This study used a qualitative design using semistructured interviews to explore barriers and facilitators to the implementation of AI in ICUs in China from the perspectives of ICU nurse managers. The CFIR was used as the guiding theoretical framework to inform the overall study design. Specifically, CFIR domains and constructs guided the development of the interview guide to ensure comprehensive coverage of multilevel implementation determinants and further provided an analytic lens for organizing and interpreting findings. The study focused on identifying perceived facilitators and barriers across CFIR domains, thereby generating a structured understanding of determinants influencing AI implementation in Chinese ICU settings. This study’s reporting followed the rules for COREQ (Consolidated Criteria for Reporting Qualitative Research; Checklist 1 [23]).

This study was conducted in ICUs of tertiary hospitals in China. To capture diverse implementation contexts across China, we adopted a maximum variation purposive sampling strategy [24]. According to the standard regional divisions commonly used in China, the country was stratified into 7 regions: Northeast, North China, East China, Central China, South China, Southwest, and Northwest. Within each region, ICU nurse managers from tertiary hospitals were invited to participate in order to capture diverse perspectives on the implementation of AI in ICU practice, including perceived barriers and facilitators across different organizational contexts.

Eligibility requirements for participants were as follows: (1) currently serving as ICU nurse managers with responsibilities for unit-level nursing administration and oversight of ICU workflows and management practices; (2) at least 5 years of ICU management experience; and (3) involvement in or knowledge of the implementation of AI or information systems within their units. Based on these criteria, the research team intentionally selected eligible nurse managers from tertiary hospitals in 7 geographic regions of China as interview participants. Potential participants were approached via WeChat (Tencent) or by phone. During the initial contact, researchers briefly explained the study’s objectives, procedures, and the voluntary nature of participation. Those who agreed to participate were subsequently scheduled for a face-to-face interview at a time and location convenient for them. Before each interview, the researcher provided additional study information, answered participants’ questions, and obtained written informed consent. Participants were assigned unique identifiers to protect confidentiality.

The sample size was determined through an iterative assessment of data saturation during data collection and preliminary analysis [25]. Data collection and coding were conducted concurrently. After each round of interviews, 2 researchers (WD and YX) independently reviewed the transcripts, preliminary codes, and CFIR construct mapping to determine whether new information continued to emerge. The research team maintained a saturation tracking grid throughout the interviews. After the ninth interview, only minor refinements to existing codes were identified, and no new CFIR constructs emerged. The tenth and eleventh interviews confirmed previously identified barriers and facilitators without generating substantially new codes, items, or constructs. Recruitment ceased after 11 interviews when the research team agreed that data saturation had been achieved.

A semistructured interview guide was created, with all questions revolving around the CFIR. The questions focused on current practices of AI in ICUs, attitudes toward AI, and its barriers and facilitators (Multimedia Appendix 1). This interview guide was reviewed by qualitative research experts and pretested with 3 ICU head nurses, followed by revisions to ensure clarity and relevance. Data were collected between August and October 2025. Interviews were conducted by trained researchers (WD and YX) with experience in qualitative methods and clinical nursing research. All interviews were conducted face-to-face in a quiet and private room. No prior relationship existed between the interviewers and participants before data collection. With participants’ consent, all interviews were audio-recorded to ensure accurate data capture. All recordings were transcribed verbatim in Chinese shortly after the interviews. Data analysis was performed using the original Chinese transcripts. Selected excerpts were translated into English after data analysis by a bilingual researcher and independently checked by a second bilingual researcher. Discrepancies were resolved through discussion to ensure accuracy and conceptual equivalence. Selected excerpts were back-translated to verify fidelity of meaning. To address potential bias, interviewers engaged in reflexive practices, including maintaining reflective journals.

