data. This study was reported in accordance with the Consolidated Criteria for Reporting Qualitative Research (COREQ) checklist (Multimedia Appendix 1).
Participants
Semi-structured interviews were conducted with healthcare professionals at three large tertiary hospitals in Chongqing, China. All interviews were conducted between April and May 2025 using a combination of in-person and online formats. The study protocol was approved by the Ethics Review Committee of the affiliated institution (No. 2025-851-01) prior to the start of data collection. All participants provided informed consent, and all protocols were conducted in accordance with the Declaration of Helsinki.
In accordance with the key methods of grounded theory and the research objectives, this study adopted purposive sampling and theoretical sampling methods to recruit participants. 37 Interview participants were selected according to their substantial expertise and active engagement in medical AI and patient safety domains. Inclusion criteria included: (1) Licensed healthcare professionals, including physicians, nurses, medical technicians, and healthcare administrators; (2) Those who have been engaged in the profession for at least one year and have direct contact with medical AI or patient safety issues; (3) Clear thinking and strong verbal communication skills; (4) Voluntary participation in the study. Exclusion criteria include those who are on leave, on a mission outside, or undergoing further study and are not at their posts.
Sampling method
During the initial phase of this study, purposive sampling was employed to recruit participants. Following the principle of maximum variation sampling, semi-structured interviews were conducted with eligible healthcare professionals, ensuring diversity in characteristics such as gender, age, professional title, and work experience. Given that nurses and junior doctors, as frontline clinical practitioners, can provide firsthand insights into risk perception in human-AI collaboration within healthcare settings, this study positioned this group as the core component of the interview sample. Meanwhile, the study also included roles such as managers, medical technicians, and healthcare informaticists. This sample composition was designed to ensure diversity and complementarity of data sources, offering multidimensional empirical support for constructing a theoretical model of medical AI risk perception. The professional roles and organizational contexts of all participants were highly relevant to the medical AI risk issues addressed in this study, ensuring their strong grasp of the research questions.
Following the initial data collection and analysis, subsequent participants were recruited using theoretical sampling. Theoretical sampling refers to the process of sampling guided by the emerging properties or dimensions of concepts, with the aim of identifying their relationships and advancing theoretical elaboration. 35 In this study, the theoretical saturation was deemed achieved after three rounds of interviews spanning the initial, data enrichment, and theory formation phases, as the newly collected data could no longer reveal novel properties or dimensions of the categories, and clear conceptual linkages had been established within the theoretical framework.
Research team
All members of our research team have backgrounds in healthcare or nursing, and all have received training in qualitative research methods and have extensive practical experience. Two female authors, a doctoral candidate (A1) and a postdoctoral researcher (A7), conducted interviews with healthcare professionals with whom they had no prior relationship, taking notes throughout the interview process. A1 and two other authors (A2, A4) were responsible for coding the interviews.
Data collection
Interview topic guide used in this study.
Data analysis
This study utilized NVivo 12.0 software to organize the interview data, which were then analyzed in accordance with the three-level coding process of procedural grounded theory (open coding, axial coding, and selective coding) combined with the constant comparative method. Initially, researchers read through the complete interview transcripts to conduct line-by-line open coding, and extracted initial categories through constant comparison. Subsequently, axial coding was applied to classify the initial categories, integrating them into main categories through further comparison. Finally, selective coding was performed to clarify the relationships among the main categories, identify the highly abstract core category, and construct a theoretical model related to the risk perception of medical AI based on the connections between the core category, main categories, and initial categories. Figure 1 illustrates the research process of the grounded theory approach in this study. Grounded theory research process.
The data analysis process required iterative review of the original interview transcripts to systematically understand participant responses and was conducted concurrently with data collection. Following the principle of investigator triangulation, three researchers (A1, A2, A4) independently analyzed the data. They compared and discussed their respective coding results and refined categories, repeatedly validating them against the original data until a consensus was reached. Both the interviews and preliminary data analysis were conducted in Mandarin Chinese, with translation taking place only during the theoretical coding phase. To ensure translation accuracy, this study adopted a two-step translation and back-translation process, which was conducted separately by two native Chinese speakers with proficient English skills (A3, A5). Subsequently, the translated versions were cross-checked and harmonized through joint consultation. Finally, the translated texts were reviewed by a researcher (A6) with extensive clinical experience and overseas academic experience, to ensure that the linguistic style and cultural context of the translations were consistent with the original texts.
Rigor
This study employed multiple strategies to maximize its credibility, dependability, confirmability, and transferability. 38 To enhance credibility, the interview protocol was reviewed by experts with extensive experience in medical AI research. Furthermore, sufficient time was allocated for interviews, and data collection continued until the point of theoretical saturation was achieved. Dependability was strengthened through rigorous documentation and a transparent audit trail of the data analysis process. During interviews, inductive or implied questioning was avoided to minimize bias. To ensure confirmability and traceability, all key decisions, from raw data collection to final coding and theoretical model formation, were systematically documented, with participants’ feedback collected. Finally, to promote transferability, the research context was described in detail, facilitating the potential application of the findings to similar contexts in studies on AI risk perception.
Results
Demographic characteristics of study participants (n=18).
