Large language medicine: defining a new paradigm in human health

Abstract

Artificial intelligence (AI) is poised to revolutionize our understanding of disease and pre-disease, which could transform the way medicine is practiced. Advances in deep learning systems, alongside the advent of generative AI and large language models, promise to usher in an era of multimodal AI systems that permeate every aspect of medicine. AI is expected to support the full spectrum of public health and clinical functions, including surveillance, monitoring, protection, health promotion, and disease prevention. In parallel, molecular biology has been revolutionized by AI solutions that can accurately model proteins and other biological molecules at scale with the potential to accelerate drug discovery and reshape our understanding of disease and pre-disease mechanisms. However, AI has yet to be incorporated into common day-to-day practice, underscoring the difficulty of clinical integration. Achieving this goal will require advancing algorithmic architecture, sourcing higher-quality multimodal data, increasing processing power and efficiency, developing secure and fit-for-purpose data infrastructures, ensuring interoperability with diverse health systems and clinical workflows, and establishing regulatory pathways to protect patient safety and clarify clinician liability. The potential emergence of autonomous systems capable of carrying out an increasing number of clinical tasks raises important ethical and legal questions about accountability and responsibility in patient care. This review explores the current state of AI in medicine, highlighting that current clinically mature and regulatory-approved products are based on deep learning architecture and are predominantly diagnostic. Robust regulatory frameworks and ethical guidelines must be established to govern the development and deployment of AI, ensuring alignment with patient safety, clinical guidelines, and public trust. Multidisciplinary collaboration among clinicians, researchers, technologists, ethicists, regulators, and policymakers is essential moving forward.

Key points

Introduction

The idea that artificial intelligence (AI) could be used to improve the practice of medicine dates back to the 1950s. The Dartmouth Summer Research Project on AI in 1956, widely regarded as the founding event of the field, introduced the term “artificial intelligence”. During this event, a small group of researchers convened to explore whether tasks that typically require human intelligence could be automated by machines. By the 1970s and 1980s, simple rule-based systems were being explored across industries, including early medical applications such as computer-aided pattern recognition of radiographs (1, 2). However, progress stalled for several decades due to limitations in computational power and data availability, as well as inherent constraints of the logic-based approach itself, including challenges in symbol grounding and limited modeling power (3). In recent years, a convergence of computational capabilities, data availability, and algorithmic innovation has led to a rapid acceleration of AI development.

Numerous sectors are now adopting AI to improve performance capacities, streamline workflows, and find innovative solutions to problems. Medicine stands to be considerably reshaped by AI, as the field offers large, diverse datasets and is a significant focus of investment and public interest (4–7). The potential applications of AI in medicine are vast, spanning not only diagnostics, screening, and personalized therapeutics but also triage, robotic surgery, remote monitoring, at-home patient management, and operational use cases such as workforce planning and resource allocation. However, its adoption into tangible and effective real-world medical solutions has so far been iterative and slow, impeded by challenges such as regulatory hurdles, ethical considerations, models that underperform or fail to generalize well across clinical settings, and, ultimately, the complexity of clinical integration.

These challenges are exacerbated by growing demographic pressures. Aging populations with complex health needs and chronic diseases are being supported by a decreasing working-age workforce (8, 9). A global shortfall of health workers is straining health resources and capacity (10). Despite increasing investment and expenditure, high medical error rates, inefficiencies, and resource waste persist (6, 11, 12). The widespread attention gathering around AI in medicine stems not just from its potential to improve clinical performance but also to reduce errors, streamline workflows, and optimize resource utilization. AI technologies offer promise not only to enhance healthcare delivery but to fundamentally transform the patient experience and improve overall health outcomes.

The term “AI” covers several types of computer systems that automate tasks traditionally requiring human intelligence in some capacity, broadly including machine learning (ML), natural language processing (NLP), computer vision, and robotics (13). One of the first demonstrable uses of AI in medicine dates back to 1995, when the United States Food and Drug Administration (FDA) approved an AI-based software (PAPNET) that automated cervical cancer screening (14, 15). Although early AI tools such as PAPNET generated some interest, their adoption was limited by underwhelming performance and poor integration into clinical workflows. These early rule-based or symbolic AI systems (so-called “good old-fashioned AI” or GOFAI) often failed to demonstrate sufficient clinical utility. In contrast, recent advances in computational power and algorithmic design have enabled the development of more capable systems (16–18). While technical capabilities have improved, challenges in adoption, integration, and implementation persist, and AI uptake in healthcare remains limited and geographically uneven. For example, while the use of AI in cervical cancer screening is gaining traction, it is not yet universally deployed due to a lack of clinical validation, variations in clinical practice, and regulatory considerations (19).

The different epochs of AI provide a framework for understanding the types of AI used in medicine (20). AI 1.0 refers to the use of symbolic AI, i.e., the early form of AI that explicitly encodes human knowledge onto rules to be followed by computer systems (21). AI 2.0 is the era of deep learning, which dominates the current clinical AI landscape and excels in accurately predicting and classifying information from data (mostly images) (22). Most recently, the development of foundation models and generative AI, including large language models (LLMs), defines AI 3.0 and promises to have a far-reaching impact on clinical decision support, administrative tasks, clinical research, and education (23–25). A potential fourth epoch, AI 4.0, may encompass hybrid architectures, agentic AI systems, and other emerging paradigms, as discussed in the Algorithmic Science and Architecture section below.

In tandem, biomedicine is witnessing a similar and equally important paradigm shift. The multimodal patient data that underpin clinical AI, including clinical information obtained from electronic health records (EHRs), investigations (e.g., imaging, laboratory tests, and pathology), biomarkers, and biosensors, are equally central to advances in the biological sciences. When combined with cellular data and “omics”, these data can further our understanding of the drivers of diseases and how to optimally treat them (26). AI techniques are increasingly utilized in drug discovery and development, accelerating the identification of novel therapeutic targets and drug products (27). Ultimately, our understanding of the links between basic pathophysiological mechanisms, the manifestations of disease in the form of clinical symptoms and signs, and the diseases themselves could usher in an era of personalized and precision medicine, where increasingly targeted treatments are tailored to the individual patient (Figure 1) (28).

Figure 1

While multimodal data have long been available in medicine (e.g., clinical notes combined with imaging), recent advances in AI architecture and processing power have substantially improved the ability to integrate diverse data sources effectively. Combined with increasingly sophisticated infrastructure, these developments suggest that the translational impact of AI may begin to be more fully realized in the coming years, though significant barriers to clinical adoption remain. Additionally, the maturation of methods that can leverage increasingly digitized healthcare data, including EHRs, provides growing opportunities for deploying useful AI tools in clinical practice. Future progress may also depend on shifting toward higher-bandwidth data sources, such as continuous physiological monitoring, surgical video, and longitudinal imaging, which offer richer substrates for learning than text-based corpora alone.

