Skip to main content
NPJ Digital Medicine logoLink to NPJ Digital Medicine
. 2026 Mar 19;9:364. doi: 10.1038/s41746-026-02552-2

Innovating global regulatory frameworks for generative AI in medical devices is an urgent priority

Jasmine Chiat Ling Ong 1,2, Yilin Ning 2,3, Mingxuan Liu 2,3, Yian Ma 2,3, Liang Zhao 4, Kuldev Singh 5,6, Robert T Chang 5,6, Silke Vogel 7, John C W Lim 7, Iris Siu Kwan Tan 8, Oscar Freyer 9, Stephen Gilbert 9, Danielle S Bitterman 10,11, Xiaoxuan Liu 12,13, Alastair K Denniston 12,13, Nan Liu 2,3,14,15,16,✉
PMCID: PMC13168469  PMID: 41857339

Abstract

The integration of generative AI (GenAI) and large language models (LLMs) in healthcare presents both unprecedented opportunities and challenges, necessitating innovative regulatory approaches. In this perspective, we discuss the risks of GenAI and LLM-based medical devices, the limitations of current medical device regulation frameworks when applied to GenAI or LLMs, and advocate for global collaboration in regulatory science research through engaging multidisciplinary expertise and focusing on the needs of diverse populations.

Subject terms: Health policy, Computational science

Main

The rapid integration of generative artificial intelligence (GenAI) in healthcare has exposed a significant regulatory gap. GenAI algorithms, such as large language models (LLMs), are now capable of producing new text, audio, image and video outputs from training data. Despite a notable increase in healthcare AI applications, existing health products regulatory frameworks across jurisdictions differ and are not specifically tailored to address the unique challenges posed by GenAI and LLMs. Current regulatory approaches were developed for AI models designed to generate specific healthcare decisions or recommendations. In contrast, LLM models are capable of a broad range of functions, such as summarisation of general medical information and suggesting differential diagnoses, often applied in tasks for which they were not purposefully designed. This represents a shift from intended use-specific medical decision-making, which was relatively straightforward to regulate, to more generalised and flexible applications of AI. This shift requires an evaluation of current medical device regulations (MDRs) to ensure that they remain relevant to GenAI/LLMs

To enable timely, safe and high accessibility to effective AI/machine learning (ML)-based medical devices, regulators, including the US Food and Drug Administration (FDA) and UK Medicines and Healthcare products Regulatory Agency (MHRA), are adapting the Total Product Life Cycle (TPLC) approach for these models1,2. This approach was also emphasised in the executive summary of the FDA’s Digital Health Advisory Committee Meeting, forming the bedrock of the committee’s recommendations for regulating GenAI-enabled medical devices3. TPLC broadly encompasses the evaluation and monitoring of a software product from its premarket development to post-market performance, which enables a rapid cycle of product improvement with safeguards. This approach enhances device safety and accelerates innovation by enabling early risk identification, continuous monitoring, and the use of real-world evidence for informed regulatory decisions. However, it poses challenges in integrating and analysing complex data and requires substantial resource commitments from both regulators and manufacturers.

The poor fit between LLM-based medical devices and current regulatory frameworks results in potential risks which necessitate oversight from stakeholders across the ecosystem, aside from regulatory authorities4. In a recent commentary, the FDA stated the need to keep a view on the broader ecosystem, such as maintaining healthy competition among start-ups, big tech and academia; reinforcing the responsibility of regulated industries; and emphasising the importance of global harmonisation of regulatory standards and data definitions. In this perspective, we provide an in-depth discussion on the outstanding challenges that are not adequately addressed by a TPLC approach, focusing on the regulation of LLM-based medical devices. We explore the emerging role of international and multidisciplinary regulatory science consortia in advancing LLM innovations responsibly, as well as proposing a call for global action to drive equitable and responsible GenAI governance.

The constraints of the TPLC approach

Unique challenges to regulation exist for LLMs embedded into medical devices5. First, LLMs are trained on extensive datasets gathered from the internet and other diverse sources, making it virtually impossible to thoroughly examine or scrutinise the training data. These can include digital books, websites, personal blogs or articles that may contain misinformation. Second, long-form LLM outputs are subject to concerns over poor repeatability even with the same prompt strategy. Decision support systems using LLMs may experience performance deterioration when used in real-world settings, particularly if LLMs are sensitive to changes in the syntax of input queries. Data fragmentation remains a significant challenge in healthcare, as different institutions use varying data structures and file formats. Additionally, inconsistencies in medical documentation, such as the use of differing terminologies, jargon, and non-standard abbreviations, can lead to variations in outcomes or system performance. Third, the risk of privacy and confidentiality breaches remains unresolved. This is especially challenging with LLM-based devices as the models are trained on vast datasets that may inadvertently include sensitive or identifiable health information. LLMs’ opaque architecture and ability to memorise or regenerate personal data make conventional privacy safeguards insufficient. In the following section, we discuss the limits of regulatory oversight using the TPLC approach for LLMs.

Ambiguity in medical device definitions for LLMs and LLM-based applications

The risk classification and degree of control required for LLM-based medical devices may need revision to cater for risks specific to LLM-based devices. The applicability of various regulatory frameworks, including the European Union’s MDR and the US FDA’s Software as a Medical Device (SaMD), is based on the AI model’s or AI-based device’s intended use and proposed indications6. However, the qualification of LLMs and LLM-based applications as medical devices remains a topic of debate. For the EU, under the EU AI Act, LLMs are classified as being of high risk or posing systemic risk. Downstream LLM-based applications with a medical purpose will likely qualify as medical devices. Table 1 summarises how different regulatory frameworks define a medical device. Many LLMs and LLM-based applications will likely fall into a grey area where their classification remains unclear. One example is LLM tools built to transcribe and summarise patient-physician interactions into a clinical document. Creation of the clinical summary requires the model to interpret clinical data and suggest medical information that may influence decisions made by subsequent reviewing clinicians. In the US, for example, various companies have marketed a variety of clinical documentation automation products without seeking regulatory approval as a medical device. The commercial presence of these products may signal a need for stricter enforcement.