Results of data analysis using the grounded theory framework
The researchers employed inductive reasoning 35 to conduct open coding of the raw interview data from 18 participants. Through systematic analysis, discussion, comparison, and synthesis, 117 initial codes were identified. After repeated comparison and relational analysis with the original data, these codes were further consolidated into 35 sub-categories through conceptual grouping. Initial codes and their corresponding sub-categories are presented in Multimedia Appendix 2. Through multiple iterations of axial coding that clustered these sub-categories’ attributes while examining their internal logic and hierarchical relationships, 12 main categories were refined. For instance, during open coding, initial codes such as “competitive leakage information” (A1), “data theft” (A2), and “data leakage” (A3) were grouped together due to their shared theme of unauthorized data exposure. Through constant comparison, these were merged into the sub-category “privacy leakage risk.” Similarly, codes like “malware attack” (A4), “system defense vulnerability” (A5), and “hacker attack” (A6) were synthesized into the sub-category “system security risks.” In the subsequent axial coding phase, these two sub-categories were further integrated under the broader main category “safety and privacy risks,” as they collectively represent critical dimensions of security-related concerns in medical AI adoption. A final iteration of selective coding and constant comparative analysis led to the emergence of three core categories from the main categories: multidimensional drivers of risk perception, dimensional attributes of risk perception, and the dual behavioral effects of benefit-risk perception. The results of axial and selective coding are presented in Multimedia Appendix 3.
This study strictly adhered to the principle of theoretical saturation, with data collection and analysis conducted iteratively across three rounds of interviews.
37
In the initial interviews, the first nine participants contributed the vast majority of initial codes. The 10th and 11th participants yielded only a small number of new codes (3 and 1, respectively), and by the 12th participant, no new codes emerged, indicating that code saturation had been achieved. Building on this, interviews with the 13th to 16th participants were conducted. Through the merging, splitting, and renaming of existing codes, the category structure was further refined, resulting in clearer connotations and more distinct boundaries for both sub-categories and main categories. To further test the theoretical saturation of the findings, two additional interviews (the 17th and 18th participants) were conducted and subjected to a full round of three-level coding. The results showed that all data could be subsumed within the existing coding system, with no new concepts, dimensions, or relationships identified. In conclusion, after completing 18 interviews, this study reached the required level of theoretical saturation. Figure 2 presents an overall classification of identified sub-categories, main categories, and core categories through a sunburst chart, and marks the number of initial codes corresponding to each category. The size of each category area intuitively shows the difference in the number of identified codes contained therein. Sunburst chart of sub-categories, main categories and core categories based on grounded theory.
Construction of a theoretical model for healthcare professionals’ risk perception of medical AI
Construction of the theoretical model: Adaptations from APCO and TAM.
Based on the complementary advantages of the above two models, this study systematically analyzes the factors influencing healthcare professionals’ risk perception and their outcome effects, and constructs a three-stage theoretical model of “antecedent drivers of risk perception - dimensional attributes of risk perception - outcome effects of risk perception”, as shown in Figure 3. In this model, antecedent drivers at the individual, technological, informational, and organizational levels collectively shape healthcare professionals’ multidimensional risk perceptions of medical AI, including safety, ethical, cost, and other risks. These perceptions are then integrated into a dynamic benefit-risk trade-off mechanism, through which healthcare professionals continuously assess and weigh relevant factors to form an overall judgment. This, in turn, drives a spectrum of mutually convertible heterogeneous clinical behaviors, ranging from technological optimism to restrictive adoption. Thus, the theoretical model comprehensively reveals the inherent logical pathway through which medical AI risk perceptions are formed and subsequently influence behavior. A theoretical model of healthcare professionals’ risk perception of medical AI.
Core category 1: Dimensional attributes of risk perception
Core category 1 examines healthcare professionals’ risk perceptions regarding medical AI. A total of 70 codes were identified and categorized into 21 sub-categories, which were further consolidated into 6 main categories. Figure 4 presents a conceptual diagram of the dimensional attributes of healthcare professionals’ risk perception of medical AI, which helps to clarify the relationships and boundaries among them. This classification not only encompasses all 8 ethical principles outlined in the Question Bank for AI Risk Assessment (QB4AIRA),
40
but more importantly, through the three dimensions of technical implementation, ethical compliance, and social impact, it comprehensively reveals the core risk issues that need to be addressed in clinical applications of medical AI, providing both theoretical foundation and empirical support for subsequent risk prevention and control measures. Conceptual diagram of the dimensional attributes of healthcare professionals’ risk perception of medical AI.
