Global physician perspectives on artificial intelligence in healthcare across 50 countries and territories

Abstract

Despite rapid advances in artificial intelligence (AI) capabilities, clinical integration remains limited, partly due to an incomplete understanding of physicians’ real-world engagement and determinants of adoption. We conducted an international cross-sectional survey between June and September 2024 using a 30-item questionnaire translated into 13 languages, analysing 1049 complete responses from 50 countries and territories. Most respondents reported fundamental to advanced understanding of AI (86.5%) and believed it would improve clinical practice (80.2%), particularly in efficiency (53.5%), timeliness (52.0%), and effectiveness (44.0%). However, only 27.8% had used AI in practice, and 17.7% had received formal training. Mann–Whitney analyses showed that positive attitudes and formal training were associated with greater AI understanding, familiarity and usage (all p ≤ 0.008). In multivariable regression, physicians with formal training were more than three times as likely to use AI (adjusted OR 3.40; 95% CI 2.31–5.01; p < 0.001), and those working in institutions with AI technologies were more than eight times as likely (adjusted OR 8.43; 95% CI 5.74–12.34; p < 0.001). These findings indicate a persistent gap between awareness and adoption driven primarily by structural factors, particularly the absence of formal AI training and limited institutional investment in AI technologies.

Introduction

Artificial intelligence (AI) is fast gaining recognition in the healthcare sector as a greater number of clinicians become increasingly AI-literate. Many medical experts have already begun to adopt AI in both research and clinical practice to improve decision-making, diagnostic precision, and clinical efficiency. These developments are largely attributed to improvements in algorithmic performance, enhanced computational infrastructure, and the availability of large, diverse healthcare datasets that enable AI systems to learn with contextual medical nuance¹. The rapid expansion of FDA-approved AI/ML-enabled medical devices has accelerated clinical adoption opportunities, particularly in imaging and diagnostics².

From the perspective of physicians, AI has demonstrated the capacity to enhance healthcare quality through improved accuracy in diagnostics, reduced medical errors, and the automation of routine tasks that allow clinicians to spend more time in direct patient care^3,4. A study highlighted that AI-enabled medical devices or algorithms is high among radiology, cardiovascular and internal or general medicine⁵. However, despite promising results, considerable debate persists within both the public and scientific communities regarding the clinical value and safe implementation of these technologies. The adoption and implementation of newly approved AI solutions in the majority of clinical practices have thus been slow as a result, predominantly due to a lack of substantial clinically proven advantages, difficulties in integration into existing workflows, and uncertain business models⁶. Some critics express concern that increasing reliance on AI may lead to a loss of clinical skills and decision-making autonomy, while others worry that its integration could diminish the human elements of care, such as empathy, trust, and personal connection, that are central to the doctor–patient relationship⁷. These concerns primarily emerge from how AI technologies are being implemented, rather than the intrinsic capabilities of the tools themselves. When thoughtfully integrated and ethically deployed, AI has the potential to meaningfully enhance patient care.

Beyond clinical precision, there is a growing consensus that AI can play a pivotal role in addressing social determinants of health (SDOH) by enabling predictive risk modeling, improving access to underserved communities via telehealth and mobile health tools, and identifying patterns in population health data that would otherwise remain obscured. Equitable deployment of AI remains a global challenge, necessitating policies to address disparities in access and outcomes^8,9. These technologies can optimize care delivery in low-resource settings by compensating for workforce shortages and infrastructure gaps⁷. While AI holds promise for system-level improvements, its success ultimately depends on those who implement it chiefly, physicians at the frontline of care delivery. Importantly, physicians are not passive recipients in this transformation. Rather, they serve as critical gatekeepers and mediators of AI adoption. Their insights, trust, and frontline experiences determine the success or failure of AI integration into clinical practice. The human-in-the-loop paradigm must be preserved, and physicians must actively guide the development, testing, and deployment of AI solutions to ensure they are clinically relevant, ethically sound, and patient-centered¹⁰. Strengthening the physician–AI partnership is essential to realize AI’s clinical potential¹¹. Therefore, we conducted an international cross-sectional study of physicians to explore perspectives and practices, identify challenges and benefits, and examine the factors influencing AI adoption and its broader impact on healthcare.