Table 1.

Broad definitions of ‘medical device’ are similar across different regions

Region General regulatory pathway for medical devices
(AI-specific guidance/legislation)
Medical device definition Risk classification and regulatory controls Does an LLM-based clinical documentation scribe meet criteria as a medical device?
USA

Food, Drug, and Cosmetic Act

(FDA AI/ML-based SaMD Action Plan)

Software, when intended to treat, diagnose, cure, mitigate, or prevent disease or other conditions, are medical device under the FD&C Act, and is called SaMD

Software functions intended (1) for administrative support of a health care facility, (2) for maintaining or encouraging a healthy lifestyle, (3) to serve as electronic patient records, (4) for transferring, storing, converting formats, or displaying data, or (5) to provide certain, limited clinical decision support are not medical devices and are not subject to FDA regulation

Class I (low-risk)

Class II (medium-risk)

Class III (high-risk)

Class I and II devices are often cleared through the 510(k) pathway, in which applicants submit a premarket notification to demonstrate safety and effectiveness and/or substantial equivalence of their proposed device to a predicate device.

Class III (high-risk) devices are defined as those that support human life, prevent impairment of health, or pose potential unreasonable health risk(s)

No.
UK

Medicines and Medical Device Act

(MHRA: software and AI as a Medical Device Change Program, Regulation Horizons Council: the Regulation of AI as a Medical Device)

An article, instrument, apparatus or machine that is used in the prevention, diagnosis or treatment of illness or disease, or for detecting, measuring, restoring, correcting or modifying the structure or function of the body for some health purpose

Decision support software is usually considered a medical device when it applies automated reasoning, such as a simple calculation, an algorithm or a more complex series of calculations

Class I (low-risk)

Class II (medium-risk)

Class III (high-risk)

Most SaMD and AIaMD are Class II or Class III devices. Conformity assessments are required prior to receiving the UKCA (UK) mark.

Probable.
Europe

MDR

(EU AI Act)

Software intended for the following medical purposes: diagnosis, prevention, monitoring, prediction, prognosis, treatment or alleviation of disease, diagnosis, monitoring, treatment, alleviation of, or compensation for, an injury or disability, investigation, replacement or modification of the anatomy or of a physiological or pathological process or state, providing information by means of in examination of specimens derived from the human body, including organ, blood and tissue donations, and which does not achieve its principal intended action by pharmacological, immunological or metabolic means, in or on the human body, but which may be assisted in its function by such means

Class I (low-risk)

Class IIa (medium-risk)

Class IIb (medium-risk)

Class III (high-risk)

AIaMD would be classified following Rule 11 of the MDR (classification rule for SaMD).

AIaMD are also subject to the EU AI Act, and if Class IIa or higher, are considered as high-risk AI systems under Article 6. Some healthcare-related AI systems that do not qualify as medical devices fall under some provisions (e.g. transparency) of the EU AI Act, but are not generally subject to conformity assessment procedures.

Possible.

However, regulations for AI-based medical devices and software as medical devices display nuanced differences. Risk classification differs across different regions, revealing a lack of international consensus.

Unlike predictive models, LLM outputs are non-deterministic in nature, even when the model temperature is set to zero. The frequency of hallucinations by LLMs under different conditions of use is poorly quantified and requires further examination and documentation.

Lack of robust evaluation for LLM performance

LLM evaluation faces significant challenges as a result of the unique capabilities of LLMs in open-ended output generation. Unlike diagnostic tests, where accuracy can be assessed, or interventional trials, where causal effects can be quantified, the evaluation of LLMs often relies on subjective and largely qualitative assessments. Multiple studies have evaluated the performance of LLMs on isolated medical tasks such as information extraction from clinical notes or answering patient enquiries7,8. The performance of LLMs in such tasks has largely been tested in structurally simple case vignettes, medical licensing exams or medical question and answer datasets7. However, there is increasing recognition that the strong performance of LLMs in these tasks is not reflective of these models’ performance in real-world clinical decision making. Hager et al. evaluated leading open-access LLMs using a comprehensive evaluation framework. LLMs were assessed for their performance in making diagnoses, gathering information and providing guideline adherence recommendations using real-world cases9. The results revealed significant limitations of LLMs, with all models performing worse than clinicians in making diagnoses, interpreting laboratory results and adherence with established treatment guidelines. On the contrary, a study introducing the Sequential Diagnosis Benchmark (SDBench) found that an LLM-based diagnostic orchestrator outperformed clinicians in diagnostic accuracy across complex case vignettes. This highlights the need for standardisation of testing methods and benchmarking10.

Robust evaluation of LLM bias is similarly critical to surface risks that may perpetuate health inequities. Before LLMs can be approved and used widely in clinical settings, contextualised evaluations with the setting up of suitable guardrails are necessary to ensure patient safety. Various approaches to LLM bias evaluation have been proposed by researchers. One approach is to rely on physician evaluation of different dimensions of LLM bias, including inaccuracy across axes of identity, lack of inclusion, stereotypical language or characterisation, omission of structural explanations for inequity, failure to challenge a biased premise and potential for disproportionate withholding of opportunities or resources11,12. Another approach involves adversarial testing to probe for different failure modes and conditions whereby LLMs will generate harmful and biased outputs11. Convergence among regulators, manufacturers and researchers is necessary to develop evaluation methodology and metrics that are clinically relevant. The complexity of developing benchmarks depends on the deployment scenario of the LLM or LLM-based application. It could be easier for applications and models with narrow clinical use, while it would be more challenging for applications with a broad intended use13.

Challenges to monitoring and regulatory enforcement

The training of LLMs on vast and diverse datasets presents challenges in monitoring and enforcement, in particular concerning data provenance. LLM models are often trained on vast, diverse, and inconsistently documented datasets, making it difficult for regulators to trace the origins and licensing of the data used. This complicates efforts to ensure compliance with regulatory standards, such as those governing data privacy, consent, and intellectual property rights. A recent large-scale audit of over 1,800 text datasets on Hugging Face, an open-source collaborative platform for datasets and ML tools, revealed frequent misclassifications. License omission was reported in 70% of datasets, and errors were detected in 50% of listed licences. This highlights a problem in the misattribution of datasets and informed use in AI and LLM development14. In addition, the open-sourced models allow modifications and retraining by third parties to be performed freely, further obscuring data provenance and undermining data authenticity. Addressing these challenges necessitates the development of robust frameworks for data auditing and provenance tracking to enable more effective oversight of LLM-based tools.