Safety and privacy risks
Over half (10 of the 18 participants, 55.56%) of healthcare professionals express significant concerns about data security threats of medical AI, which mainly center on unauthorized access to and leakage of sensitive information. Such concerns include data exposure risks in active scenarios, such as inputting patient privacy or institutional core data into AI systems, as well as passive data breaches like hacking attacks. “If we input certain data into AI... there is a possibility that information could be leaked inadvertently, leading to data breaches.” (N4) “Honestly, my biggest worry is about patient privacy leaks. You know how it is these days - just give out your phone number once when shopping, and suddenly you're getting random calls from insurance agents, credit card scams... who knows where they got your info!” (N13) “If it's on the external network, your data can easily get leaked... Even if it's on the internal network, it... might still get hacked.” (N3)
It’s worth noting that the uniqueness of medical data amplifies the risks. Leaking patient privacy could not only lead to legal troubles but also result in misuse by commercial firms or competitors. “Hackers intentionally target a group... and use the relevant data... for the benefit of other organizations. That kind of thing is really scary.” (N11)
Additionally, system security risks and privacy leakage risk can create a cumulative effect. Technical defense gaps directly increase the likelihood of data breaches, while the high-value nature of medical data further induces more frequent cyberattacks, forming a vicious cycle of security dilemmas. “AI’s being used in healthcare in an embedded way right now… The data usage processes are unstandardized and irregular. And because these systems are built in a simplistic way, with not enough thought put into security, it’s really easy to end up with data leaks or other cybersecurity risks.” (N5) “AI adoption in healthcare is the big trend, but the core data here is incredibly valuable – that’s exactly why it becomes a target for cyberattacks.” (N7)
Technical efficacy risks
Regarding the technical efficacy of medical AI, a predominant majority (15/18, 83.33%) of healthcare professionals express concerns, such as those over diagnostic errors, judgment failures due to training blind spots, and discrepancies in human-machine decision-making. These issues are especially pronounced when dealing with rare cases or complex conditions, where AI systems are more likely to make mistakes due to insufficient data or model limitations, potentially leading to a cascade of erroneous diagnoses and treatments. “Yeah, we're using this AI for assisted diagnosis, but let's be real—it still messes up sometimes. I'm talking wrong diagnoses, missed cases... the whole deal.” (N13) “Let me give an example. We had a patient with a relatively rare condition, and according to a guideline we could access, the recommended blood pressure control target was around 120/60 mmHg. But the AI pulled up a more general hypertension control range, which didn't quite match our actual situation.” (N8) “When it comes to rare diseases, its diagnostic ability is still kinda weak—it gets easily thrown off by one-sided data.” (N15)
However, one participant remains optimistic about the application of AI in rare diseases, regarding it as a new solution to the challenges of diagnosis and treatment in this field. “Although our hospital is quite strong within the Chongqing area, on a national level, we still see relatively few rare disease cases… AI could help us access medical resources or data from other top-tier hospitals, and this is a really great direction for driving the development of our department and the hospital as a whole.” (N16)
The application of medical AI in clinical settings poses certain adaptability risks. When confronted with scenarios that require flexible responses, such as the rapid identification of critical illnesses and the formulation of personalized treatment plans, the system may exhibit a mechanistic processing pattern that fails to meet the complex demands of actual medical practice. “When it comes to the specific conditions and workflow of each hospital... an AI's response could easily be wrong in that situation.” (N5) “Cause clinical situations are really complicated, there are some things AI just can't replace human thinking for.” (N2)
Moreover, the content generated by medical AI may have three potential risks in terms of authenticity, accuracy, and relevance. The system may produce “hallucinatory” outputs such as fabricated content or grafted data, or it may generate misleading content due to the influence of polluted training data. “The first thing I came across was our VTE AI automatic scanning assessment, which required us to manually review it. During our use, we found that it was talking nonsense in a very serious manner [a colloquial expression meaning the AI provided absurd or unreasonable results while appearing formally authoritative].” (N9)
Ethical and social risks
Over half of the healthcare professionals (10/18, 55.56%) acknowledge ethical and social risks associated with medical AI. The ethical challenges potentially stemming from its application are centered on the diminished human agency in clinical decision-making. Meanwhile, the problem of algorithmic fairness is gradually coming to light. AI systems trained on specific data may generate implicit biases along dimensions such as race, gender, or age, leading to the risk of differential diagnosis and treatment. “We all know AI is an algorithm, but once this algorithm becomes so powerful that it can do anything... then people might be controlled by AI instead.” (N7)
The application of medical AI may exacerbate existing health inequalities. Individual differences in the acceptance of technology can lead to a digital divide, placing certain patient groups at a disadvantage. Meanwhile, the deployment of AI systems often tends to concentrate in resource-rich areas, which may further widen the gap in medical resources between regions. “Because each hospital has different procedures. Take a patient as an example—if I'm from a county-level hospital, the situation here is different from a big hospital. So, after AI goes live, this gap will only get wider.” (N6)
The use of AI in medical settings has sparked concerns over the weakening of humanistic care. Four healthcare professionals noted that AI interactions lack warmth and cannot replace the humanistic care based on empathy and emotional connection between doctors and patients. “Well... when it comes to nursing... it's really more about the human side of things, you know? Because AI... it just doesn't have that human touch... that warmth.” (N8)
Moreover, the introduction of medical AI may pose challenges to the traditional doctor-patient relationship model. AI’s involvement in the diagnostic process may lead to reduced communication between doctors and patients, decreased trust, and even conflicts arising from inconsistencies between AI recommendations and doctors’ judgments. “Maybe everyone will come to rely on AI so much that they become indifferent. Yeah, there might be less interaction with patients...” (N4) “A lot of patients are bringing AI-generated information to their doctors to check its accuracy. And that can lead to a situation where the standard treatment plan the doctor comes up with based on their experience doesn’t match what the AI says – these are all risks right there.” (N7)
The automation capabilities of AI have raised concerns about the potential displacement of certain medical positions, which may impact job security and professional value recognition. “I'm really worried radiology techs might lose their jobs over this. I mean, that AI can read like 100 scans in a second? Our whole team combined can't even match that in a day.” (N1) “If you let AI make all the decisions, then how are we medical staff supposed to show our worth?” (N7)
Legal and liability risks