Results

Participants demographics

Our study included 1,049 participants from 50 countries and territories. Of these, 614 were female (58.5%), 421 were male (40.1%), and 14 preferred not to disclose their gender (1.3%). Participants ranged in age from 20 to 75 years (M = 33.66, SD = 8.57). A total of 413 participants (39.4%) were trainees or residents, while 636 (60.6%) were practicing physicians. Most respondents worked in public hospitals or academic institutions (n = 764, 72.8%), while the remainder were in private healthcare settings (n = 285, 27.2%). Participants’ demographics are summarized in Table 1. Geographically, one country or territory contributed ≥100 participants (2%), six contributed 50-99 (12%), twelve contributed 10–49 (24%), and 31 contributed fewer than 10 (62%). The geographic distribution is illustrated in Fig. 1. By clinical specialty, Radiology had the highest representation (n = 305, 29.1%), followed by General Practice (n = 200, 19.1%) and Pediatrics (n = 85, 8.1%).

Table 1.

Participant demographics (N = 1049)

Demographics	N = 1049 (100%)
Age (years), SD, range	M = 33.66, SD = ± 8.57
Gender
Male	421(40.1%)
Female	614(58.5%)
Prefer not to say	14(1.3%)
Primary workplace
Public	764(72.8%)
Private	285(27.2%)
Distribution of participants by country
≥100 participants	1 (2%)
50–99 participants	6 (12%)
10–49 participants	12 (24%)
<10 participants	31 (62%)

Fig. 1. Geographic distribution of survey participants.

Country-level participant counts are provided in Supplementary Information 1.

General experience and perspectives

Participants reported a generally high level of understanding and familiarity with AI in healthcare. Overall, 907 participants (86.5%) indicated a fundamental to advanced understanding of AI, while 770 (73.4%) reported at least a basic to very high level of familiarity with AI technologies. Trust in AI for assisting with diagnostic and treatment decisions varied across the sample. Specifically, 66 participants (6.3%) reported no trust, 305 (29.1%) expressed minimal trust, 499 (47.6%) indicated moderate trust, 148 (14.1%) reported substantial trust, and 31 (3.0%) expressed extensive trust. Perceptions of whether AI products are aligned with healthcare needs were nearly evenly divided: 518 participants (49.4%) believed that available AI tools were tailored to clinical requirements, while 531 (50.6%) disagreed. Similarly, views on AI’s practical impact on clinical care were split, with 527 participants (50.2%) reporting little or no improvement in medical practice following AI integration. In terms of practical exposure, only 292 participants (27.8%) reported prior experience using AI tools in healthcare, while 757 (72.1%) had no such experience. A majority (863 participants, 82.3%) also indicated that they had not received any formal AI-related training. Ethical and privacy concerns were prominent: 948 participants (90.4%) expressed some level of concern regarding patient data privacy in the use of AI for clinical care. Among those with direct AI experience (n = 292), approximately 76 (26.0%) reported encountering data privacy issues in practice (see Supplementary Information 2). The overall distribution of responses is illustrated in Fig. 2.

Fig. 2. General experience and perspectives on AI in medicine among physicians (N = 1049).

Physicians’ experience and perspectives on artificial intelligence in medicine. This figure summarizes responses from 1,049 physicians regarding their experience, familiarity, and perceptions of AI in clinical practice.

Barriers and challenges

Key barriers to AI implementation included limited access to AI technologies (29.8%, n = 313) and difficulties integrating these tools with existing systems (24.7%, n = 259). The most prominent obstacle, however, was the lack of infrastructure and resources, identified by 62% (n = 650), followed by concerns related to data privacy and security (48%, n = 504). Among participants with prior experience using AI tools, 67.8% (n = 198) reported encountering instances in which the technologies did not meet clinical expectations. Despite these challenges, physicians expressed a clear desire to be more actively involved in the adoption process. Collaboration among healthcare providers (65.6%, n = 688) and capacity-building and training initiatives (55.2%, n = 579) were identified as key strategies to support effective AI integration (Supplementary Information 2).