Post-marketing surveillance of authorised LLM-based tools has similar challenges to monitoring the effects of medicinal products post licensure15,16. First, signal detection of public health risks is very challenging given the pervasiveness of LLM tools with diverse applications and uses. Reliance on voluntary reporting of adverse events by clinicians may lead to potential under-reporting and delays in signal detection, especially when risks are not fully known17–19. LLM-based tools, such as Ambient AI medical scribes, heighten the challenge of identifying potential adverse events attributed to the use of these tools. Unedited “hallucinations” or inappropriate redaction of pertinent information may be carried forward to subsequent clinical notes, resulting in errors of unknown provenance. Second, the inability to track “off-labelled” use of domain-specific LLMs, for example, when applying the models for unapproved indications and/or patient populations17. Finally, high cost and extensive resources are required to develop a robust vigilance program due to the diverse nature of LLM applications and the unknown profile of risk to patients.

510(k) clearance process poses additional challenges for monitoring programs. This process for medical devices is based on “substantial equivalence” to devices cleared pre-1976 or legally marketed thereafter, known as predicate devices20. A ‘predicate creep’ has been described as a cycle of technology change through repeated clearance of devices based on predicates with slightly different technological characteristics. While the 510(k) pathway has facilitated faster clearance of devices, these may have referenced multiple iterations of predicates, resulting in devices with significant differences in features and deviating from the design intent of the first rigorously reviewed and approved reference model; this could pose risks to patient safety21,22. Depending on risk assessment, some GenAI tools may be approved via the 510(k) pathway on the premise of ‘low-risk, design iterations’ built upon the same pre-trained foundation models, such as fine-tuning or employing a retrieval-augmented generation framework. These ‘design iterations’ enhance performance but may also significantly change the nature of the model and therefore its risks23. From the perspective of the user, even variations in user prompts may significantly impact the reproducibility and reliability of responses from LLMs24,25. Some models have been cleared with predicates that were not ML/DL (deep learning) devices. In a review of AI/ML-based medical devices that have been cleared through the 510(k) pathway, up to one-third of devices were approved on the basis of non-AI/ML-based medical devices in the first generation20. In addition, safety concerns were raised for AI/ML tasks that changed frequently along a device’s predictive network, whereby the intended use differs significantly from earlier generations.

The rising pervasiveness of the technology makes enforcement challenging. Growing evidence suggests that open-sourced LLMs are fast approaching the performance of proprietary LLMs in medical tasks26,27. As a result of dramatic cost reductions per token of proprietary LLMs such as GPT-4o by OpenAI28, the access barrier is lowered, and we anticipate widespread adoption of LLMs for both medical and non-medical purposes in the absence of formal approval and/or robust evidence, posing challenges for enforcement. Developers are releasing unapproved, consumer-facing LLM-based applications that provide medical advice, even though these tools may meet the criteria for medical devices under EU and US regulations. This highlights a critical gap in regulatory oversight and safety evaluation10,29.

Ethical considerations beyond current regulatory oversight

The widespread availability and rapid adoption of LLM tools, either directly applied to or adapted for healthcare, are raising numerous ethical dilemmas and posing potential risks to the broader public. Risks that remain inadequately addressed include low trust of LLM-based health applications due to risks of hallucinations and poor reproducibility of output30–32; poorer performance of LLMs in under-represented populations, exacerbating health inequities33,34; and bioethical concerns such as patient privacy arising from the use of healthcare data in model pre-training35. Despite data protection laws, bad actors may rush to market with unapproved solutions without ensuring robust data protection36. Despite its promise, current evidence is still lacking on the impact of LLMs on patient outcomes. In addition, the impact of LLM applications on population health and minority populations is still largely unknown.

Regulatory guidance for responsible and ethical LLM use is critical. The issue of bias being demonstrated or exacerbated by AI algorithms is an increasing concern within the healthcare industry. Several initiatives, such as the STANDING Together initiative (standards for data diversity, inclusivity, and generalizability)37 and the FDA’s “Artificial Intelligence/ML (AI/ML)-Based SaMD Action Plan”, serve to bridge the gaps inherent in existing research and regulatory standards/guidance. However, there has been little formal consideration of the impact of patient interactions with AI programs, including the impact of LLM medical tools on patient autonomy, dignity, the patient-clinician relationship and trust38. Conversational applications (chatbots) powered by LLMs allow dynamic, context-aware exchanges in patient education. However, LLM interactions are nuanced, thus model fine-tuning is typically needed, and requires post launch monitoring and policy optimisation to reduce risks of embedded cognitive biases such as confirmation bias, automation bias and automation complacency, as well as avoiding model drift. Figure 1 shows different phases of TPLC, and overall regulatory considerations specific to LLMs.

Fig. 1. An illustration of the different phases of the total product life cycle for LLM-based medical devices.

Fig. 1

This figure illustrates the TPLC of an LLM-based medical device, comprising different phases from planning and design to real-world performance evaluations. However, LLMs pose unique considerations and regulatory challenges at each phase, such as a lack of standardised evaluation metrics to assess performance.

To strengthen data quality, transparency, and patient privacy in LLM development and deployment, federated learning and blockchain technologies may play a future role in enabling secure, privacy-preserving data collaboration across institutions39. Federated learning distributes model training and aggregates weights without sharing individual patient records, allowing LLMs to benefit from diverse datasets while keeping sensitive data on-site, with novel low-parameter federated approaches further reducing communication overhead and enabling few-shot prompt learning40. Blockchain adds an immutable, auditable layer for exchanging tokenised health data, safeguarding data integrity and patient autonomy, and even enabling health-data non-fungible tokens for provenance verification41. By integrating these methods, stakeholders can uphold rigorous standards of data provenance, auditability, and security that are critical for regulated, high-stakes environments.