The data were analyzed using directed content analysis [26,27], combining data-driven inductive coding with theory-driven deductive interpretation guided by the CFIR framework. Researchers repeatedly reviewed transcripts to familiarize themselves with the data, extracting meaningful text units, such as sentences, paragraphs, and words for coding through inductive methods. Similar codes were subsequently grouped into items. These inductively derived items were deductively mapped onto the domains and constructs of the CFIR framework. For example, a participant statement, such as “Third, it reduces labor costs by freeing nurses from repetitive, mechanical tasks,” was initially coded and grouped as the “Mitigating labor shortages” item and subsequently mapped to the CFIR construct relative advantage. Following the coding and mapping of data to CFIR constructs, the researchers assigned ratings to each construct, indicating whether it was a barrier, a facilitator, or a neutral factor [18,28]. Construct ratings were determined according to both valence and strength [18,29]. Valence indicated whether a construct predominantly facilitated or hindered AI implementation, whereas strength reflected the degree of its influence. When both positive and negative views were identified within the same CFIR construct, all relevant coded segments were retained and reviewed. The overall valence of the construct was determined first, followed by its strength. The researchers considered the level of agreement among participants, the strength and consistency of the language used, the presence of concrete examples, and the stated influence of the construct on implementation or related constructs. If positive and negative influences were approximately balanced, or if divergent perspectives did not indicate a predominant direction, the construct was rated as neutral (0). If one direction predominated but the supporting evidence was limited, mixed, or general, the construct was rated +1 or –1. If one direction clearly predominated and was supported by consistent, specific, and concrete evidence, the construct was rated +2 or –2. The rating criteria used in this study are presented in Table 1. Information extraction, transcript coding, and construct rating were all carried out independently by 2 researchers (WD and YX). In cases of disagreement, a third researcher (LW) was consulted to reach consensus. Interrater agreement for construct ratings was assessed using weighted κ, which was 0.87. The final results were reviewed and confirmed at a meeting of the research team. NVivo software (Lumivero) was used to support qualitative data analysis by organizing interview transcripts, managing coding structures, grouping codes within CFIR domains and constructs, and facilitating the retrieval and comparison of coded segments across participants. The software served as an analytical support tool, while all coding, interpretation, and rating decisions were made by the research team.

This study was conducted in accordance with the principles of the Declaration of Helsinki. Ethical approval was obtained from the Ethics Committee of Shandong Provincial Hospital (SWYX: number 2025‐888). All participants were informed of the study purposes and signed consent forms. Individual interviews were audiotaped and transcribed if participants agreed. Identifiable information was removed during transcription to protect participants’ anonymity. All research data were stored on password-protected laptops. To prevent disclosure of respondents’ identities, the transcribed data exclude potentially identifying information such as sex, age, professional title, and specific job roles.

A total of 13 eligible ICU nurse managers were invited to participate in this study; 11 completed face-to-face interviews and 2 did not participate: one due to a scheduling conflict and the other due to illness. The interviews ranged from 27 to 78 minutes. The average age of participants was 45.4 (SD 4.5) years (ranging from 34 to 51 y). Half of the participants held a master’s degree, and 90.9% (10/11) were female participants. Overall, 90.9% (10/11) of participants had more than 15 years of management experience and 36.4% (4/11) were associate chief nurses. Participants N1-N3 reported experience with or exposure to several AI-related technologies in their ICUs, including early risk warning systems, consumables management systems, logistics robots, large language models, wearable devices, virtual simulation training systems, intelligent access control systems, intelligent call systems, and others. Detailed demographic characteristics of the participants are presented in Table 2.

This study identified 36 items mapped across the 5 domains and 20 constructs of the CFIR framework. Overall, these constructs were categorized into 5 barriers, 13 facilitators, and 2 neutral influencing factors. Figure 1 and Table 3 summarize these findings. Detailed information can be found in Multimedia Appendix 2. The following sections present the results in detail, with particular attention to factors specific to the Chinese context.

In this study, intervention characteristics referred to nurses’ perceptions and experiences regarding the intrinsic attributes of AI systems. Six constructs were identified as influencing the implementation of AI in ICUs, including 3 barriers, 2 facilitators, and 1 neutral influencing factor (Table 3).