Seven healthcare professionals (7/18, 38.89%) expressed concerns that the practical application of medical AI may give rise to a series of legal and liability risks. From a legal perspective, they raised worries that the use of existing medical knowledge or technical solutions by AI may involve intellectual property infringement; meanwhile, if AI generates false information, the identification and accountability of relevant responsible entities have become key issues that need to be clarified legally. “What if the stuff generated by AI is already patented or registered by someone else? Could we be infringing on their intellectual property?” (N18) “We know that sometimes AI hallucinates and gives false information. If that misleads our clinical decisions, could we be held legally responsible? I mean, would we be on the hook for that?” (N9)
Currently, the application of medical AI faces a significant dilemma in determining liability. Throughout the chain from technological development to clinical application, there is a blurring of boundaries among the responsibilities of multiple parties, including developers, hospitals, and healthcare professionals. This issue is particularly pronounced when AI-driven decisions result in medical errors, making it difficult to distinguish between technical flaws and human negligence. “If this AI really makes a mistake, it's probably unclear right now who should be held responsible for the liability assessment - the manufacturer, the doctor, or someone else?” (N12) “When using AI to assist in healthcare, if the technology isn’t reliable and it becomes a standard practice, then when something goes wrong, who’s to blame? The doctor? Or someone else? It’s about how responsibility should be assigned.” (N13)
More seriously, the existing regulatory framework lags to some extent behind the pace of technological advancement. The clinical application of AI may be in a risky state of “practice first, regulate later,” which is not conducive to the protection of patients’ rights and interests and also increases the compliance risks for medical institutions. “Right now, how to regulate AI—I think because it's such a new thing and developing so fast—I doubt even the national regulatory bodies have solid measures in place to handle it properly.” (N7)
Capacity development risks
Half of the healthcare professionals (9/18, 50.00%) mentioned that AI may impede the development of individual capabilities. The widespread use of medical AI is reshaping the professional capability structure of healthcare workers. Some healthcare professionals exhibit a tendency toward excessive reliance on AI systems, including blind trust in AI outputs, neglect of independent clinical judgment, and weakened risk prevention awareness. “My first concern is that people might become overly reliant on it. I do think AI has its strengths, but it definitely has its shortcomings too.” (N8) “I have a friend who's really dependent on and trusts AI now. Whenever we run into any kind of problem, she has to first take a look at what AI has to say about it. She doesn't even bother to think about whether the AI's output is correct or not anymore.” (N17)
Meanwhile, the standardized outputs of AI may trigger dual risks of cognitive thinking rigidity and professional competence weakening among healthcare professionals. Long-term dependence on AI could lead to formulaic clinical thinking patterns, undermining independent judgment and innovative capabilities. More critically, with foundational tasks replaced by AI, healthcare professionals may experience reduced practical opportunities, causing a gradual decline in core professional competencies and forming a vicious cycle of “technological dependence-competence weakening.” “When it comes to clinical applications of AI, I'm mainly worried that people will form a dependency—like, they stop thinking... Anytime there's a problem, they immediately input it into AI without checking if it's correct. Over time, everyone might lose a lot of ability to think independently...” (N4) “So like... if 90% of the people training the AI are just... average, y'know? Doesn't that mean all the answers we get might just stay stuck in this little bubble? Like, how are we ever gonna get some real groundbreaking innovation if everyone feeding the system is just... meh?” (N1) “Here's an issue – if we rely too heavily on AI, it might cause our clinical doctors' diagnostic and treatment capabilities to decline. After all, medicine is an empirical science that also needs to keep evolving. But if we over-rely on this kind of AI, will our own abilities gradually diminish?” (N16)
Additionally, the introduction of AI has imposed new requirements for human-AI collaboration on healthcare professionals, including the ability to discriminate AI outputs and adapt to standardized work processes, which in effect raises the threshold for professional competency. “Now that AI is being introduced into clinical practice, I'm also worried about whether it will change the... well, the workflow we're already used to. It means we'll have to adapt.” (N3) “Sometimes you can actually catch the AI faking it – like, straight-up making up data. Honestly, I think that's where our own judgment really comes into play.” (N16)
Resource consumption risks
Seven healthcare professionals (7/18, 38.89%) indicated that the deployment and application of medical AI may face certain resource consumption risks. In terms of financial investment, localized deployment costs and competitive investments among hospitals may lead to resource waste, and even issues such as the transfer of costs to patients, thereby exacerbating the healthcare burden. “Then there's the funding issue. For medical AI, you have to deploy it locally, and that kind of setup costs an arm and a leg.” (N7) “It's about the charging issue. If the fees are too high, people might as well just look things up in books themselves.” (N13) “Every hospital has to invest in buying software, and that cost will eventually be passed on to patients, which means the cost of medical services will also go up.” (N5)
Meanwhile, the application of medical AI may also give rise to new time efficiency issues. Healthcare professionals need to invest time in adapting to new tools, and the operational time consumption and efficiency performance of some AI systems may be inferior to traditional manual processes, potentially hindering overall medical efficiency. “Take us emergency triage nurses for example—sometimes before a patient even gets to speak, I can quickly tell what's going on with them... No need for too many questions. The AI is just slower than us in that situation.” (N2) “Then there’s its processing speed. If you use it and wait ages for the results to come out, that kind of efficiency just won’t cut it, right?” (N13) “Before I can use this AI tool, I need to spend time learning and getting used to it. So at first, it might actually slow me down instead of helping.” (N18)
On top of the financial and time-efficiency burdens, the deployment of medical AI may also give rise to the risk of cost-benefit reversal in hospitals and increase the complexity of clinical workflows, thereby constituting a notable risk of operational burden. Specifically, the manual review of AI-generated outputs not only adds extra procedural steps but also may lead to treatment delays. “The tricky part is, clinicians might think, ‘If I'm the one making the final call anyway, what's the point of this AI?’ But especially when we're just starting to use it, we really don't feel confident relying on it completely. We still need a human to double-check everything.” (N7) “We can triage patients directly to the resuscitation room when we need to. But if we rely on AI completely, it will just stick to the system’s checklist of questions during triage – and that could delay the treatment of critically ill patients.” (N2)
Core category 2: Multidimensional drivers of risk perception
Core category 2 analyzes the multidimensional factors influencing healthcare professionals’ risk perception of medical AI. A total of 35 codes were identified, categorized into 10 sub-categories, and further consolidated into 4 main categories: individual agency factors, technical characteristic factors, information dissemination factors, and organizational environment factors.