AI researchers and developers have suggested that technological capacity may eventually steer toward artificial general intelligence (AGI), in which AI could approach human-level capabilities across a broad range of tasks, though both the feasibility and timeline of AGI remain highly uncertain and a matter of active debate (29). Indeed, some researchers have argued that current AI systems may be fundamentally limited in achieving genuine understanding or general intelligence, noting that computers lack the embodiment, tacit knowledge, and causal reasoning that characterize human cognition (30), while surveys of leading AI researchers reveal deep divisions on whether current approaches will lead to AGI at all (29). However, medicine is a complex field comprising several domains that span the basic sciences of biochemistry, physiology, and anatomy; screening, diagnostics, and prognostics; behavioral, medical, and surgical therapeutics; human factors such as clinician–patient interaction; and the World Health Organization (WHO) Essential Public Health Operations (EPHOs). The EPHOs span surveillance, hazard response, health protection and promotion, disease prevention, governance, workforce development, financing, communication, and public health research (31).

The path to realizing the benefits of AI across these domains will likely be challenging. Critical issues such as upholding data privacy and security, mitigating biases in AI algorithms, ensuring equitable access to AI globally and for disadvantaged populations, and maintaining the trust of both patients and healthcare professionals must be considered in unison. The ethical considerations surrounding AI in medicine, including questions of fairness, accountability, and transparency, are paramount and require careful deliberation by all key stakeholders (32, 33).

This article outlines key areas of promising AI applications in medicine, while acknowledging the cyclical nature of AI development and the possibility that we may still be in the early stages of a Gartner hype cycle (34). Historically, “AI winters” have illustrated how optimism can outpace practical considerations, mandating a thoughtful evaluation of current progress. We begin by reviewing the current landscape of clinically approved AI devices and emerging generative AI tools, followed by an examination of how AI is reshaping molecular biology and the life sciences through advances that are directly relevant to clinical medicine through translational pipelines for drug discovery, precision diagnostics, and targeted therapeutics. We then identify the key technical, infrastructural, and human factor domains that must advance for AI to achieve meaningful clinical integration, before discussing the ethical, regulatory, and legal challenges that lie ahead. Emphasis is placed on the necessity for multidisciplinary collaboration, robust regulatory frameworks, and the development of systems that ensure the integration of AI into medicine enhances, rather than undermines, the quality of healthcare.

Current applications of AI in medicine

AI models have already demonstrated impressive capabilities in several medical applications, highlighting the versatility of AI-based technologies. To better conceptualize these functions, we use the descriptive, diagnostic, predictive, and prescriptive (DDPP) framework. In analytics, this established framework helps to understand what has happened (descriptive), why it happened (diagnostic), what may happen next (predictive), and recommended future actions (prescriptive) (35). Within medicine, the DDPP framework has been effectively applied to bibliometric analysis (36, 37). However, it can also be used to help understand AI functionality in medicine. In a medical AI context, descriptive tools use automated calculations of data to perform discrete tasks such as image analysis or decision support; diagnostic tools aim to establish potential pathologies, including identification of diseases or areas of interest; predictive tools forecast events or stratify risk; and prescriptive analytics recommend or initiate clinical actions such as autonomous decision-making and surgical navigation (38). To further extend this framework, two additional domains, critique and creative, are proposed to capture the evolving scope of AI in medicine. The critique domain refers to the ability of AI systems to critically appraise clinical management plans or evaluate medical literature. Unlike diagnostic reasoning, which aims to identify or classify a condition, critique involves a higher-order evaluative function that assesses the quality, appropriateness, or evidence base of existing clinical decisions, management plans, or published research. The creative domain encompasses solutions that may generate novel applications, though the degree to which AI can produce genuinely original outputs remains actively debated within the field, with arguments both for and against this possibility (39, 40). This expanded DDPPCC framework offers a comprehensive tool to understand both current and future uses of AI in healthcare (Figure 2).

Figure 2

In parallel, AI tools can also span a functional spectrum ranging from advisory to co-pilot to navigator functions (Figure 2). At the advisory level, AI provides basic suggestions or decision support, typically operating as an informational aid that supplements clinical judgment without being deeply embedded in clinical workflows. The predominantly advisory role of current AI systems in clinical settings reflects not only the technical limitations of earlier generations of AI tools but also, importantly, an appropriate caution toward autonomous decision-making in contexts that directly affect human lives. In its co-pilot function, AI operates as a supportive partner, sharing tasks and remaining embedded within existing clinical workflows. Although it provides valuable input, clinical decision-making or execution remains fundamentally steered by humans. At the navigator end, AI adopts a more autonomous or high-level advisory function, synthesizing complex data, offering recommended actions, and potentially initiating clinical steps with minimal human oversight. One notable example of a more autonomous system approaching this navigator category is LumineticsCore (formerly IDx-DR; Digital Diagnostics), which received United States FDA regulatory approval in 2018 for detecting diabetic retinopathy without requiring interpretation by a clinician (41), with other autonomous retinal AI products such as CLAiR (TOKU) since receiving similar approval (42).

The degree to which AI deployment can shift toward these autonomous roles is partly shaped by governing legal and regulatory frameworks, which influence the acceptable scope of AI-driven decision-making. This is closely linked to regulatory device classification, as discussed further in the Regulation section below. Medical AI tools will often exist within this continuum, as illustrated in Figure 2, and their position on it may vary depending on the intended clinical use and end-user needs. Given the high-stakes nature of healthcare, co-pilot functionalities are likely to be more readily accepted in the near term. Nevertheless, as AI technology continues to advance and demonstrate safety and efficacy, it is plausible that future iterations of AI tools will shift increasingly toward the navigator role, delivering broad strategic insights and operating with a higher degree of autonomy within clinical settings. This framework can help stakeholders understand how AI might effectively integrate into healthcare operations, clarify respective responsibilities, and identify the level of autonomy suitable for different clinical domains. Figure 3 illustrates the relationship between AI architectural maturity (epochs), data volume and quality, and functional autonomy, situating these factors within a unified three-dimensional framework.