Expectations and use of AI in addressing healthcare challenges

Participants reported several persistent challenges and solutions in healthcare delivery and their impacts on wider health determinants in Fig. 3. Limited resources (64.9%; n = 681) and shortages of healthcare professionals (54.5%; n = 572) were the most frequently cited. Additional concerns included limited access to healthcare services (42.5%; n = 446) and inadequate infrastructure (40.5%; n = 425). When asked about areas where AI could enhance quality of healthcare, participants identified efficiency (53.5%; n = 561), timeliness (52.0%; n = 545), and effectiveness (44.0%; n = 461) as top priorities. The most widely supported AI-driven solutions were telemedicine and remote patient monitoring (67.5%; n = 708), and decision-support systems for diagnosis and treatment (57.7%; n = 605), reflecting their perceived potential to address gaps in service delivery. Regarding AI’s impact on social determinants of health, 43.7% (n = 458) believed AI could help reduce healthcare inequalities, whereas 25.8% (n = 271) expressed uncertainty. Among specific applications, medical imaging analysis was viewed as the most useful AI-enabled tool (68.1%; n = 714), followed by electronic health record (EHR) management and analysis (60.5%; n = 635). Only 18.4% (n = 193) reported that their institution had implemented AI-based tools, and 40.6% (n = 426) rated the alignment between available AI solutions and actual healthcare needs as moderate (Supplementary Information 2).

Fig. 3. Current challenges, potential solutions and their impact on healthcare.

Challenges, potential solutions, and their impact on healthcare systems. This figure illustrates key barriers to AI adoption, proposed solutions, and their perceived impact on healthcare delivery.

Determinants of physicians’ perspectives and engagement with artificial intelligence

We examined factors influencing physicians’ attitudes and behaviours toward AI across demographic and professional subgroups (n = 1,049). Statistically significant associations (p < 0.05) were identified across multiple factors as shown in Fig. 4. Gender differences were particularly notable. Male and female participants differed significantly in their understanding of AI (U = 113,587, n = 1,035, p < 0.001), familiarity with AI potential (U = 116,226.5, n = 1035, p = 0.003), and trust in AI applications (U = 119,261, n = 1035, p = 0.023). Institution type was associated with differences in AI engagement. Respondents from public and private healthcare settings differed significantly in their understanding (U = 99,978, n = 1,049, p = 0.025), current usage (U = 102,226, n = 1049, p = 0.049), and familiarity with AI applications (U = 97,333, n = 1049, p = 0.005). However, as most participants were from public hospitals or academic institutions (n = 764, 72.8%), these results should be interpreted cautiously due to unequal group representation.

Fig. 4. Physicians’ Experience, Familiarity, and Trust in AI (n = 1049).

*Gender-based analysis includes 1035 respondents; 14 participants who selected “prefer not to say” were excluded. Physicians’ experience, familiarity, and trust in artificial intelligence. This figure presents survey results from 1049 physicians; gender-based analysis includes 1035 respondents, excluding 14 who selected “prefer not to say”.

Physicians who believed AI is improving healthcare reported greater understanding (U = 119,845, n = 1049, p < 0.001), current use (U = 127,139, n = 1049, p = 0.006), familiarity (U = 125,456, n = 1049, p = 0.008), and trust (U = 118,040, n = 1049, p = 0.001), as well as fewer ethical concerns (U = 118,112, n = 1049, p < 0.001) and stronger perceptions of AI alignment with healthcare needs (U = 102,831, n = 1,049, p < 0.001). Formal AI training was a major determinant of engagement. Participants who had received formal training reported greater understanding (U = 58,632, n = 1049, p < 0.001), usage (U = 47,622, n = 1049, p < 0.001), and familiarity with AI (U = 46,930.5, n = 1,049, p < 0.001), underscoring the value of structured education in fostering digital competency. Institutional involvement in AI purchasing was similarly associated with higher understanding (U = 71,019.5, n = 1049, p = 0.001), current usage (U = 38,401, n = 1049, p < 0.001), familiarity (U = 58,101, n = 1049, p < 0.001), and perceived alignment between AI tools and healthcare priorities (U = 50,445.5, n = 1049, p < 0.001). These findings indicate that demographic, attitudinal, and institutional factors, particularly training and organizational investment play substantial roles in shaping physicians’ understanding, trust, and adoption of AI in healthcare.

Predictors of AI use in clinical practice

In multivariable logistic regression, several factors were associated with physicians’ use of AI-powered tools, as shown in Fig. 5. Age showed a nonsignificant trend toward higher AI use (OR 1.36; 95% CI, 0.96–1.92; p = 0.082), while female gender showed a nonsignificant trend toward lower AI use (OR 0.85; 95% CI, 0.64–1.12; p = 0.243). Working in a public hospital versus a private hospital was not significantly associated with adoption (OR 0.88; 95% CI, 0.62–1.24; p = 0.461), nor was holding a trainee or resident position versus practicing physician (OR 0.90; 95% CI, 0.63–1.29; p = 0.573). Believing that AI tools are tailored to healthcare needs also showed no significant effect (OR 1.09; 95% CI, 0.80–1.49; p = 0.591).