Individual agency factors
Drawing on the perspectives of eleven healthcare professionals (11/18, 61.11%), this study finds that individual agency factors shape their risk perception regarding medical AI. The level of understanding of AI, relevant knowledge reserves, and personal awareness of risk identification among healthcare professionals may affect their risk perception ability. Individuals with higher information literacy tend to evaluate the potential risks of the technology more rationally. “How risky it is really depending on how much people understand AI. Some folks have never even used AI—they might think it’s some kind of miracle tool. Then one day, they realize, ‘Wow, this thing (AI) actually works really well… maybe too well.’ If they start relying on it for everything, even thinking it could replace them, that’s when the risk gets really high.” (N5) “If you haven't worked on medical AI research and you encounter this kind of tool for the first time, you'd probably be blown away by its capabilities. You'd be really excited about it. But for those of us who've been researching this for years, we also see the flaws. Even though the technology is advancing fast, we're still not 100% confident in it—we're aware of the risks.” (N12)
Long-term accumulated professional experience, practical experiences, and professional orientations may shape unique risk perception frameworks, leading to differences in the acceptance of and concerns about AI among healthcare professionals with different professional backgrounds. “Well, it comes down to experience... For a general, ordinary case, I'd probably just go by my own experience.” (N2) “I think it really comes down to experience. When you actually use medical AI—like, you follow its recommendations, you try its suggested approach—and you see the results? That’s when you start trusting it.” (N11)
Individual self-efficacy, including confidence in technology, sense of control, and adaptability, also implicitly moderates their sensitivity to risks. “You know, using this medical AI thing is actually pretty easy for us. It's got a set way of doing things - like a clear step-by-step process. Honestly, anyone could learn to use it if they tried.” (N14)
Technical characteristic factors
Six healthcare professionals (6/18, 33.33%) indicated that technical characteristics may influence their risk perception of medical AI. The interpretability of technology is a key factor for healthcare professionals to assess AI risks, including the black-box effect of algorithms, theoretical foundations, and transparency. When the decision-making logic of an AI system is difficult to understand, healthcare professionals often have doubts about its reliability and safety, thereby enhancing risk perception. “One of the reasons people feel anxious is that it's just... unclear. It's like buying a phone: you buy one for 3,000 yuan, I buy one for 9,000 yuan, but in reality, you don't even use 10% of its features before it becomes obsolete.” (N1) “Take the technical side for example - with all these algorithms and models... See, most AI models today have this 'black box' effect going on. Like, you can't really fully explain how they work, or make them completely transparent, you know?” (N12)
Healthcare professionals believe that medical AI needs to define clear boundaries of technical applicability, such as application scenarios and service positioning. If the scope of AI’s applicability is ambiguous or exceeds its actual capabilities, healthcare professionals may be more alert to the risks of technical misuse or failure. “Even among doctors and nurses in the same clinical setting, the AI tools they use and the scenarios where they apply AI are different. So we really need to distinguish—like, draw the line—based on the application scenario.” (N5) “Like during training, it's crucial to define what it CAN and CAN'T do - y'know, is it just assisting doctors, or meant for something else? We gotta clearly outline its scope. Because if you start using it for everything without boundaries, things will get messy really quick.” (N13)
Information dissemination factors
Nine healthcare professionals (9/18, 50.00%) suggested that information dissemination factors could influence their risk perception of medical AI, primarily manifested in two aspects: social interpersonal communication and institutionalized communication. Social interpersonal communication can shape healthcare professionals’ cognition through diverse channels, including online discussions, peer reputation, interdisciplinary exchanges, etc. These informal information sources often influence individuals’ risk perception of AI technology in a subtle way through forms of experience sharing or opinion exchange. “So, you use it, I use it, everyone uses it. Actually, people are all talking about it (AI) by word of mouth, right?” (N1) “The main way people use it is through online services. Honestly, because the online world is everywhere these days.” (N17)
In contrast, institutionalized communication defines the legitimacy and reliability of AI technology in a more systematic manner through authoritative channels such as national policy advocacy, official media reports, and professional literature. “Honestly, whichever one the government approves - that's the one we'll go with. It's not like we're in any position to judge this stuff ourselves.” (N1) “You know, I think info from proper academic sources matters too—like peer-reviewed journals or official conference reports. When legit platforms publish these findings, it helps me cross-check and assess the risks better before using AI.” (N15)
Organizational environmental factors
Organizational environmental factors may play a critical role in shaping risk perception toward medical AI among healthcare professionals (5/18, 27.78%). The position-power structure influences healthcare professionals’ decision-making participation in AI technology and their cognition of risk responsibility attribution through work duties, responsibility division systems, and other mechanisms. “Here's the thing—if something goes wrong, who do we turn to? Who can actually fix it? If we just dive in headfirst and roll this medical AI out everywhere, then suddenly there's an emergency and... crickets. No one steps up to help.” (N7)
Workload pressure moderates risk perception from the perspectives of task volume and complexity. Healthcare professionals under overload may be more inclined to accept AI to alleviate their burdens, but they may also overlook potential technical risks due to time constraints. “You know what? I think doctors are just so swamped with work that they often end up blindly trusting AI. And honestly, that's kinda worrying.” (N14)
Organizational technical support provides capacity assurance for healthcare professionals through creating an innovative atmosphere and systematic training. Good organizational support can significantly reduce anxiety about technology use, while the lack of it may amplify concerns about risks. “Actually, our hospital's been pretty proactive about this—I've personally attended at least two AI training sessions they organized. Honestly, I think that's one of the big reasons why we're more open to using it now.” (N16)
Core category 3: The dual behavioral effects of benefit-risk perception
The differences in healthcare professionals’ perception of the benefits and risks of medical AI can lead to vastly different behavioral patterns. Core category 3 analyzes the specific impacts of such perceptual differences on the behaviors of healthcare professionals. A total of 12 codes were identified, categorized into 4 sub-categories, and further consolidated into 2 main categories: perceived benefit advantages and perceived risk advantages.