Figure 3

Approved AI-enabled medical devices

Regulatory approval of AI-enabled medical devices demonstrates how AI technologies are entering formal clinical pathways (43, 44). Several regulatory organizations, including the United Kingdom Medicines and Healthcare products Regulatory Agency (MHRA), the European Medicines Agency (EMA), and the United States FDA, have established guidelines to regulate the development and deployment of AI as a medical device (45–48). In particular, the FDA maintains the AI-Enabled Medical Devices List to enhance transparency by releasing publicly accessible device clearance summaries. To date, the list includes more than 1,500 AI-enabled medical devices that have received regulatory approval (49). Moderate- or high-risk devices (class II or III) that directly influence clinical decision-making, such as AI tools for lesion detection on imaging, are usually subject to FDA oversight. However, several other AI applications operate at lower-risk tiers or serve administrative functions and therefore fall outside the FDA’s purview due to exclusions outlined in the 21st Century Cures Act and other internal agency guidance (50). Consequently, the number of AI systems subject to formal regulatory oversight represents only a portion of the broader AI landscape in healthcare. Mapped onto the DDPPCC framework, most of these FDA-approved devices are descriptive or diagnostic imaging-based applications providing co-pilot functions (Table 1). These leverage the strengths of deep learning, and particularly convolutional neural networks (CNNs), in detecting and characterizing abnormal patterns within medical images (51). Other influential deep learning architectures have also shaped the field, including residual networks (ResNets) that enable training of substantially deeper models, and generative adversarial networks (GANs), which serve as important precursors to current generative AI approaches and have found applications in medical image synthesis and augmentation (52, 53). A smaller number of devices are predictive or prescriptive, while none currently perform critique or creative functions. Notably, many of these AI models have been developed and evaluated in single-center settings, which limits confidence in their generalizability across diverse clinical environments and patient populations.

An example of an AI-enabled device that has achieved regulatory clearance through large clinical trials is Transpara (Screenpoint Medical), a CNN architecture used in breast cancer mammography screening. In a randomized controlled trial involving over 80,000 women, AI-supported mammography screening was non-inferior to standard double-reading in the cancer detection rate and significantly reduced clinician workload (54). While results for the completed trial are awaited, Transpara and similar technologies are currently being used in several United States and European health systems and hospitals (55, 56). These devices could be used as an alternative to standard double reading, as an extra reader, or to help triage high-risk cases for urgent review, highlighting their potential in reducing detection error rates, streamlining workflows, and reducing workload. Other pivotal and prospective trials for descriptive and diagnostic devices have demonstrated statistical accuracy and robustness in lesion detection across a wide range of conditions and specialties, such as diabetic retinopathy (41), prostate cancer (57), skin cancer (58), endoscopy (59), and electrocardiogram interpretation (60). However, a persistent gap remains between technical performance and demonstrated benefit for patients, clinicians, and health systems. Most approved devices have been evaluated primarily on technical accuracy metrics, such as sensitivity and specificity for lesion detection, rather than on downstream health outcomes including patient survival, quality of life, clinician efficiency, or cost effectiveness (61–63). Few have been subject to rigorous prospective evaluation across diverse clinical settings, and post-deployment evidence on real-world performance, safety, and clinical impact remains sparse. Bridging this evidence gap is essential before the benefits observed in controlled research settings can be assumed to translate to routine clinical practice and should be a priority for the field as it matures. Additional image-classification and lesion-detection applications of deep learning devices in medical images have been extensively discussed in other reviews (22, 64, 65).

Predictive and prescriptive AI

Aside from lesion detection on images, AI applications are expanding into other areas, such as triage and surgical support. The FDA-approved Viz.ai LVO (large vessel occlusion) system assists in the automated identification of suspected ischemic strokes from computed tomography (CT) angiograms, coupled with real-time alerts to clinicians that expedite endovascular therapy. A randomized controlled trial demonstrated faster diagnosis and coordination of stroke teams, reduced times for stroke intervention, and improved patient outcomes (66). The Senhance Surgical System (Asensus Surgical) is an example of a prescriptive device that incorporates AI for instrument sensing and robotic assistance during minimally invasive surgery. It provides haptic feedback and eye-tracking camera control and demonstrates comparable outcomes to traditional laparoscopic techniques while also offering ergonomic benefits to surgeons (67). Another example, the StealthStation S8 (Medtronic) neurosurgical navigation system, maps instruments virtually, enhancing surgical precision and reducing operative times (68, 69).

These developments suggest a trend toward AI devices diversifying beyond descriptive and diagnostic-based technology into predictive and prescriptive domains. A scoping review of randomized controlled trials assessing AI in clinical practice revealed various other clinically useful applications, such as improving clinical decision-making, care management, clinician behaviors, and patient symptoms (70). Automated insulin dose adjustments by an AI-based decision tool in patients with type 1 diabetes were comparable in efficacy and safety to physician-led titration, highlighting the potential for remote and automated patient monitoring that demonstrates navigator uses (71). An ML mortality prediction tool for patients with cancer encouraged oncologists to have more serious illness conversations with patients, positively influencing clinician behavior (72). Similarly, the “Stanford death” algorithm has been used to identify palliative care needs in patients, prompting earlier end-of-life discussions and better care aligned with patient goals (73). Mobile health tools and apps using reinforcement learning for patients with chronic pain showed improvements in function and pain scores (74, 75). However, as most studies were single-center trials, further studies in broad and diverse populations are necessary to ensure algorithms generalize to different clinical settings. More broadly, even where AI tools have shown promising results in individual trials, the field has yet to establish a robust evidence base demonstrating consistent and reproducible improvements in patient-centered outcomes (61, 62). Establishing this evidence will be critical for building the case for sustained clinical adoption of AI.

Generative AI in medicine

Large language models

Although CNNs primarily excel at narrow tasks such as image classification and segmentation, the introduction of the transformer architecture in 2017 led to an important paradigm shift. This enabled the application of AI to non-imaging modalities and general-purpose tasks, while also enabling their increased adoption within imaging analysis itself (76). Transformers underpin modern generative AI tools, including popular LLMs such as the generative pre-trained transformer (GPT). LLMs have already shown impressive capabilities in replicating human-like language, encoding medical knowledge, and approximating clinical reasoning. For example, LLMs such as Med-PaLM (Google) have demonstrated high accuracy in answering complex medical questions and can provide step-by-step reasoning and transparent explanations, leading to an interest in their use for both medical education and clinical decision support (77, 78). More recent work has shown that advanced reasoning LLMs can match or exceed physician baselines on complex diagnostic and management reasoning tasks, including real, unstructured emergency department cases, although it is emphasized that these results are largely from curated benchmarks and highlight the urgent need for prospective clinical trials (79).