Within the outer setting domain of the CFIR, all 3 identified constructs (patient needs and resources, cosmopolitanism, and external policies and incentives) were identified as facilitators (Table 3).

Patients’ needs and resources emerged as important facilitators. Participants indicated that the rapidly changing and complex conditions of ICU patients placed high demands on nurses’ clinical knowledge and judgment. In addition to clinical demands, participants noted that patients’ emotional experiences and subjective feelings were not always adequately addressed in current practice.

Cosmopolitanism was also identified as a significant facilitator. Most participants emphasized the importance of engaging with manufacturers and suppliers and strengthening collaboration with external organizations, such as technology companies and universities, to support AI development and implementation.

External policies and incentives further facilitated the implementation of AI. Participants emphasized that national-level policy guidance, together with strong leadership attention, provided the confidence to adopt AI technologies in ICUs.

Within the inner setting domain, 4 constructs were identified, including 2 barriers and 2 facilitators influencing AI implementation in ICUs (Table 3).

The characteristics of individuals domain focused on personal attributes of nurses that influence AI implementation. Three constructs were identified as facilitators, including self-efficacy, individual stage of change, and other personal attributes (Table 3).

Self-efficacy was identified as a facilitator. Participants reported that nurses generally showed a positive attitude toward AI and actively sought opportunities to use it, reflecting high confidence in engaging with new tools.

Individual stage of change was also identified as a facilitator. Participants indicated that the broader organizational and societal momentum toward AI adoption created shared expectations among nurses, encouraging proactive engagement and uptake of AI practices.

Other personal attributes further facilitated the implementation of AI. Participants reported that younger nurses were particularly quick to adapt and proactive in using AI technologies, with some even pursuing additional education to strengthen their expertise in AI.

The process domain encompassed activities and strategies related to AI implementation in ICUs. Four constructs were identified, including 3 facilitators and 1 neutral influencing factor (Table 3).

Guided by the CFIR framework, this study explored the barriers and facilitators to implementing AI in ICUs from the perspective of nurse managers across 7 geographic regions in China. A total of 5 barriers, 13 facilitators, and 2 neutral influencing factors were identified across all 5 CFIR domains. Overall, facilitators were reported more frequently than barriers, suggesting that nurse managers generally perceived substantial potential for AI to support ICU nursing practice in China. Beyond identifying barriers and facilitators within the CFIR domains, the study findings suggest that there may be a gap between the positive attitudes toward the implementation of AI in ICUs and the practical readiness. Many of these determinants appeared to reflect contextual characteristics specific to the Chinese health care environment, including strong policy support for digital health, hierarchical leadership structures, and the central role of nurse managers in coordinating interdisciplinary collaboration and guiding frontline adoption.

Within the intervention characteristics domain, relative advantage and design quality and packaging emerged as the most prominent facilitators of AI implementation. These findings were consistent with previous studies in general nursing and clinical settings, which have similarly reported perceived benefits of AI in improving efficiency, clinical decision support, and workflow optimization [13,30-32]. Similar positive perceptions have also been reported in studies on AI-enabled wearable technologies [33] and AI−internet of things health care services [16]. However, our findings further indicated that perceived usefulness alone was insufficient for ICU implementation. This finding suggests that CFIR intervention characteristics may require contextual interpretation in AI-enabled critical care, where perceived advantage is closely linked not only to efficiency but also to algorithmic reliability, patient safety, clinical accountability, and continuous system evolution. Compared with non-ICU settings, uncertainty regarding algorithmic reliability may be perceived more seriously in ICUs because even small errors or delays may affect patient safety [34,35]. In ICUs, AI systems must be reliable, interpretable, adaptable, and compatible with fast-paced clinical workflows. Participants were particularly concerned that algorithmic errors, delayed alerts, or poor contextual adaptation could directly affect patient safety. This concern may be more pronounced in ICUs than in general wards because clinical deterioration can occur rapidly and decisions are often time-sensitive [36,37].