Perceived benefit advantages
Five healthcare professionals (5/18, 27.78%) indicated that when perceived benefits dominate, they may exhibit dual psychological and behavioral responses. On the one hand, it manifests as risk compensation complacency, including neglect of potential risks, reduced verification, and technological dependency. Such blind trust may pose a threat to medical safety. “My main concerns are probably about errors and risk identification. Maybe at first, we trust it 80%, then it goes up to 90%, 100%. Eventually, I might stop thinking for myself and just trust it unconditionally, which would lower my awareness of risk prevention.” (N12) “Here's the thing—after using AI for a while, if its feedback consistently matches my own experience and verification, and it proves reliable... well, I'd probably start trusting it completely. Before you know it, I'd be double-checking less on specialized questions—both in research and clinical practice.” (N15)
On the other hand, it gives rise to technological optimism behaviors, such as proactive adoption and active promotion of AI technologies. Although this is conducive to the diffusion of innovation, potential hazards caused by irrational optimism need to be addressed vigilantly. “People will only push for its widespread use if everyone thinks it's actually good. They need to try it first and have a positive experience with it. Take me for example—if I see it's cost-effective and the benefits far outweigh the risks, then absolutely, everyone's gonna want to use it.” (N13)
Perceived risk advantages
Seven healthcare professionals (7/18, 38.89%) indicated that when risk perception dominates, corresponding reactive measures are likely to be adopted. Specifically, they tend to implement more cautious coping strategies, which typically involve defensive measures such as cross-validation using multiple AI systems, seeking alternative solutions, and strengthening professional reviews. These behaviors reflect a rational, risk-averse professionalism. “I don't trust it, but I'm willing to give it a try... If there are aspects where it might not fit or even pose risks, then in those cases, I'll try to avoid them, use it as little as possible, or find alternative solutions.” (N5) “For example, I’ve got tons of AI tools on my phone right now. If I don’t trust one of them, or if I think its answers aren’t all that accurate, I’ll just check out other AI tools instead.” (N10)
Meanwhile, restrictive adoption behaviors, characterized by conservative tendencies such as technological avoidance, usage anxiety, and cautious participation, may reduce risks but could hinder the full realization of technological innovation value. “For example, if the perceived benefits and risks of AI are both 50%, I definitely won't use it. I'll only use it if the benefits far outweigh the risks...” (N13) “For healthcare workers who don't have a high level of trust in AI, I don’t believe they would simply stop using it. In fact, they'd probably use it in a more targeted and cautious way.” (N15)
Discussion
Principal findings
This study employed the grounded theory methodology proposed by Corbin and Strauss 35 to explore how healthcare professionals perceive risks associated with medical AI in clinical settings. Based on interview data from healthcare professionals working in tertiary hospitals in urban China, the study identifies a structured process through which risk perceptions are formed and translated into behavioral responses toward medical AI. This study makes four key contributions to the field. First, it reveals that healthcare professionals’ risk perceptions of AI are shaped not only by individual agency factors but also by the interplay of AI technical characteristics, information dissemination mechanisms, and organizational environmental influences. Second, it identifies six key dimensions of medical AI risks and establishes a prioritized ranking based on the level of concern among healthcare professionals. Third, the findings indicate that risk perceptions do not lead simply to acceptance or resistance; rather, they give rise to a dual behavioral orientation characterized by a dynamic balance between perceived benefits and perceived risks. Finally, integrating these insights, the study proposes a theoretical model of “antecedent driving factors-dimensional attributes-consequent effects,” which explains how healthcare professionals interpret AI-related risks and translate these perceptions into clinical behavioral responses. Notably, these findings reflect the experiences of healthcare professionals working in tertiary hospitals in Chongqing, China—settings characterized by relatively advanced digital infrastructure and increasing institutional support for AI implementation. As such, the results provide valuable insights into the risk perception processes of AI in high-resource clinical settings.
Multidimensional driving factors analysis
A consensus has been reached that healthcare professionals face uncertainties and risks in adopting medical AI technologies. 41 Building on this, the present study reveals that healthcare professionals’ risk perceptions of AI are shaped not only by the objective characteristics of the technology but also by their cognitive frameworks and the social information environment. Existing research highlights the central role of trust in shaping technological risk perception. A theoretical model of risk perception regarding genetic technologies indicates that trust influences how individuals interpret risk, which in turn affects their acceptance of emerging technologies. 42 In the context of AI systems, trust is closely associated with system transparency and interpretability. 31 As noted by Miller, 43 interpretability should be regarded as a prerequisite for achieving fair and trustworthy AI. Therefore, XAI emphasizes dimensions such as transparency, accountability, security, and fairness,44,45 which are essential for users to understand and critically evaluate the algorithmic decision-making process.46–48 Our findings further suggest that explainability is a key technological antecedent in shaping healthcare professionals’ risk perception of AI, thereby complementing existing frameworks on AI trust and human-AI interaction.