Building on this, visual language models (VLMs) represent an advancement in multimodal AI, combining visual and textual understanding within a single framework. These models can integrate image data with natural language prompts to perform more complex and clinically relevant tasks, such as image interpretation, automated reporting, and surgical assistance. For example, VLMs have recently been applied to tasks such as spectral CT reconstruction from limited-angle measurements (80), computer-aided diagnosis in endoscopy (81), and delineation of radiotherapy treatment targets (82). Their ability to reason across modalities highlights the potential of VLMs to extend beyond traditional LLMs, particularly in imaging-dominant fields such as radiology, endoscopy, and surgical robotics.

However, these models currently face challenges when applied to real patient information in clinical environments. These include hallucinations, in which the model confidently generates plausible but incorrect information, and non-deterministic outputs, in which the model produces varying outputs from identical inputs (83). Rather than isolated technical flaws, these phenomena may be manifestations of deeper underlying issues, including limited data quality, bias, weak grounding in factual knowledge, and unreliable reasoning processes. Furthermore, regulatory oversight of generative AI remains uncertain. The United States FDA has recently suggested that broadly applicable generative AI models without a specific intended medical use may not qualify as devices, potentially placing them outside the FDA’s regulatory jurisdiction (84). This adds further complexity to the governance of generative AI, emphasizing the need to clarify whether, and how, broadly applicable systems will be regulated as they continue to be applied within various healthcare contexts.

Visual and multimodal models

While multimodal approaches have a longer history in AI and medicine, recent foundation models have substantially enhanced the ability to integrate diverse data sources at scale, marking a potentially significant advance and likely representing a next frontier in this field. Generalist multimodal models such as GPT-4 Vision and Claude 3 have demonstrated varying accuracy in clinical tasks (85, 86), so interest has turned toward specialized and fine-tuned medical models (26). These models aim to combine diverse data inputs to augment medical tasks, mirroring the inherently multimodal nature of medical practice. Med-PaLM Multimodal and, more recently, Med-Gemini are generalist medical AI models that can integrate clinical text, imaging, and genomic information, achieving state-of-the-art results on various medical tasks such as medical question answering, diagnostic challenges, two- and three-dimensional image interpretation, text summarization, and health-outcome prediction (87–89). While real-world evidence is certainly required to ascertain clinical efficacy and the impact on clinical workflows in this high-stakes field, these models provide a glimpse of the possible future of everyday clinical AI systems—namely, models that work within a more autonomous navigator role and perform tasks toward the latter end of the DDPPCC framework. In the meantime, current language models could be more immediately useful in lower-risk NLP and workflow tasks such as medical summarization, documentation, and coding (90, 91).

AI-driven molecular biology

The clinical applications discussed above represent the most immediate and tangible impact of AI in medicine, but an equally consequential transformation is occurring upstream in the life sciences. Advances in molecular biology, including protein structure prediction, drug discovery, and genomic analysis, form the translational pipeline that will increasingly feed into clinical AI tools for diagnosis, prognosis, and treatment selection. Understanding these developments is therefore essential to the central thesis of this review. AI has shown particular promise in these fields, where it has been rapidly adopted. AlphaFold 2 (developed by Google DeepMind) represents a significant breakthrough in accurately predicting protein structures without experimental validation (92). The protein-folding problem, a longstanding challenge in biology, determines the three-dimensional structure of a protein from its amino acid sequence. Following the release of roughly 200 million protein structures without experimental input, structural biology is poised to be reshaped by AlphaFold and similar technologies, such as RoseTTAFold (93, 94). Building on this foundation, AlphaMissense has demonstrated accurate proteome-wide prediction of missense variant effects, offering a powerful tool for identifying potentially pathogenic mutations at scale (95). By accurately modeling protein structures, we could better understand the human proteome, disease mechanisms, and novel targets for drug discovery (96). This is still a relatively nascent field, yet AlphaFold has already been used to uncover the structure of proteins related to tuberculosis (97), understand pathogenic mutations of early-onset Parkinson’s disease (98), and engineer a protein delivery system to cells that could have therapeutic potential (99). So far, AI-developed molecules and drugs in early clinical trials (phase I) have significantly higher success rates than historic industry comparators, highlighting the promise of AI-enabled drug discovery and pipelines (100).

Notably, the success of AlphaFold, which was trained on shared resources including the Protein Data Bank and UniProt (101), also illustrates the importance of broad access to comprehensive datasets and collaboration between industry, academia, and government institutions. Building on these foundations, AlphaFold 3 employs a diffusion-based generative model to predict more complex biological structures and their interactions (102), while insights from protein modeling are being harnessed to streamline drug discovery and accelerate identification of novel therapeutic candidates.

Foundation models may also significantly alter the life sciences landscape. Evo is a foundation model that can decode genomes; predict DNA, RNA, or protein sequences and functions; and generate DNA sequences at the scale of entire genomes (103). Large-scale sequence analysis and generation can accelerate the development of genome therapies and precision medicine strategies, particularly in rare and poorly understood genetic diseases. Similarly, AI-driven approaches have been instrumental in the Human Cell Atlas initiative, which aims to construct a comprehensive biological map of human cells (104). Single-cell datasets have been used to train several foundation models, including SCimilarity, which can predict cell profiles and generate novel biological insights from these data (105). Recent work has extended this paradigm by combining transcriptomic foundation models with active learning to accelerate drug discovery and identify compounds that modulate complex cellular processes (106). Taken together, these innovations (protein modeling, genomic editing, and cell mapping) are converging to reshape the landscape of both the life sciences and translational medicine.

Future of AI in medicine

The evolution of AI may take multiple forms, with some industries undergoing rapid and disruptive transformations while others progress more gradually through incremental advances. Andrew Ng, a leading figure in AI, famously referred to AI as “the new electricity”, highlighting its transformative potential across several industries (107). However, this analogy may be somewhat flawed, as AI functions more as a facilitator than as the content itself. We compare AI to a more fundamental process like photosynthesis. Just as photosynthesis utilizes different inputs, including carbon dioxide and water (which can be thought of as data types) to produce oxygen, AI processes discrete or continuous data types to generate medical advancements and other innovations. This process is sustainable and has the potential to fundamentally change our world, much like how photosynthesis transformed the Earth’s biosphere despite the existence of biochemistry beforehand. Similarly, AI has the capacity to sustainably drive significant advancements by facilitating new solutions and innovations.