Fig. 5. Predictors for the use of AI.

Predictors of artificial intelligence use in clinical practice. This figure presents factors associated with AI adoption based on multivariable regression analysis.

In contrast, AI-related training and institutional exposure were strong predictors of AI adoption. Physicians who had received formal AI training were more than three times as likely to use AI (OR 3.40; 95% CI, 2.31–5.01; p < 0.001), and those working in institutions that had purchased AI technologies showed the strongest association, being over eight times more likely to use AI (OR 8.43; 95% CI, 5.74–12.34; p < 0.001).

Consistent with these findings, physicians also identified demonstrated effectiveness and accuracy (36.0%, n = 378) and ease of integration with existing systems (30.4%, n = 319) as the strongest self-reported drivers of adoption, while cost-effectiveness (12.2%, n = 128) and regulatory approval (7.82%, n = 82) were less frequently cited. These findings indicate that institutional investment and formal AI training are the most influential determinants of AI adoption among physicians, whereas demographic and perceptual factors show comparatively modest associations.

Discussion

This international cross-sectional survey of 1049 physicians offers practical insights into how clinicians perceive, experience, adopt and engage with AI in clinical practice. The findings indicate that physicians are not only increasingly aware of AI’s expanding role but are actively shaping its integration into healthcare practices, particularly when supported by robust institutional readiness, comprehensive training programs, and strategic infrastructure improvements. This aligns with broader evidence emphasizing the critical necessity of institutional preparedness, targeted professional training, and ensuring AI technologies closely reflect practical clinical realities.

In our 2024 global survey, most physicians (86.5%, n = 907) reported at least a fundamental to advanced level of understanding of artificial intelligence in healthcare. This shows improvement compared with the findings of a 39-country survey by Chen et al.¹², in which 53% (n = 758) of respondents reported lacking basic knowledge of clinical AI, highlighting a marked increase in baseline awareness and understanding of AI among physicians over the past few years. Optimism regarding AI’s potential impact has also remained high: Chen et al¹². reported that 77% of respondents were optimistic about the role of AI in clinical practice, and our survey revealed a similar outlook, with 80.17% (n = 841) of physicians anticipating that AI will improve their clinical practice. However, nearly half of our respondents also emphasized that existing AI products still require better alignment with real-world healthcare needs. This aligns with studies highlighting the need to meet clinical, patient, market, and regulatory demands for success in e-health. Global trends show rising awareness but persisting variability in physicians’ readiness to adopt AI¹³. Importantly, the increase in awareness and optimism has not been matched by greater adoption in practice. In our survey, only 27.8% (n = 292) of physicians reported prior experience using AI tools in clinical practice, comparable to the 20% usage rate reported in the cross-sectional survey¹². This persistent gap between rising awareness and limited real-world adoption underscores the need for targeted training, supportive environment, and context-specific validation of AI tools to translate clinicians’ enthusiasm into meaningful integration within everyday clinical workflows. However, large-scale implementation requires robust governance frameworks to balance potential benefits and risks¹⁴.

Our study identified proven effectiveness (36.0%, n = 378) and ease of integration with existing systems (30.4%, n = 319) as the strongest drivers of AI adoption, while infrastructural limitations (61.9%, n = 650) and restricted access to data and computing power (58.8%, n = 617) were the most prominent barriers. Multivariable regression further demonstrated that physicians with AI training and those working in institutions that had procured AI systems were substantially more likely to use AI, underscoring the importance of organizational resources and AI education. In our sample, only 17.7% (n = 186) of respondents reported receiving AI training, while 82.3% (n = 863) had not, consistent with a prior study¹², which highlighted a widespread deficit in formal AI education across healthcare populations. Similar to our findings, a global review identified a lack of AI training as a consistent barrier to adoption¹⁵.