At the individual level, factors such as AI information literacy and professional capital accumulation together constitute the complex cognitive mechanisms through which healthcare professionals respond to medical AI. Previous studies have shown that significant professional differences among healthcare professionals in different departments, especially between medical technologists and clinicians, may lead to substantially different perspectives on new technologies. 49 Experiences with mHealth services may reduce healthcare professionals’ risk perception of medical AI. 50 This study also finds that social information environments such as organizational innovation climate can influence healthcare professionals’ risk perception of AI—innovative healthcare professionals tend to better perceive the ease of use of AI rather than merely focusing on potential risks. 51 The above findings echo the Almere model, 52 which integrates theoretical frameworks such as TAM 25 to comprehensively demonstrate how factors including the functional characteristics of medical AI, human-machine trust, and environmental characteristics collectively influence technology adoption behavior. Taken together, these results suggest that risk perception toward medical AI emerges from the interaction between technological characteristics, individual cognition, and the broader socio-organizational context in which AI is embedded. Understanding this interaction is essential for designing governance strategies that support safe and effective human-AI collaboration in clinical practice.
The dimensional attributes of medical AI risks
Despite the release of the Guidelines for Reference Scenarios of Artificial Intelligence Applications in Healthcare 53 by the National Health Commission of China at the end of 2024, which has facilitated the deeper integration of AI into medical settings, a series of risks, including technical reliability, data security, and ethical concerns, has gradually emerged. We find that healthcare professionals’ perception of the multidimensional risks of medical AI exhibits a priority structure centered on clinical consequences and responsibility. Specifically, risks related to technical efficacy, due to their direct impact on diagnostic accuracy and patient safety, have triggered concern among the vast majority of healthcare professionals (15/18, 83.33%). Following these were safety and privacy risks, along with ethical and social risks (both 55.56%, 10/18), which mainly affect the establishment of trust and regulatory compliance during clinical application. In contrast, risks related to competence development, legal liability, and resource consumption were perceived as having long-term implications for organizational adaptation and sustainability. These findings provide a theoretical basis for the scenario-based governance of AI and hierarchical risk communication. In advancing the clinical implementation of medical AI, priority should be given to addressing core efficacy risks directly related to patient safety, followed by a systematic response to institutional risks such as ethical and privacy issues, so as to form a well-tiered, context-adapted pathway for risk mitigation.
Consistent with prior studies, the greatest risks associated with medical AI lie in potential harm to patient rights.47,54 These risks primarily fall into three categories: data sensitivity, irreversibility of outcomes, and complexity of liability for subjects. First, regarding safety and privacy risks, models require access to and processing of massive sensitive information during training, including clinical records, molecular characteristics, and demographic data, which to some extent increases the risk of data breaches. Second, AI systems may generate “hallucinations,” producing plausible yet inaccurate outputs in complex clinical scenarios, which are often difficult to detect and may adversely affect patient treatment. 55 Finally, if AI-involved decision-making causes a medical accident, the liability subjects may include model developers, deployers, users, and other parties, making it difficult to define the degree of responsibility.
Regarding ethical risks, the reliance of AI models on real-world medical data can embed and amplify biases related to gender, race, and other social factors, potentially leading to inequitable outcomes. A comparative study by Stanford Medical Center showed that LLMs tend to underestimate the pain level of Black patients during pain assessment, demonstrating racial bias. 56 Such biases not only undermine health equity for patients but may also compromise the accuracy and authenticity of clinical diagnosis and treatment. Additionally, some healthcare professionals expressed concerns about professional replacement; however, AI is not intended to replace healthcare professionals; rather, those who use AI effectively may replace those who do not. 57
The double-edged sword effect of risk perception
This study indicates that healthcare professionals’ attitudinal responses to medical AI reflect a dynamic balance between perceived benefits and perceived risks. This pattern can be jointly explained by the Valence framework 58 and the EPPM. 20 The former emphasizes the cognitive basis of benefit-risk trade-offs, while the latter further elucidates how these evaluations are translated into behavioral responses through the interaction of perceived threat and perceived efficacy, thereby jointly shaping the polarized behavioral responses in this study. When perceived benefits dominate, healthcare professionals’ technological optimism may accelerate innovation diffusion but also induce risk compensation complacency. From the perspective of patient outcomes, while such benefit-driven adoption can enhance healthcare accessibility and continuity, 59 the accompanying technological dependence or risk neglect may lead to long-term accumulation of systematic errors, compromising the quality of care and patient safety. Conversely, when risk perceptions prevail, defensive strategies—such as repeated AI verification—can reduce error rates but may also result in excessive caution or technology avoidance, limiting gains in diagnostic and treatment efficiency and precision, ultimately impacting recovery processes and health outcomes. Empirical studies support this mechanism. Szabó et al. 60 found that respondents’ risk perceptions were negatively correlated with their trust in and willingness to accept robotic surgery. Arfi et al. 61 also confirmed through structural equation modeling that the risk-trust relationship influences the intention to use the Internet of Things in healthcare settings. These findings echo the behavioral polarization identified in this study and suggest that traditional binary technology-acceptance framework should be integrated with a dynamic benefit-risk balance mechanism, particularly considering the inherent safety-efficiency tension in medical AI.
To address this imbalance, healthcare institutions should design targeted intervention strategies while comprehensively guiding healthcare professionals in the use of AI devices and enhancing their technological literacy. At the training level, hierarchical and differentiated risk education courses should be developed for personnel with diverse professional backgrounds, incorporating case-based reflection and scenario-based exercises to enhance awareness of AI limitations and prevent misuse or inappropriate application. At the management level, institutions should establish a multidisciplinary review mechanism for in-hospital AI applications, clarifying human-AI accountability boundaries to alleviate adoption anxiety while ensuring safe exploration. At the policy level, regulators should promote industry-wide frameworks for error tolerance and accountability. Drawing on the European Union (EU)’s Artificial Intelligence Act, 62 which defines liability principles through accountability lists, regulators should refine the tripartite liability allocation for medical AI to support safe clinical implementation. The ultimate goal is to guide healthcare professionals in developing a dialectical and balanced technological cognition of AI, enabling them to embrace the efficiency gains brought by such technologies with an open mindset while consistently maintaining patient safety-centered professional prudence.