As discussed, significant progress has been made in developing and implementing AI solutions with potential impact across the medical field. However, we are still far from realizing the full potential of AI in clinical practice. Bridging this gap will require a sustained and multidisciplinary effort across several key domains, informed by principles of implementation science (108). We identify the following fundamental domains to focus on: (i) algorithmic science and architectures, (ii) datasets, (iii) processing power and efficiency, iv) infrastructure, (v) human factors through co-design, and (vi) interoperability with both legacy and modern healthcare systems (Table 2). Importantly, progress in these domains alone will not be sufficient without addressing the practical enabling conditions for real-world deployment, including sustainable reimbursement and payment models, procurement pathways, governance structures within health systems, post-deployment monitoring and maintenance, and the institutional capacity needed to support safe and long-term adoption. Collectively, these efforts could pave the way for AI systems to reshape healthcare delivery, clinical research, education, workflow efficiency, and patient outcomes.

Algorithmic science and architecture

The transitions between the epochs of AI are mainly defined by evolution of algorithmic architecture and advances in computational platforms, such as graphics processing units (GPUs) enabling deep learning. These transitions are characterized by shifts from symbolic AI to deep learning and foundation models (20). The next epoch, AI 4.0, may consist of optimized existing architectures, hybrid architectures containing desirable features of different algorithms, or entirely novel types of architectures. Techniques to optimize current language models include retrieval-augmented generation, which integrates external knowledge databases to improve the reliability of outputs and reduce hallucinations (109, 110). Hybrid architectures, such as neuro-symbolic AI, combine the strengths of neural networks (learning patterns in data) with symbolic AI (rule-based processing) and are an active field of research (111). Although it is not possible to predict which architecture will dominate the next epoch, researchers are actively exploring novel approaches that could surpass current benchmarks in performance, efficiency, and scalability. One particularly promising area is agentic AI, which refers to autonomous agents capable of learning through iteration, coordinating with other agents, and undertaking complex planning tasks (112, 113). This approach offers a practical pathway toward generalist AI systems. Other important directions include recent advances in graph neural networks, a field with roots in earlier work (114, 115), which have shown renewed promise in drug discovery and repurposing (116); and embodied AI, a concept with a history spanning decades (117), which continues to evolve through interactions with its environment (118). While a full exploration of these architectures is beyond the scope of this review, they represent important directions in the evolution of algorithmic design.

The preceding discussion focuses on AI applications for which published evidence or early-stage validation exists. The following developments, by contrast, remain largely theoretical and have not yet been validated in clinical or applied settings. Recent theoretical advances point to increasingly autonomous and adaptive AI systems. Self-adapting language models can iteratively generate their own fine-tuning data and update directives using reinforcement learning (119). These systems can improve performance on new tasks without external supervision, offering a pathway to more efficient and generalizable knowledge integration. Extending this concept further, the Darwin-Gödel Machine combines elements of Gödel machines, capable of optimal self-modification, with Darwinian evolutionary strategies to construct AI agents that can self-improve both their learning algorithms and underlying architectures over time (120). While still largely theoretical, these frameworks illustrate a hypothetical future in which AI systems could continuously adapt and evolve. Whether such systems would be clinically appropriate, feasible within existing health systems, and acceptable to end users (including patients and clinicians) remains entirely open. Significant technical, regulatory, and safety barriers would need to be addressed before any transition from fixed-model architectures toward self-directed improvement could be considered in healthcare.

Datasets enabling enhanced, reliable, and fairer AI

The performance of an AI model is critically dependent on the quality and diversity of the datasets used to train and test it. Comprehensive and representative datasets not only allow models to capture a wider range of real-world patient characteristics and disease variations but also facilitate more reliable, accurate, and equitable predictions across populations. For example, a comprehensive dataset for developing a chest radiograph AI tool would ideally include images acquired from multiple institutions across different geographic regions, representing a range of patient demographics (including age, sex, and ethnicity), imaging equipment types, and disease prevalence rates, so that the resulting model can perform reliably across the clinical settings in which it is intended to be deployed (121). Several types of health data are being collected continuously and are rich with information about patients and populations, offering ample opportunities to train deep learning models, which require large, curated datasets to be clinically useful (6). However, sourcing high-quality datasets that are representative of intended-use populations is essential to reduce biases and improve the generalizability of AI models, which has been a concern with existing models (122).

Furthermore, data are often not collected equally across individuals or communities, leaving many marginalized populations underrepresented. True representativeness should go beyond statistical parity to account for structural inequities, social determinants of health, and the complex interplay between social and biological factors. To address this issue, data-sharing frameworks such as the Medical AI Data for All initiative can encourage collaboration between global institutions in amassing comprehensive and representative datasets to improve generalizability of models across diverse populations and geographical regions (123). National-scale initiatives such as the Scottish Medical Imaging archive, which links over 57 million radiology studies to their medical records, illustrate what such infrastructure can achieve (124). However, in practice, data sharing is frequently impeded by a range of barriers. These include legal restrictions, such as varying interpretations of data protection regulations across jurisdictions; technical challenges, including a lack of standardized data formats and limited interoperability between institutional systems; institutional reluctance to share data viewed as proprietary assets; and commercial barriers arising when datasets are treated as sources of competitive advantage. An alternative method that obviates the need for sensitive data sharing is federated learning, in which data remain decentralized in disparate organizations and are used to train a centralized model (125). Efforts to utilize both approaches will help to improve the performance of AI models in their intended populations while maximizing model fairness.

Processing power and efficiency

Using state-of-the-art AI algorithms requires significant processing power. The computational demands of deep learning models scale with the number of parameters and data points used (126). In particular, transformer models impose substantial computational costs owing to the quadratic scaling of self-attention mechanisms with sequence length (127). An active area of research involves increasing processing power and efficiency within specialized hardware such as GPUs and, more recently, tensor processing units (128). However, advanced chip production is highly concentrated in a small number of countries and is subject to growing geopolitical tensions. This raises strategic concerns about the sustainability and resilience of the AI development pipeline. For medical AI specifically, deploying large models in clinical settings requires real-time inference, often on resource-constrained hardware; privacy-preserving computation (such as federated learning or on-device processing) adds further demands; and the cost of training and maintaining models on sensitive health data may be prohibitive for many institutions. Quantum computing, though still in its nascent stages, holds promise for exponentially increasing future processing capabilities (129).