Developing effective regulatory policies which take into consideration the vast number of ethical principles is challenging. Interpretation of ethical principles can be highly subjective, and may vary in different contexts of LLM application. In addition, most ethical guidance offers little assistance when trade-offs need to be made between ethical principles. Collaboration across disciplines is now mandatory, bringing together bioethicists, regulators, end-users, including healthcare providers and patients, legal experts and developers. Equipping regulatory authorities and healthcare professionals through interdisciplinary training or capacity-building programs is essential to ensure they are well-prepared to manage and engage with AI technologies effectively.

Regulatory science in LLM innovation—opportunities and trends

Regulatory science is the science of developing new tools, standards, and approaches to assess the safety, efficacy, quality, and performance of health products that are assessed by regulatory agencies42. Regulatory authorities such as the FDA, European Medicines Agency (EMA), Australian Therapeutic Goods Administration (TGA) and the Singapore Health Sciences Authority (HSA) conduct research or partner with regulatory research groups in the development of guidance material. Such regulatory collaborative groups include Centres of Excellence in Regulatory Science and Innovation in the United States, Centres for Regulatory Science and Innovation in the United Kingdom, the Duke-NUS Centre of Regulatory Excellence in Singapore, and at a global level, the Global Coalition for Regulatory Science Research and the World Health Organisation (WHO). In this section, we discuss the priorities of regulatory science research in advancing the goals of responsible LLM innovation and balanced regulation.

Applying adaptive regulatory approaches

Adaptive regulatory approaches may be adopted for policies characterised by a fast pace of innovation. LLM-based tools are prime candidates for this regulatory approach and could catalyse advances in this area, commanding the attention of regulators and industry stakeholders to keep up to date with growing scientific evidence of risks and benefits of the technology43. Adaptive approaches are those that empower regulators to be less restrictive in the absence of negative outcomes demonstrated in clinical trials, and reinstate restrictions should evidence of harm emerge in the process of post-marketing surveillance44. For example, accelerated approval (in the US) and conditional marketing authorisation/approval (in the EU, Singapore and Japan) pathways are in place for emerging therapeutics45. The predetermined change control plan proposal put together by FDA, Health Canada and MHRA supplements the TPLC, allowing certain changes that were predicted and predefined in a submitted plan, to be made without requiring another approval process; these changes could be a result of information from real-world use. A more flexible approach is an anticipatory regulatory approach, whereby regulatory rules are iteratively developed alongside the development of new products or services. Regulatory sandboxes are outcomes-oriented tools used to guide anticipatory regulation, whereby new services, health products or digital health tools can be tested in a constrained environment with fewer regulatory requirements. Examples include the United Kingdom’s MHRA AI Airlock and Singapore’s Infocomm Media Development Authority (IMDA) Privacy Enhancing Technology Regulatory Sandbox46,47. Similarly, the EU AI Act calls for establishing regulatory sandboxes for improving regulatory compliance with the act (detailed in Article 57 of the EU AI Act). These sandboxes allow developers and researchers to work closely with regulators to identify, quantify and prospectively develop solutions to mitigate potential risks in early stages.

Regulatory authorities can work with local and global regulatory research groups in the conceptualisation, conduct and assessment of outcomes from such regulatory sandboxes. Such a setup enables regulatory experimentation to gain a rapid understanding of the risks and benefits of different frameworks and policies. The main characteristics of regulatory sandboxes include the following: (1) they are adaptive; (2) they require close collaboration and iteration between stakeholders such as users, regulators and industry; and (3) they are evidence-based, with a trial and error approach46. “Global regulatory sandboxes” can be developed to study the international interoperability of regulatory policies and the impact of policies on cross-border innovation and competition.

Convergence of regulatory frameworks: medical device and health services

AI agents powered by LLMs and agentic systems show promise in enabling a greater level of automation in clinical tasks48. AI agents in healthcare are intelligent systems designed to assist, automate, and enhance various aspects of medical care. Agentic systems have demonstrated improved capabilities over single LLM models in medical information processing, planning and making clinical or operational decisions through interacting and collaborating among different agents48. These capabilities are shifting LLM agents from being pure assistive tools into service agents with elevated levels of autonomy, much like a skilled worker performing a service. The development of AI agents powered by LLMs, in particular agents with reasoning capabilities49, is introducing a paradigm shift. For example, OpenAI o1 model was designed to answer complex questions through a series of intermediate reasoning steps, a strategy known as “chain-of-thought”50. The scaling of automation beyond repetitive, low-skilled tasks lends to a ‘Software as a Medical Service’ (SaMS) model. As opposed to a tangible product with defined functionalities, SaMS is focused on service-oriented delivery, much like a digital health service. In Singapore, health services regulation is coming under the ambit of HSA, which had previously only regulated health products; this will result in greater convergence of products and services regulation in future, providing a more optimal framework for regulating healthcare AI51. Such novel regulatory frameworks will need to be co-developed with regulatory science groups and studied before LLM agents and systems can be adopted at scale.

Strengthening resiliency in AI service delivery

The recent global IT outage incident that started with a single software update by the cybersecurity company CrowdStrike is a prominent example of how a single failure point in the digital delivery system can lead to a global disruption of critical health services, including individual products52. Widespread digitisation, adoption of electronic medical records, mHealth and proliferation of AI-based tools add to our dependency on digital solutions and growing complexity, which cannot afford any disruptive downtime. Beyond predictive, diagnostic and decision support algorithms, it is important to recognise that AI is likely to be pervasive in operating systems, software, and even embedded within hardware such as computer chips53. Ensuring continuity of care delivered to patients requires a scrutiny of vulnerabilities in the use of AI across the process of provision of healthcare delivery, both in the business and service value chain. Figure 2 shows the healthcare AI supply chain and examples of governance and guidance frameworks.

Fig. 2. The healthcare AI supply chain.

Fig. 2

Data, model (or product), IT implementation capabilities and infrastructure and hardware integration all form parts of the healthcare AI supply chain. Different regulations, framework and governing bodies oversee parts, if not all segments of the supply chain.