In contrast, demographic variables such as gender and professional role were not significant predictors, showing that although physicians can develop AI competencies on their own, structural and organizational factors ultimately determine widespread integration. Da Costa et al.¹⁶ examined both health professionals and managers, finding that knowledge and perceived benefits strongly predicted intention to adopt AI. Their results emphasize cognitive and attitudinal factors, whereas our findings highlight the decisive role of systemic enablers such as infrastructure, procurement, and training in translating intentions into practice. At the same time, our participants emphasized their role as proactive contributors, with 63.2% (n = 663) providing input on clinical needs, 58.7% (n = 616) participating in design and testing, and 58.7% (n = 616) advocating for ethical use. This dual perspective highlights that successful AI adoption requires not only institutional capacity but also active provider engagement in co-production.

The responses show a clear expectation that AI should make healthcare faster and more efficient, with more than half highlighting timeliness (52.0%, n = 545) and efficiency (53.5%, n = 561) as its main contributions. In comparison, Oh et al.¹⁷ reported that 83.4% of physicians considered AI clinically useful, especially in diagnostic efficiency. Alzghoul et al.¹⁸ further support these findings in their systematic review and meta-analysis, which showed a significant positive impact of AI on healthcare quality despite high heterogeneity (I² = 92%, Z = 1.93, p ≤ 0.05), with subgroup analyses confirming consistent benefits across objectives. Also, the most potential AI solutions for healthcare challenges were telemedicine and remote monitoring (67.5%, n = 708), decision support for diagnosis and treatment (57.7%, n = 605), and disease surveillance (58.1%, n = 609). However, far fewer linked AI to equity (19.0%, n = 199), a reminder that technology alone will not correct structural disparities. Previous studies have shown that poorly designed algorithms can perpetuate or even exacerbate healthcare disparities¹⁹. Recent advances such as adversarial training have been proposed to address bias in medical AI models²⁰. These results indicate that AI is expected to enhance healthcare delivery through faster processes and stronger decision support, but its limited association with equity suggests a risk of widening disparities if fairness is not prioritized in implementation.

Physicians were divided on AI’s role in social determinants of health: 43.7% (n = 458) expected reductions in inequalities, 12.1% (n = 127) feared exacerbation, 25.8% (n = 271) were uncertain, 17.7% (n = 186) lacked familiarity with SDOH, and 0.7% (n = 7) provided other responses. From the patients’ perspective, a global scoping review reported that over 75% identified gains in efficiency and diagnostic accuracy, yet 50–70% also raised concerns about loss of human connection in healthcare delivery, reliability, and safety uncertainties, privacy and data security risks, and issues with trust, transparency, and patient autonomy²¹. These perspectives together indicate that equity is not automatic; AI requires human-centered design from ideation through evaluation. Ultimately, technology must serve people, not the reverse.

This study has several limitations that should be acknowledged. First, although the survey yielded 1049 completed responses, representing 59.33% of those who started (n = 1768) and 49.90% of those who viewed the survey (N = 2102), the voluntary nature of participation introduces the potential for response bias. Physicians with a pre-existing interest in artificial intelligence may have been more inclined to participate, limiting the representativeness of the sample. Second, although responses were collected from 50 countries and territories, geographic and specialty distribution was highly uneven, with 64% of countries contributing fewer than 10 participants and radiology being disproportionately represented. The survey was translated into 13 languages; however, the absence of additional translations limited participation from some regions. As a result, the findings may not be generalizable to all regions or clinical disciplines, and a systematic review or larger national and multinational study would be needed to establish more representative global estimates. Third, most participants had limited hands-on exposure to AI (only 27.8% had used AI tools), and most had not received formal AI training (82.3%). Consequently, many responses may reflect perceptions or expectations rather than practical experience, potentially affecting the accuracy of insights regarding real-world implementation. Fourth, the cross-sectional and self-reported nature of the survey introduces the possibility of recall and social desirability bias and does not allow for causal conclusions. For example, while AI training and institutional procurement were associated with higher adoption, the temporal direction of these relationships cannot be determined. Future longitudinal or mixed-method studies may provide deeper insight into how physician engagement with AI evolves over time. Fifth, the survey employed ‘artificial intelligence’ as a broad term, consistent with prior multinational physician surveys^22–24, and did not distinguish between predictive, generative or foundation model-based, and agentic AI systems; future studies should consider characterising AI type explicitly to capture more nuanced differences in physician experience and trust²¹. Finally, this study was conducted in 2024. Since data collection, ambient AI tools for clinical documentation have entered routine practice, exposure to foundation model-based tools has been shown to meaningfully increase public trust and acceptance of AI in healthcare²⁵, agentic systems capable of autonomous multi-step clinical task execution have emerged as a distinct paradigm, and governments are now committing to sovereign AI infrastructure as a matter of national health policy²⁶. Findings should be read with this rapid evolution in mind.