We argue that healthcare professionals’ risk perception is not a fetter but a cornerstone for the sustainable development of technology. Given the specificity of the healthcare industry, it is essential to deeply integrate human-AI alignment principles into the entire lifecycle of medical AI, ensuring that AI’s objectives, behaviors, and outputs are consistent with human values and social norms. 54 At the technical level, addressing algorithmic transparency must be a priority. Developing XAI tools can assist healthcare providers and patients in understanding decision-making logic. Currently, the Artificial Intelligence Act 62 mandates that medical AI systems provide technical documentation and transparency information, thereby upgrading interpretability from a technical option to a compliance standard—an approach worthy of reference. At the ethical level, an “ethical-clinical dual verification” framework should be adopted, integrating medical ethical principles into model training while incorporating real-time self-check mechanisms to ensure outputs comply with both ethical norms and clinical needs. While derived from tertiary hospitals in urban China, the identified multidimensional risk structure and benefit-risk balancing mechanism may offer analytical insights for other healthcare contexts. However, their applicability depends on contextual factors such as resource availability, the level of AI implementation, and regulatory environments.
Limitations
This study has several limitations. First, participants were recruited from tertiary hospitals in the same region with relatively resource-rich environments and growing AI deployment. Therefore, the findings mainly reflect experiences of healthcare professionals in urban tertiary hospitals and may differ in primary-care or rural healthcare settings, or in non-Chinese healthcare systems with different AI governance frameworks. Second, interview-based research may be subject to social desirability bias, particularly when discussing emerging technologies that are strongly supported by institutional policies. Third, this study has an absence of triangulation with observational or document-based data, which may limit deeper understanding of how AI technologies are used in clinical workflows. Finally, participants’ prior exposure to AI varied across departments and professional roles. While this enhanced the diversity of perspectives, it may also have influenced the nature and intensity of risk perceptions. Future research could expand the scope of investigation by including institutions across different regions and resource levels, as well as by integrating mixed-methods approaches combining interviews, observations, and quantitative surveys to further validate and extend the proposed theoretical model.
Conclusions
The risk perception of medical AI is a complex construct shaped by both technological attributes and social cognition. Utilizing grounded theory methodology, this study uncovers the multidimensional driving mechanisms, six risk dimensions, and dual behavioral effects that shape healthcare professionals’ risk perceptions of medical AI in tertiary hospitals in urban China, and develops a corresponding three-stage theoretical model. This addresses a key limitation in existing research, which has predominantly focused on quantitative prediction while overlooking the dynamic perception process and value trade-offs, thereby enriching qualitative insights in this field. The proposed framework also offers analytical value for understanding AI risk perception in other contexts, although specific risk priorities and responses may vary. Building on these findings, the governance of medical AI should move beyond a binary technology-acceptance paradigm toward a synergistic technology-risk-humanity framework guided by the hierarchy of clinical risks. This entails enhancing interpretability and ethical integration at the technological design level; establishing tiered training, multidisciplinary review, and fault-tolerant accountability mechanisms at the organizational level; and developing industry-wide accountability frameworks and transparency standards at the policy level. Future research should further examine variations in risk perception across diverse healthcare systems to improve the transferability of these insights and promote interdisciplinary collaboration to translate them into actionable clinical governance tools.
Supplemental material
Supplemental material - How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes
Supplemental material for How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes by Haoning Shi, Huan Wang, Shuangjiang Zheng, Wenna Xiao, Mingyuan Ju, Qinghua Zhao, and Huanhuan Huang in Digital Health.
Supplemental material
Supplemental material - How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes
Supplemental material for How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes by Haoning Shi, Huan Wang, Shuangjiang Zheng, Wenna Xiao, Mingyuan Ju, Qinghua Zhao, and Huanhuan Huang in Digital Health.
Supplemental material
Supplemental material - How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes
Supplemental material for How healthcare professionals perceive artificial intelligence risks: A grounded theory exploration of antecedents, dimensions, and outcomes by Haoning Shi, Huan Wang, Shuangjiang Zheng, Wenna Xiao, Mingyuan Ju, Qinghua Zhao, and Huanhuan Huang in Digital Health.
Footnotes
Acknowledgements
We would like to express our gratitude to our participants for sharing their time with us and allowing us to learn from their experiences. Without their contribution, our research would not have been possible.
Ethical considerations
This study protocol was approved by the Ethics Review Committee of the First Affiliated Hospital of Chongqing Medical University prior to the start of data collection. Respondents gave signed informed consent before starting the interviews.
Author contributions
HS contributed conceptualization, methodology, writing and revising. HW contributed data collection, data analysis, and writing. SZ contributed software, writing, and visualization. WX and MJ contributed data collection and writing. QZ contributed conceptualization, and revising the paper. HH contributed conceptualization, data collection, and revising the paper. All authors critically reviewed and revised the initial draft. All authors have read and approved the final version of the manuscript.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and publication of this article: This work was supported by the General Program of Chongqing Natural Science Foundation (project No. CSTB2025NSCQ-GPX1118) and the Key Nursing Research and Innovation Project of the First Affiliated Hospital of Chongqing Medical University (project No. HLPY2025-04).
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
Research data is not shared in compliance with ethics measures in place to protect participant confidentiality and privacy.
Guarantor
HH.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