A global collaborative effort in regulatory science enhances and strengthens the following: internal and global visibility on various components of the supply chain; frameworks to guide identification, testing and validation of critical components and pipelines; sharing good practice for resilience teams and business continuity plans (BCP); facilitating communications strategy across supply chain networks. We also propose this global effort to support health systems in building AI development capabilities, and developing infrastructure for local deployment (on-premises or edge) of these models. This serves as a fail-safe strategy, reducing dependency on commercially available, leading technologies. In addition, such deployment strategies improve the availability of services and lower dependence on commercial cloud-based products. This approach could lower privacy risks while having its own cybersecurity challenges. There is a need to ensure that capacity building and participation in a global collaborative effort are not limited to researchers from high-income countries and representing top-tier research institutions30. Along the same line, continued efforts to improve open-sourced models (e.g. open-sourced LLMs) align with commercial competition laws54. Government policies and health system investments need to align with these strategies, emphasising diversity in technology investments to build a more resilient healthcare AI infrastructure.

Future directions for regulatory science and regulators

Beyond MDR: responsible AI in health product development

AI has significant potential to influence various aspects of medical product development, and this impact is already in progress. The drug development industry has undergone a revolution with the introduction of computational approaches, including deep learning, foundation models and quantum computing55. Deep GenAI has shown huge promise in de novo molecular design and early development stages, e.g. predict protein folding, molecular interactions, and cellular disease processes56,57. ADMET (absorption, distribution, metabolism, elimination and toxicity) properties of small molecules can now be predicted with a high degree of accuracy58. As these AI-designed targets move into clinical testing, the industry is embracing another revolution in the clinical research process to accelerate and streamline the process of clinical trials. In particular, the advent of LLMs has brought about rapid advancements in AI-aided clinical trial design, participation selection and recruitment, as well as tracking and promoting adherence to trial protocols4. Such applications present immense promise in shortening the cost and development time for new drug entities. However, unknown risks associated with using the technology to augment or automate critical processes warrant close scrutiny by the research community and regulators.

We believe that LLMs can bring about significant value in reducing administrative burden within regulatory agencies, and increase sensitivity and reliability of post-marketing surveillance and evaluation. Studies suggest promising performance in early identification and classification of adverse drug events and automation of drug approval processes59. Such tools are of low clinical risk, but bring about significant returns in reducing documentation burden and turnaround time for new drug applications. More funding and collective support for research in this area will bring about benefits in accelerating the time to market.

Global call to action

The International Medical Device Regulators Forum (IMDRF) is an example of a global effort to promote harmonisation of MDRs, including ML/AI-enabled devices60. The committee is represented by regulatory authority representatives from different jurisdictions, including Australia’s TGA, Health Canada, Singapore’s HSA, the EU’s European Commission - Directorate-General for Health and Food Safety and the US FDA. Global regulatory research organisations, such as the Global Coalition for Regulatory Science Research and the Global Network for Regulatory Science, are pivotal in advancing regulatory research through fostering international collaboration. By engaging with industry and regulatory stakeholders from various regions, they play a crucial role in establishing cross-industry, cross-national conversations, global harmonization and drive outcomes research to facilitate the formulation of regulatory guidelines for the responsible adoption of AI and LLMs. The worldwide, interconnected nature of digital health and AI technologies underscores the importance of adopting a global perspective and collaboratively addressing pressing issues at hand.

One such pressing issue is the lack of robust standardisation and harmonisation for LLM-based tools. A recent literature review of 142 LLM studies in medical applications revealed significant gaps in overall reliability, generalizability, and applicability of model evaluation practices61. In a separate systematic review concerning current AI reporting and evaluation methods, 220 distinct items were recommended for detailing model performance or describing source data characteristics through “model cards” and “data cards”62. “Model cards” encourage transparent reporting of AI model performance characteristics while “data cards” describe the diversity of AI model training datasets63. Several groups have attempted to address this gap, notably CONSORT-AI, SPIRIT-AI, TRIPOD-AI, PROBAST+AI, MI-CLAIM, and, most recently, TRIPOD-LLM for transparent reporting of LLMs in scientific publications64–66. The TREGAI and CARE-AI checklists address ethical reporting of AI models67,68 while health AI assurance labs such as CHAI, VALID-AI and HAIP69,70 facilitate private-public partnerships and boost collaborative governance of medical AI models at a local level69,71. As these different checklists and governance groups start to form, alignment in directions and consolidation of functions will be critical. The Partnership for Oversight, Leadership, and Accountability in Regulating Intelligent Systems–Generative Models in Medicine (POLARIS-GM) is one such initiative72. Global regulatory research groups can take on such a function, establishing common grounds for conversations among different stakeholders while allowing room for localisation of regulations.

Advancing the collective goals of health equity

Global divides between high-income countries (HICs) and low- and middle-income countries (LMICs) are leading to disturbing health inequities73,74. Regulatory bodies and international organisations are taking actions to push health equity to the forefront in regulatory strategies, e.g. the World Health Organisation's global Guidance for best practices for clinical trials to promote equitable clinical trials, and the FDA’s Health Care at Home Initiative to drive equity in digital medicine75. LLMs have the potential to both bridge and exacerbate health inequities. Potential harms can arise if social and structural health determinants, geographic, linguistic, and demographic biases in datasets76, misconceptions tied to patient identity, prioritisation of privileged perspectives, and systemic disparities in system performance and accessibility across populations are not appropriately addressed8. There is also a need to identify and understand the various nuanced failure modes of LLM tools when implemented at scale, to allow health delivery systems to devise more effective real-time monitoring programs. For example, repeated testing (conducted over a thousand times) of LLMs using hypothetical clinical scenarios revealed clear gender and racial biases36. These may not be apparent when tested in a clinical trial, but emerge as critical public health concerns when deployed in a real-world setting. LLMs and GenAI risk perpetuating biases in clinical applications at the production scale, with biases accumulating throughout the data processing and modelling process.