In conclusion, physicians worldwide demonstrate growing awareness of AI’s potential in clinical practice, yet a persistent gap between awareness and real-world adoption remains. This gap is not primarily explained by attitudes or demographic factors, but by structural determinants, particularly the absence of formal AI training and limited institutional investment in AI technologies. Empowering physicians as active collaborators through targeted education, workflow integration, and organisational readiness is essential to translate awareness into meaningful and equitable adoption across diverse health systems.

Methods

Ethical approval

This prospective, anonymous online survey was approved by the Human Research Ethics Committee of Intermed Hospital (Ref No. 01/223) and conducted in accordance with the principles of the Declaration of Helsinki. All participants provided electronic informed consent prior to participation. Clinical trial number: not applicable.

Survey design

Data were collected between June and September 2024 via the QuestionPro platform. The survey was developed by a radiologist and pediatrician with MSc-level training in Digital Health at the University of Bristol, based on existing surveys and published sources related to artificial intelligence in healthcare^22–24. The questionnaire comprised 30 items with an informed consent section (Supplementary table 1), and was divided into four domains: (1) demographics and clinical background; (2) experience, familiarity, and trust in AI; (3) perceived clinical impact of AI tools; and (4) broader views on AI’s influence on healthcare quality, adoption barriers, training needs, and social determinants of health. Consistent with the study’s aim to capture real-world physician perspectives across diverse global settings, the survey employed ‘artificial intelligence’ as a broad term without distinguishing between predictive, generative or foundation model-based, and agentic AI systems. This approach was intentional, reflecting the heterogeneous AI exposure among participants across countries and territories, and consistent with the terminology used in prior multinational physician surveys on which this instrument was based^22–24. The survey included both closed-ended and Likert-scale questions. The full questionnaire is available in Supplementary Table 1. Also, it was translated into 13 languages (English, Burmese, Chinese, Hindi, Indonesian, Japanese, Laotian, Latin American Spanish, Mongolian, Thai, Turkish, and Vietnamese) by native speakers to ensure linguistic and contextual accuracy and subsequently circulated among the research co-authors for validation and pilot testing. The questionnaire was disseminated through Facebook, LinkedIn, and professional networks including the Chevening Scholars cohort (2023/2024) and the Youth Club of the Asian Oceanian Society of Radiology.

Statistical analysis

Descriptive statistics were used to summarize participant demographics and survey responses as proportions. The Mann-Whitney U test, a non-parametric method, was employed to assess differences in AI perceptions and usage across demographic groups (e.g., age, gender, institution type, professional position). Logistic regression analyses were conducted to identify predictors of AI adoption, using age, gender, institution type, professional role, belief in AI, prior AI training, and AI-related purchases as covariates. Results are reported as adjusted odds ratios (ORs) with 95% confidence intervals (CIs). Survey items measuring attitudes and trust in AI were rated on Likert scales. P-values were used to assess associations between demographic variables and AI-related perceptions, ethical concerns, and trust. P-values less than 0.05 were considered statistically significant. OpenAI’s ChatGPT models 5.0 and 4.1 were used for language editing.

Author contributions

B.B. and O.S. led the conceptualization, with additional contributions from T.L., S.-E.D., M.S.K., C.W., R.E.J., L.G., C.R., N.T., N.K.W., F.F., U.K., T.N.N., B.T., N.P., P.S.W.N., R.K. and P.A.K.-B. B.B., M.S.K. and N.K.W. drafted the main manuscript, while W.L. reviewed the manuscript. Data analytics were conducted by B.B., O.S., M.S.K. and W.L. Survey development was carried out by B.B. and O.S. Survey translation, validation, and distribution were performed by T.L., S.-E.D., M.S.K., C.W., R.E.J., L.G., C.R., N.T., N.K.W., F.F., U.K., T.N.N., B.T., N.P., P.S.W.N., R.K. and P.A.K.-B. All authors have read and approved the manuscript

Data availability

All data supporting the findings of this study are available within the paper and its Supplementary Information.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02726-y.