Current frontier language models ingest on the order of 20–30 trillion text tokens (~10¹⁴ bytes) and compress this into 3–6 trillion parameters, so model size is now roughly 20–30% of the raw corpus. Under classical scaling heuristics (~20 tokens per parameter), a 6T-parameter model trained on a 20T-token text corpus would have only ~3.3 tokens/parameter, i.e., be under-trained by a factor of about 6, exposing a structural data–capacity mismatch. In parallel, a human child is also exposed to ~10¹⁴ bytes in the first few years of life but via dense visuomotor experience rich in causality and physics rather than sparse text, indicating that our current training substrate is intrinsically low density for learning world models. These constraints imply that merely scaling text-only models to multi-trillion parameters will show diminishing returns, and that future systems could consider raising “intelligence per byte” or “sub-domain intelligence per byte” as a measure of delivering even more clinical pathway-relevant solutions by shifting toward higher-bandwidth data (video, embodied interaction, or simulation) rather than further re-scraping an already saturated text internet. For medicine, this shift could mean training on richer clinical data modalities, such as continuous physiological monitoring, surgical video, and longitudinal imaging sequences, which may yield more clinically relevant models than those derived primarily from text-based corpora.

Infrastructure and energy efficiency

Suitable infrastructure must also be developed to support the deployment of AI systems in healthcare. In addition to the hardware and software requirements for AI model development, these systems necessitate complex and energy-intensive support structures, including model deployment pipelines, data management, and ongoing maintenance processes (130). Secure data centers and environments are also important for storing and maintaining health data used to train and test AI models, while robust cybersecurity measures are essential to safeguard patient information and ensure compliance with privacy regulations (131). However, the expansion of data centers and infrastructure to meet the computational demands of AI also raises significant concerns regarding energy consumption and environmental sustainability. The International Energy Agency estimates that data centers consumed approximately 460 terawatt-hours of electricity in 2022, accounting for almost 2% of global electricity demand; this is expected to double by 2026 (132). To illustrate, a single training run of an earlier frontier model such as GPT-4 has been estimated to consume energy roughly equivalent to the annual electricity usage of a small United States town (133). Large data centers also require vast quantities of water for cooling, prompting growing public concern over and resistance to the expansion of computing facilities, including experimental approaches such as underwater data centers (e.g., Microsoft’s Project Natick) (134). Within healthcare, these demands are particularly consequential because clinical AI systems must operate continuously, process sensitive patient data securely, and scale across dispersed health systems with varying levels of infrastructure.

Intensive energy requirements have profound implications for carbon emissions, illustrated by a recent modeling study examining the operational greenhouse gas emissions and global warming potential of the widespread use of deep learning in digital pathology—a single use case of AI in health (135). Furthermore, inference for LLMs is substantially more compute-, energy-, and cost-intensive on a per-query basis than a conventional web search (136). This underscores the importance of focusing on energy capacity, efficiency, and renewable energy sources in the development and operation of AI-related infrastructure. Efforts to improve the energy efficiency of AI algorithms themselves, such as through model compression techniques including pruning, quantization, and knowledge distillation, can also substantially reduce model size and inference cost while maintaining performance (19). In this way, infrastructure design and sustainability must evolve in tandem with the AI technology itself to ensure responsible long-term integration into healthcare. More broadly, leading AI researchers have voiced concerns about the sustainability and wider societal risks of continued scaling of AI systems (137).

Human co-design

Input from end users (healthcare professionals and patients) throughout the design, development, and deployment phases is an oft-neglected factor that is important in successfully integrating AI models. Understanding the specific needs, workflows, and challenges of healthcare professionals helps ensure that AI tools are user friendly and meet clinical requirements (138). In parallel, patient input in the form of patient and public involvement and engagement promotes patient-centered care and improves outcomes (139). Both steps also foster trust in the use of AI tools, which is a significant barrier to clinical adoption (140).

Importantly, end-user expectations, preferences, and concerns vary substantially. They are likely to differ across ethnic, racial, cultural, linguistic, and religious contexts, shaped by varying levels of trust in healthcare institutions, differing attitudes toward technology and data sharing, and diverse experiences of care (32, 141). For instance, communities with historical experiences of medical mistreatment may approach AI-driven clinical tools with greater skepticism, while linguistic barriers may limit meaningful engagement with systems designed primarily in English or for majority-language populations (142, 143). These differences must be considered throughout the design and evaluation of AI tools to ensure equitable adoption and avoid marginalizing populations whose needs and perspectives may not be reflected in systems developed primarily within well-resourced, Western healthcare settings. Ensuring diversity in co-design processes, including representation from underserved and minority communities, is therefore not only an ethical imperative but a practical requirement for building tools that are trusted and effective across populations.

Explainable AI (XAI) has emerged as one approach to increase acceptance regardless of the type of architecture used. However, the types of explanations commonly featured in research, such as SHAP (SHapley Additive exPlanations) values or feature importance scores, may not necessarily align with the information that clinicians would find useful for real-world clinical decision-making (144). More intuitive explanations might highlight specific elements of a patient’s EHR data or present similar cases, thereby offering practical insights rather than abstract technical metrics. Concurrently, it is important to remain cognizant of inherent limitations of XAI. Many existing XAI solutions are post hoc approximations of algorithms that do not necessarily mirror the functions of the “black box” (145, 146). Therefore, small variations in input data can produce large discrepancies in the explanations provided, raising questions about their robustness and reliability. Nevertheless, promising examples have been proposed. For instance, a recently developed graph neural network model can predict drug-repurposing candidates for thousands of diseases (including those with limited current treatments) and provide an explanation for why the model selected a particular drug (147). This knowledge could strengthen or reframe our understanding of the biological basis of disease processes and provide clinicians and patients with a clearer explanation of complex model predictions, helping to bridge the gap between model performance and human interpretability. A further dimension of explainability that requires attention is causality. While most current AI models learn statistical correlations, causal reasoning seeks to identify why an outcome occurs rather than simply predicting it. Incorporating causal inference methods into medical AI could help distinguish genuine disease mechanisms from spurious associations, support more robust clinical decision-making, and enhance the interpretability and trustworthiness of model outputs (148).

Interoperability with healthcare systems

Many healthcare systems rely on legacy infrastructure based on outdated technology. A global survey conducted in 2021 found that 73% of participating healthcare cybersecurity professionals reported that their organizations were using legacy operating systems (149). Additionally, according to 200 global health system executives, the most significant barrier to implementing an AI transformation is legacy systems that render a digital switch too challenging (150). While AI systems should ideally be designed with interoperability in mind, achieving this across the full diversity of legacy health system architectures may not be feasible in practice. Efforts to harmonize data standards and modernize system architectures should therefore proceed in parallel with AI development rather than expecting AI tools alone to bridge the interoperability gap. Although there is a concerted effort to modernize and digitalize health systems, many low- and middle-income countries and rural areas will continue to rely on legacy infrastructure. For equitable access and impact, a combination of AI adaptability, standards harmonization, and targeted system modernization will be needed to avoid exacerbating existing healthcare disparities and ensure the benefits of AI are delivered globally.