Global regulatory science groups can advance the goals of health equity beyond country borders and geopolitical regions. Intentional inclusion of LMIC perspectives in global discussions to tailor AI solutions that address specific regional health challenges and resource constraints. The goal of such efforts is to democratize resources and extend support to regions and ecosystems with limited access. Knowledge and resource transfer enable accessible and affordable digital and AI interventions that are safe, and effective. Translational gaps exist for innovative technologies in LMICs, primarily arising from complex economic, infrastructural, and political barriers77. Economic and political interests from industry, nongovernment organizations and private funders within LMICs may drive agendas that influence regulatory decisions. AI-supported screening programs for chronic diseases such as diabetic retinopathy (DR) have reported varying degrees of success in low resource settings78. Various implementation and adoption barriers were reported, such as the inability to provide timely reports and lack of patient adherence to program recommendations79,80. Research on LLMs in LMIC has demonstrated potential benefits. In one prospective study, patients randomised to the intervention arm received diabetes self-management recommendations from primary care physicians assisted by an LLM integrated tool81. Patients in the intervention group showed better self-management behaviour compared to patients in the control group, where primary care physicians were unassisted. However, the authors' note raised the need to address various implementation challenges, including data quality and barriers to integration with existing healthcare information technology infrastructure.

Conclusion

As GenAI and LLM continue to transform healthcare and diversify its applications, there is a need to recognise current challenges and uncertainties, and develop regulatory policies to address these concerns. There will be a paradigm shift in how we view the TPLC approach, and a need for an innovative and responsive regulatory framework. A global independent body can establish voluntary frameworks that evolve into trusted industry standards, in a similar fashion to NIST functions in cybersecurity. These standards guide vendor onboarding through clear due diligence processes, helping build customer trust, support product adoption, and sustain long-term relationships.

Acknowledgements

There was no funding for this work.

Author contributions

Author J.C.L.O. conceptualised the paper, performed formal analysis and wrote the original draft. Author N.L. conceptualised the paper, reviewed and edited the manuscript and provided overall supervision. Authors Y.N., M.L., Y.M., L.Z., K.S., R.T.C., S.V., J.C.W.L., I.S.K.T., O.F., S.G., D.S.B., X.L. and A.K.D. contributed to the formal analysis of the paper, reviewed and edited the manuscript. All authors have read and approved the final manuscript.

Data availability

No datasets were generated or analysed during the current study.

Competing interests

Authors Y.N., L.Z., K.S., R.T.C., S.V., J.C.W.L., I.S.K.T., O.F., D.S.B., X.L. and A.K.D. declare no financial or non-financial competing interests. Author J.C.L.O. serves as associate editor of this journal and had no role in the peer-review or decision to publish this manuscript. Author J.C.L.O. declares no financial competing interests. Author S.G. serves as an editor of this journal and had no role in the peer-review or decision to publish this manuscript. Author S.G. declares no financial competing interests. Author N.L. serves as an editorial board member of this journal and had no role in the peer-review or decision to publish this manuscript. Author N.L. declares no financial competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