Ethical and humanistic concerns emerged as important issues in this study. Previous studies have highlighted privacy risks, ethical uncertainty, overreliance on AI, and threats to professional autonomy as barriers to implementation [30,38,39]. In this study, nurse managers raised similar concerns. Participants emphasized that AI should not replace nurses’ clinical judgment or weaken humanistic care. Instead, AI should support nurses in making timely and informed decisions while preserving patient-centered care [40]. This finding highlights a tension that may be particularly salient in ICU settings between technological efficiency and ethical-humanistic responsibilities. Meanwhile, the boundaries of human-machine collaboration need to be clearly defined before AI can be safely embedded into ICU workflows.

Cosmopolitanism and external policies and incentives emerged as facilitators. Participants emphasized the value of collaboration with technology companies, universities, and interdisciplinary teams. This finding is consistent with international literature, indicating that successful AI implementation in health care depends on cooperation among clinical, technical, managerial, and regulatory stakeholders [41,42]. In China, strong governmental support for digital health and AI innovation provides legitimacy and momentum for AI adoption in ICUs. Such policy support may accelerate early-stage implementation by mobilizing resources and encouraging institutional engagement. However, the transferability of these findings to other health care systems should be interpreted cautiously. China’s health care system is characterized by strong policy guidance and relatively hierarchical organizational structures, which may facilitate top-down implementation.

Nurse managers emphasized that leadership vision strongly influenced the prioritization of AI implementation and resource allocation. These findings align with evidence from non-ICU settings, where leadership and organizational culture are consistently identified as critical determinants of innovation uptake [13,38]. Our findings add to this perspective by suggesting that nurse managers serve not only as organizational leaders but also as boundary-spanning actors who connect multiple implementation domains. ICU work is characterized by high workload intensity, rapid decision-making, and close interprofessional dependency. Nurse managers function as boundary spanners who connect multiple implementation domains, translating organizational priorities into clinical practice through coordination, resource allocation, and capability building. In this sense, AI implementation in ICUs is not only a technical process but also an organizational change process.

Training and education were also identified as essential facilitators. Consistent with prior studies, insufficient AI literacy [43] and limited digital competence [44] can hinder adoption among nurses and nursing students. Targeted training and continuing education programs can effectively promote the implementation of AI in nursing practice [45,46]. Notably, participants proposed incorporating AI applications into departmental requirements and mandating their use in specific scenarios to promote implementation. However, mandatory implementation strategies should be applied cautiously, as premature or overly rigid requirements may provoke resistance if AI systems are not yet sufficiently mature, interpretable, or trusted by frontline nurses. Generational differences were also observed. Younger nurses appeared more adaptable to AI, whereas senior nurses tended to be more cautious. This suggests the need for tailored training strategies and intergenerational learning models [13], with younger nurses potentially serving as AI champions. At the same time, the lack of nursing informatics professionals was identified as a barrier. Currently, few regions have incorporated nursing informatics into the required curriculum for the foundational nursing stage [40]. Nursing administrators should therefore advocate for the systematic integration of informatics and AI-related competencies into nursing education and continuing professional development.

Overall, while many facilitators and barriers are shared with general health care settings, AI implementation in ICUs is not simply an extension of general nursing AI adoption. ICU environments amplify concerns about safety, reliability, time sensitivity, ethical accountability, and workflow compatibility. These findings may inform the contextual application of the CFIR framework by illustrating how implementation determinants can be interpreted in high-risk clinical environments such as ICUs. Specifically, this study suggests a potential gap between positive implementation attitudes and actual organizational preparedness for AI adoption. Although supportive policies, leadership engagement, positive implementation climate, and professional self-efficacy were identified as important facilitators, challenges related to technological adaptation, resources, governance, and interdisciplinary collaboration may limit practical implementation. At the same time, issues such as algorithmic reliability, accountability mechanisms, transparency, and human-centered care may constitute cross-dimensional factors that affect multiple areas of implementation. Traditional classification frameworks may not be fully sufficient to provide a comprehensive explanation. Further research is needed to clarify how AI-specific dimensions can be integrated into implementation frameworks. Interestingly, this study identified more facilitators than barriers. This finding should be interpreted in light of the participants’ managerial roles. These participants may be more familiar with organizational strategy and innovation initiatives and therefore hold a more positive attitude toward the application of AI. Their managerial position provided unique insights into organizational readiness, leadership engagement, and contextual conditions that are often less visible from frontline perspectives. Consequently, while frontline nurses may place greater emphasis on usability, workload, and workflow disruption, nurse managers highlighted broader organizational and implementation issues that are essential for successful AI integration.