Conditions for AI development and adoption may also vary substantially across clinical specialties. Fields such as oncology and radiology may attract disproportionate attention owing to the availability of large, structured imaging datasets, well-defined diagnostic endpoints, and strong commercial interest (65, 151). By contrast, specialties focused on rare diseases, prevention, and longitudinal care management or those relying on less-structured data (such as free-text clinical notes or patient-reported outcomes) may face distinct and potentially greater barriers to AI integration (152, 153). These include smaller patient populations, limited training data, fewer commercial incentives, and clinical workflows that are less amenable to automation. A realistic appraisal of where and how AI can be most effectively deployed should account for this uneven landscape rather than assuming that advances in data-rich specialties will generalize uniformly across medicine.

Clinical applications

Clinical decision support

Collectively, progress in these areas could lead to the development of algorithms that are heavily integrated into workflows and enhance patient care. The following discussion outlines possible near-term scenarios that, while grounded in current technological trajectories, would require substantial further validation, infrastructure development, and regulatory clearance before clinical implementation. In such scenarios, a primary or secondary care clinician could use an optimized multimodal model that has been trained on global health data and finetuned to a local population. The clinician would be able to input a variety of clinical information obtained from patients, including demographics, clinical history and examination findings in EHRs, vital signs, laboratory tests, biochemical markers, medical imaging, and “omics” data such as genomics, proteomics, and metabolomics. While elements such as integrated clinical decision support and population-level analytics are already available to varying degrees in existing health information technology platforms, the advance envisaged here lies in the ability of multimodal foundation models to synthesize heterogeneous data sources in real time and deliver contextualized, patient-specific outputs that go beyond the rule-based alerting and structured querying of current systems (26, 154).

Healthcare systems could incorporate diverse sensor data including implantable cardiac monitors and brain stimulation devices; external wearables that track vital signs, blood glucose levels, and sleep patterns; and in-home or remote monitoring systems to provide further real-time patient information (155–158). Continuous video streams from clinical devices and wearables, as well as smartphones or in-home cameras, may also enable a non-invasive “video biopsy” paradigm, in which optical and motion-derived digital biomarkers could facilitate an accurate virtual assessment at the point of care. By integrating these data sources, the model could return a likely differential diagnosis and personalized management plan for the clinician’s consideration. In theory, combining multiple forms of clinical information should improve model accuracy and confidence in the diagnostic, prognostic, or management recommendations generated. Importantly, these tools will only augment clinician decision-making and not yet replace it or act autonomously.

This system could be one AI agent among multiple agents integrated into clinical pathways. Instead of a single monolithic program, different clinical and non-clinical domains could be supported by multiple AI agents; each agent specialized for a specific use case and proactively interacting with its environment to accomplish tasks, gather data, and provide real-time feedback (113). Feedback could include providing clinical decision support, assisting surgical procedures, scheduling patient appointments and investigations, optimizing bed management, and managing patient flow.

Research and translational applications

Beyond direct clinical applications, the agentic AI paradigm could also extend to clinical research through “virtual labs” that deploy various specialized AI agents to perform different parts of a study, including data collection and cleaning, hypothesis testing, and manuscript writing (159). While distinct from patient-facing clinical tools, such capabilities could accelerate the generation and testing of novel hypotheses and ultimately feed back into clinical practice. The advent of agentic AI would underscore the shift toward increasingly integrated and autonomous (navigator role) solutions across both clinical and research domains in healthcare.

Another potentially transformative use is to reconstruct our understanding of disease drivers and mechanisms, as depicted in Figure 1. Biobank datasets have been used to develop ML models that can reveal disease signatures and predict over 3,000 diseases undiagnosed at the time of recruitment (160). More recently, a generative transformer model (Delphi-2M) was trained on large-scale longitudinal biobank data and demonstrated the ability to model disease trajectories and predict more than 1,000 diseases, offering interpretable outputs that reflect the natural history of disease (161). The hope is that future tools can unveil markers of pre-disease and lead to accurate disease predictions before any clinical symptoms manifest. This could precipitate a shift of focus from curing diseases to preventing them, which could lead to healthier populations, reduced chronic disease rates, and decreased healthcare expenditure (162). Other AI tools could be used to uncover novel therapeutic targets and assist in drug discovery, as well as aid in education and administrative tasks. In addition, AI-enabled synthetic biology, spanning generative protein design and sequence-to-function modeling, could lead to programmable therapeutics (e.g., de novo enzymes and antibodies), rapid vaccine development, and engineered cell therapies that translate early prediction into targeted intervention (163, 164). Indeed, the versatility of AI means that it will likely permeate every facet of medicine.

Toward AGI

Separately, some of these domains could contribute to the development of proto- or near-AGI systems, though the feasibility and timeline of AGI remain highly uncertain and a matter of significant debate within the research community (165). If realized, such systems could possess the ability to understand, learn, and apply knowledge across several medical tasks, potentially facilitating complex clinical decision-making. They could also adapt to new medical knowledge and guidelines to improve their performance over time. However, continuous learning and model updating in clinical settings present distinct challenges that should not be understated. These include distribution shift, in which the data encountered during deployment diverge from those used in training; model drift, in which performance gradually degrades over time as patient populations or clinical practices evolve; the risk of data poisoning, whether accidental or adversarial; and the need for robust governance frameworks to oversee continuous deployment (166). How these systems would be monitored, validated, and reauthorized on an ongoing basis remains an open question with significant implications for patient safety. Furthermore, whether autonomous AI decision-making would be clinically appropriate, safe, or desirable in practice are questions that extend well beyond technical capability and require careful consideration by clinicians, patients, regulators, and society more broadly. Directing progress first toward domain-specific AGI models that perform well-defined tasks may be a more pragmatic approach and provide more clinical utility. Within the proposed DDPPCC framework, these models could be specialized agents designed to excel at specific tasks within established clinical and non-clinical workflows. In either case, the pursuit of such technological advancement in healthcare raises significant ethical, legal, and regulatory considerations.

As AI systems take on increasingly complex roles in medicine, particularly roles that move beyond narrow diagnostic or assistive tasks, it becomes important to consider the cognitive processes and reasoning capabilities they may be expected to emulate. Our team proposes a framework for conceptualizing these capabilities, called DELMAR, which encompasses six core domains of cognition: Decision-making, Executive functions, Language, Memory, Attention and perception, and Reasoning and problem-solving. These capabilities mirror key components of human intelligence that are often relied upon in clinical reasoning and interaction.