  • 1.Ng, M. Y., Kapur, S., Blizinsky, K. D. & Hernandez-Boussard, T. The AI life cycle: a holistic approach to creating ethical AI for health decisions. Nat. Med.28, 2247–2249 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Wellnhofer, E. Real-world and regulatory perspectives of artificial intelligence in cardiovascular imaging. Front. Cardiovasc. Med.9, 890809 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Food and Drug Administration, U.S.A. Total product lifecycle considerations for generative AI-enabled devices. in Executive Summary for the Digital Health Advisory Committee (FDA, 2024); https://www.fda.gov/media/182871/download.
  • 4.Warraich, H. J., Tazbaz, T. & Califf, R. M. FDA perspective on the regulation of artificial intelligence in health care and biomedicine. JAMA333, 241–247 (2024). [DOI] [PubMed] [Google Scholar]
  • 5.Ong, J. C. L. et al. Ethical and regulatory challenges of large language models in medicine. Lancet Digit. Health6, e428–e432 (2024). [DOI] [PubMed] [Google Scholar]
  • 6.Aboy, M., Minssen, T. & Vayena, E. Navigating the EU AI act: implications for regulated digital medical products. NPJ Digit. Med.7, 237 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Bedi, S. et al. Testing and evaluation of health care applications of large language models: a systematic review. JAMA333, 319–328 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Sandmann, S. et al. Benchmark evaluation of DeepSeek large language models in clinical decision-making. Nat. Med.31, 2546–2549 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Hager, P. et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat. Med.30, 2613–2622 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Nori, H. et al. Sequential diagnosis with language models. Preprint https://arxiv.org/abs/2506.22405 (2025).
  • 11.Pfohl, S. R. et al. A toolbox for surfacing health equity harms and biases in large language models. Nat. Med.30, 3590–3600 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Alber, D. A. et al. Medical large language models are vulnerable to data-poisoning attacks. Nat. Med.31, 1–9 (2025). [DOI] [PMC free article] [PubMed]
  • 13.Freyer, O., Wiest, I. C. & Gilbert, S. Policing the boundary between responsible and irresponsible placing on the market of LLM health applications. Mayo Clin. Proc. Digit. Health3, 100196 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Longpre, S. et al. A large-scale audit of dataset licensing and attribution in AI. Nat. Mach. Intell.6, 975–987 (2024). [Google Scholar]
  • 15.Peter, J. E. Algorithmovigilance-advancing methods to analyze and monitor artificial intelligence-driven health care for effectiveness and equity. JAMA Netw. Open4, e214622 (2021). [DOI] [PubMed] [Google Scholar]
  • 16.Alan, B., Mehdi, B., Theodoros, E. & Philippe, R. Algorithmovigilance, lessons from pharmacovigilance. NPJ Digit. Med.7, 270 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Vandana, A. Off-label medication use: a double-edged sword. Indian J. Crit. Care Med.25, 845–846 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Caterina, P. et al. Limitations and obstacles of the spontaneous adverse drugs reactions reporting: two “challenging” case reports. J. Pharmacol. Pharmacother4, S66–S72 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Harmark, L. & van Grootheest, A. C. Pharmacovigilance: methods, recent developments and future perspectives. Eur. J. Clin. Pharmacol.64, 743–752 (2008). [DOI] [PubMed] [Google Scholar]
  • 20.Atul, K. Pharmacovigilance: importance, concepts, and processes. Am. J. Health Syst. Pharm.74, 606–612 (2017). [DOI] [PubMed] [Google Scholar]
  • 21.Urs, J. M., Paola, D. & Kerstin, N. V. Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015-20): a comparative analysis. Lancet Digit. Health3, e195–e203 (2021). [DOI] [PubMed] [Google Scholar]
  • 22.Charlotte, L. & Sandra, R. Identification of predicate creep under the 510(k) process: a case study of a robotic surgical device. PLoS One18, e0283442 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Urs, J. M., Christain, B. & Kerstin, N. V. FDA-cleared artificial intelligence and machine learning-based medical devices and their 510(k) predicate networks. Lancet Digit. Health5, e618–e626 (2023). [DOI] [PubMed] [Google Scholar]
  • 24.Kushal, T. et al. Use of recalled devices in new device authorizations under the US Food and Drug Administration’s 510(k) pathway and risk of subsequent recalls. JAMA329, 136–143 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Li, W. et al. Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs. npj Digit. Med.7, 41 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Bertalan, M. Prompt engineering as an important emerging skill for medical professionals: tutorial. J. Med. Internet Res.25, e50638 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Ke, Y. H. et al. Retrieval augmented generation for 10 large language models and its generalizability in assessing medical fitness. Npj Digit. Med.8, 187 (2025). [DOI] [PMC free article] [PubMed]
  • 28.Ong, J. C. L. et al. Large language model as clinical decision support system augments medicationsafety in 16 clinical specialties. Cell Rep. Med.6, 102323 (2025). [DOI] [PMC free article] [PubMed]
  • 29.Ng A. DeepLearning.AI. Falling LLM Token Prices and What They Mean for AI Companieshttps://www.deeplearning.ai/the-batch/falling-llm-token-prices-and-what-they-mean-for-ai-companies/ (2024).
  • 30.Freyer, O. et al. A future role for health applications of large language models depends on regulators enforcing safety standards. Lancet Digit. Health6, e662–e672 (2024). [DOI] [PubMed] [Google Scholar]
  • 31.Gongbo, Z. et al. Closing the gap between open source and commercial large language models for medical evidence summarization. NPJ Digit. Med.7, 239 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Wu, S. et al. Benchmarking open-source large language models, GPT-4 and Claude 2 on multiple-choice questions in nephrology. NEJM AI1, 2 (2024).
  • 33.Ong, J. C. L. et al. Medical ethics of large language models in medicine. NEJM AI. 1, 10.1056/AIra2400038 (2024).
  • 34.Omiye, J. A., Lester, J. C., Spichak, S., Rotemberg, V. & Daneshjou, R. Large language models propagate race-based medicine. NPJ Digit. Med.6, 1–4 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Soroush, A. et al. Large language models are poor medical coders—benchmarking of medical code querying. NEJM AI1, 5 (2024).
  • 36.Zack, T. et al. Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study. Lancet Digit. Health6, e12–e22 (2024). [DOI] [PubMed] [Google Scholar]
  • 37.Danielle, S. B., Hugo, J. W. L. A. & Raymond, H. M. Approaching autonomy in medical artificial intelligence. Lancet Digit. Health2, e447–e449 (2020). [DOI] [PubMed] [Google Scholar]
  • 38.Shaswath, G. et al. Tackling bias in AI health datasets through the STANDING together initiative. Nat. Med.28, 2232–2233 (2022). [DOI] [PubMed] [Google Scholar]
  • 39.Mirja, M., Marium M. R. & Joseph C. K. Bias in AI-based models for medical applications: challenges and mitigation strategies. NPJ Digit. Med.6, 113 (2023). [DOI] [PMC free article] [PubMed]
  • 40.Teo, Z. L. & Ting, D. S. W. Federated learning in ophthalmology: retinopathy of prematurity. Ophthalmol. Retin.6, 647–649 (2022). [DOI] [PubMed] [Google Scholar]
  • 41.Jiang, J., Liu, X. & Fan, C. Low-parameter federated learning with large language models. Web Information Systems and Applications: 21st International Conference, WISA, 2024Yinchuan, China 319–330 (2024).
  • 42.Tan, T. E. et al. Retinal photograph-based deep learning algorithms for myopia and a blockchain platform to facilitate artificial intelligence medical research: a retrospective multicohort study. Lancet Digit. Health3(5), e317–e329 (2021). [DOI] [PubMed] [Google Scholar]