The findings of this study have important implications for ICU nursing management, policymakers, and technology developers. For nursing leaders, strengthening AI-related education and cultivating interdisciplinary talent are essential. Nursing leaders should develop interdisciplinary collaboration mechanisms and ensure that clinical requirements are effectively communicated to technical teams. Incentive mechanisms and role modeling strategies can further foster a supportive climate. Nursing organizations should also develop clear guidelines for the safe and appropriate use of AI technologies in clinical practice. For policymakers, supportive regulatory frameworks, targeted funding incentives, and standardized guidelines for AI governance and data security are critical to addressing cost barriers and privacy concerns. At the developer level, improving algorithm adaptability, interpretability, usability, and clinical integration may be critical to enhancing trust and sustainable adoption.

Future research should incorporate broader stakeholder perspectives, including frontline nurses, physicians, IT professionals, and patients, and should combine qualitative insights with quantitative evaluations of implementation outcomes. Longitudinal studies and mixed methods approaches will also be valuable for examining how AI implementation experiences change as technologies mature and are integrated into routine clinical practice. Theoretical research and empirical validation are also needed to clarify how these AI-specific dimensions can be integrated into implementation frameworks. Intervention studies testing context-sensitive implementation strategies will also be valuable in advancing evidence-based integration of AI in critical care settings.

Several limitations should be noted. First, the sample size was relatively small, and all participants were from China. Although we included participants across 7 geographic regions to capture variation in implementation contexts, selection bias may still exist. As health care systems, organizational structures, and implementation environments vary across countries, the transferability of these findings to other settings requires further investigation through studies conducted in diverse health care contexts. Second, this study focused specifically on the perspectives of nurse managers and did not include other key stakeholders, such as frontline nurses. As key decision-makers and coordinators within ICU nursing systems, nurse managers play an important role in shaping implementation priorities, facilitating interdisciplinary collaboration, allocating resources, and supporting organizational change. Their perspectives are therefore particularly valuable for understanding organizational and contextual determinants influencing AI implementation, which was the primary focus of this study. However, the role and position of nurse managers may introduce a hierarchical or managerial bias. Future research should incorporate multiple stakeholder groups, particularly frontline nurses, to obtain a more comprehensive understanding of the implementation of AI determinants in ICU settings. Third, AI implementation remains at an early stage, and perceptions may evolve as technologies mature. Finally, CFIR-guided inquiry and cross-language translation may have influenced theme emergence and nuanced meaning despite efforts to ensure analytic rigor.

This study explored facilitators and barriers to AI implementation in ICUs from the perspective of nurse managers. The findings suggest that AI implementation is influenced by multiple determinants across technical, organizational, workforce, and process domains, as reflected in the CFIR framework. In particular, intervention characteristics and inner setting factors appear especially salient in ICUs. Nurse managers held a cautiously optimistic attitude toward AI implementation in ICU settings. They recognized its potential to improve clinical efficiency, decision support, workflow optimization, and patient safety. However, they also expressed concerns regarding limited humanistic care, ethical risks, and insufficient regulatory oversight. These challenges highlight the need for stronger governance systems, structured interdisciplinary collaboration, and targeted training for ICU nursing staff to support safe and sustainable AI integration in critical care.