  • 43.Food and Drug Administration. U.S.A. Focus Areas of Regulatory Science—Introduction (FDA, 2024); https://www.fda.gov/science-research/focus-areas-regulatory-science-report/focus-areas-regulatory-science-introduction
  • 44.Grandis, G. D., Brass, I. & Farid, S. S. Is regulatory innovation fit for purpose? A case study of adaptive regulation for advanced biotherapeutics. Regul. Govern.17, 810–832 (2023).
  • 45.Eichler, H. G. et al. From adaptive licensing to adaptive pathways: delivering a flexible life-span approach to bring new drugs to patients. Clin. Pharmacol. Ther.97, 234–246 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Emily, L., Dalia, D., Jacoline, B. & Pall, J. The sandbox approach and its potential for use in health technology assessment: a literature review. Appl. Health Econ. Health Policy19, 857–869 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Infocomm Media Development Authority, Singapore. Privacy Enhancing Technology Sandboxes—Infocomm Media Development Authorityhttps://www.imda.gov.sg/how-we-can-help/data-innovation/privacy-enhancing-technology-sandboxes (2024).
  • 48.OECD. Regulatory Sandboxes in Artificial Intelligence, Vol. 356 (OECD Digital Economy Papers, 2023).
  • 49.Qiu, J. et al. LLM-based agentic systems in medicine and healthcare. Nat. Mach. Intell.6, 1418–1420 (2024). [Google Scholar]
  • 50.Mehandru, N. et al. Evaluating large language models as agents in the clinic. NPJ Digit. Med.7, 1–3 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.OpenAI. Learning to Reason with LLMs. Vol. 16 December 2024 (https://openai.com/index/learning-to-reason-with-llms/, September 12, 2024).
  • 52.Lim J. C. W., Tan Koi W. C. & Vogel, S. Regulating the Future of Health: CoRE’s Tenth Anniversary Perspective. in CoRE Regulatory Perspective, Vol. 2024 (Duke-NUS Medical School, Centre of Regulatory Excellence, 2024); https://www.duke-nus.edu.sg/core/think-tank/core-regulatory-perspective/regulating-the-future-of-health-core-s-tenth-anniversary-perspective
  • 53.Association, A. H. CrowdStrike Technology Outage Causing Global Disruption, Including Impacts on Hospitals (AHA, 2024); https://www.aha.org/advisory/2024-07-19-crowdstrike-technology-outage-causing-global-disruption-across-industries-including-impacts-hospitals-and
  • 54.Biglari, A. & Wei, T. A review of embedded machine learning based on hardware, application, and sensing scheme. Sensors4, 2131 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 55.Riedemann, L., Labonne, M. & Gilbert, S. The path forward for large language models in medicine is open. NPJ Digit. Med.7, 339 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 56.Chiranjib, C., Manojit, B. & Sang-Soo, L. Artificial intelligence enabled ChatGPT and large language models in drug target discovery, drug discovery, and development. Mol. Ther. Nucleic Acids33, 866–868 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57.Druedahl, L. C. et al. Use of artificial intelligence in drug development. JAMA Network Open7, e2414139 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58.Amit, G. & Antonio, L. Unleashing the power of generative AI in drug discovery. Drug Discov. Today29, 103992 (2024). [DOI] [PubMed] [Google Scholar]
  • 59.Swanson, K. et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scalechemical libraries. Bioinformatics40, btae416 (2024). [DOI] [PMC free article] [PubMed]
  • 60.Ong, J. C. L. et al. A scoping review on generative AI and large language models in mitigating medication related harm. NPJ Digit. Med.8, 182 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 61.International Medical Device Regulators Forum. Working Groups: Artificial Intelligence/Machine Learning-enabled, Vol. 2024 (IMDRF, 2024); https://www.imdrf.org/working-groups/artificial-intelligencemachine-learning-enabled
  • 62.Thomas, Y. C. T. et al. A framework for human evaluation of large language models in healthcare derived from literature review. NPJ Digit. Med.7, 258 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 63.Jonathan, H. L. et al. Assessment of adherence to reporting guidelines by commonly used clinical prediction models from a single vendor: a systematic review. JAMA Netw. open5, e2227779 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 64.Anmol, A. et al. The value of standards for health datasets in artificial intelligence-based applications. Nat. Med.29, 2929–2938 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 65.Gallifant, J. et al. The TRIPOD-LLM reporting guideline for studies using large language models. Nat. Med.31, 60–69 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 66.Cruz Rivera, S. et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat. Med.26, 1351–1363 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 67.Liu, X. et al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat. Med.26, 1364–1374 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 68.Yilin, N. et al. Generative artificial intelligence and ethical considerations in health care: a scoping review and ethics checklist. Lancet. Digit. Health6, e848–e856 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 69.Yilin, N. et al. An ethics assessment tool for artificial intelligence implementation in healthcare: CARE-AI. Nat. Med.30, 3038–3039 (2024). [DOI] [PubMed] [Google Scholar]
  • 70.Nigam, H. S. et al. A nationwide network of health AI assurance laboratories. JAMA331, 245–249 (2024). [DOI] [PubMed] [Google Scholar]
  • 71.Swanson, T. UC Davis Health: Health and Technology. UC Davis Health, NODE.health, and Leading Health Systems launch VALID AIhttps://health.ucdavis.edu/news/headlines/uc-davis-health-and-leading-health-systems-launch-valid-ai/2023/10 (2023).
  • 72.Ong, J. C. L. et al. International partnership for governing generative artificial intelligence models in medicine. Nat. Med.31, 2836–2839 (2025). [DOI] [PubMed] [Google Scholar]
  • 73.Price, W. N., Sendak, M., Balu, S. & Singh, K. Enabling collaborative governance of medical AI. Nat. Mach. Intell.5, 821–823 (2023). [Google Scholar]
  • 74.Rui, Y. et al. Disparities in clinical studies of AI enabled applications from a global perspective. NPJ Digit. Med.7, 209 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 75.Michael, D. et al. Considerations for addressing bias in artificial intelligence for health equity. NPJ Digit. Med.6, 170 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 76.Brückner, S., Brightwell, C. & Gilbert, S. FDA launches health care at home initiative to drive equity in digital medical care. NPJ Digit. Med.7, 1–3 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 77.Kevin, W. D. et al. The translational gap for gene therapies in low- and middle-income countries. Sci. Transl. Med.16, eadn1902 (2024). [DOI] [PubMed] [Google Scholar]
  • 78.Tadeusz, C.-H., Ritvij, S., Miriam, A., Stephan, B. & Sandra, B. Artificial intelligence for strengthening healthcare systems in low- and middle-income countries: a systematic scoping review. NPJ Digit. Med.5, 162 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 79.Wanjiku, M. et al. Impact of artificial intelligence assessment of diabetic retinopathy on referral service uptake in a low-resource setting: the RAIDERS randomized trial. Ophthalmol. Sci.2, 100168 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 80.Noelle, W. et al. Feasibility and acceptance of artificial intelligence-based diabetic retinopathy screening in Rwanda. Br. J. Ophthalmol.108, 840–845 (2024. [DOI] [PubMed] [Google Scholar]
  • 81.Li, J. et al. Integrated image-based deep learning and language models for primary diabetes care. Nat. Med.30, 1–11 (2024). [DOI] [PMC free article] [PubMed]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

No datasets were generated or analysed during the current study.


Articles from NPJ Digital Medicine are provided here courtesy of Nature Publishing Group

RESOURCES