Complementing this, we also propose the CHAPTERLINES taxonomy, which outlines a broad range of reasoning modes that future AI systems may need to reflect or interface with: Causal and counterfactual, Heuristic and commonsense, Analogical and case-based, Physical and spatial, Temporal, Emotional and social, Reflective and critical, Logical or formal, Imaginative and hypothetical, Normative and moral, Equations and mathematical, and Statistical and probabilistic reasoning. While current AI systems are largely limited to statistical and pattern-recognition approaches, this broader framing may help guide the evaluation of more generalizable or agentic AI systems as they emerge and develop.

However, even the currently most advanced models show inconsistent performance across these dimensions. For example, tasks requiring physical or causal reasoning, such as river-crossing puzzles or chain-of-events problems, often expose fundamental gaps in model understanding, as do conceptual tasks that require an understanding of real-world dynamics (167). Some successes, such as solving complex mathematical Olympiad problems within hours of training, raise questions about data contamination, where models may memorize the underlying data rather than reason (168). This phenomenon is sometimes referred to as Potemkin reasoning and highlights the danger of overinterpreting benchmark performance as actual understanding (169). Moving forward, the evaluation of AI in medicine should account not only for performance metrics but also for the depth and robustness of the reasoning involved.

Challenges ahead

Ensuring patient safety, recognizing ethical considerations, and mitigating the risk of harm are equally as pivotal as improving capacity and performance. Robust regulatory frameworks and ethical guidelines will be necessary to govern the development and deployment of clinically applicable AI systems, but these must also be accompanied by attention to the practical conditions of implementation, including reimbursement structures, procurement processes, institutional governance, and post-deployment surveillance. The gap between technical AI capability and actual clinical benefit remains wide; closing it will require sustained investment in the organizational and systemic infrastructure that enables safe adoption alongside model development. Navigating these challenges and safely harnessing the full potential of AI will require multidisciplinary collaboration among clinicians, academic researchers, computer scientists, industry leaders, bioethicists, regulators, funders, policymakers, and patients. Various frameworks and governance strategies have been proposed to guide the ethical implementation of AI in medicine.

International organizations such as the WHO have listed six core principles of safe and ethical AI in health: (i) protect autonomy; (ii) promote human well-being, human safety, and the public interest; (iii) ensure transparency, explainability, and intelligibility; (iv) foster responsibility and accountability; (v) ensure inclusiveness and equity; and (vi) promote AI that is responsive and sustainable (170). The Coalition for Health AI has also outlined principles for trustworthy AI, which are similar to those of the WHO (171). These principles have been adapted into easily followed frameworks such as TURBO (Testable, Usable, Reliable, Beneficial, and Operable) (64).

As an example of how supranational governance can promote ethical and safe AI in medicine, the EU Artificial Intelligence Act establishes a comprehensive, risk-based regulatory framework for AI systems across sectors, including medicine (172). It sets out obligations for developers and deployers, particularly for high-risk applications, and is expected to influence global regulatory alignment (173). In addition, the WHO Regional Office for Europe Technical Advisory Group on AI for Health was recently established to provide expert guidance on integrating ethical principles, governance, regulation, and oversight into national AI strategies (174).

Ethical considerations have been explored in depth in other reviews (175–177). The preceding sections have identified several interdisciplinary challenges, including gaps in outcome-level evidence, data-sharing barriers, infrastructure demands, sustainability concerns, and the need for human co-design. Rather than restating these in full, the following subsections focus on specific areas that warrant particular emphasis as the field matures.

Bias

Bias can arise at the data, algorithm, and system levels (178), and the deployment of intrinsically biased AI in clinical environments could entrench population biases or risk patient safety. Beyond the data representativeness challenges already outlined, algorithmic bias can be particularly subtle, as models encoding demographic knowledge have been shown to propagate biases in disease classification by applying unintended demographic shortcuts (179), underscoring the need for techniques to identify and mitigate hidden biases prior to clinical deployment.

Reporting guidelines and quality assessment tools help stakeholders evaluate risk of bias, reproducibility, clinical applicability, and generalizability of models in published studies of AI tools. These include reporting guidelines of CONSORT-AI (clinical trials) (180), SPIRIT-AI (clinical trial protocols) (181), STARD-AI (diagnostic accuracy) (182), TRIPOD+AI (prediction models) (183), and DECIDE-AI (decision-support systems) (184). Quality assessment tools such as the upcoming QUADAS-AI (diagnostic accuracy) (185) and PROBAST-AI (prediction tools) (186) will help to determine risk of bias and clinical applicability in secondary research, which will become increasingly important as more studies of similar algorithms are published (187). Initiatives such as the STANDING Together (STANdards for data Diversity, INclusivity, and Generalizability) recommendations are important in ensuring that datasets do not introduce biases into models and can effectively generalize to intended-use populations (188).

Regulation

Further along the product pipeline, once AI tools are ready to be deployed in clinical environments, regulatory organizations should develop fit-for-purpose and evidence-based pathways for clinical approval. There is concern that current frameworks, such as the United States FDA 501(k) premarket notification pathway, have introduced the risk of predicate creep, in which changes in the underlying AI technology or intended use occur without sufficient evaluation (44). Although this encourages innovation, regulatory organizations need to increase transparency around evidence standards and develop robust post-market surveillance strategies to ensure patient safety for moderate- or high-risk devices. We welcome the FDA initiative in publishing an online database of all approved AI devices (49) and urge other regulatory organizations, such as the EMA and MHRA, to follow suit. This transparency will facilitate better monitoring of AI technologies and their long-term impacts on clinical care, while also promoting public trust in their deployment within healthcare systems. More thought should be given as to how adverse events are defined and measured when AI is involved (189).

An additional regulatory challenge concerns models that are designed to learn continuously or undergo periodic retraining after deployment. Current regulatory frameworks largely evaluate models at a fixed point in time and may not adequately address the iterative changes in performance, safety, and intended use that accompany continuous learning (190). Developing adaptive regulatory approaches, including ongoing post-market performance monitoring, predefined change protocols, and clear thresholds for when reauthorization is required, will be important for ensuring that continuously evolving models remain safe and effective throughout their operational lifespan. More broadly, it is important to note that the vast majority of AI models in healthcare have not been evaluated prospectively. Silent-pilot or shadow deployment frameworks, in which AI outputs are generated and monitored alongside routine clinical workflows without directly influencing patient care, offer a pragmatic approach to evaluating real-world performance and safety before full clinical deployment and could be more widely adopted (191).

A critical consideration is the “locus of review” for certain devices, which may extend beyond regulatory agencies. In many regions, only a fraction of AI applications in clinical environments undergo formal review by regulatory bodies, partly due to limited throughput, various internal policies, and exemptions for certain classes of software. Therefore, alternative oversight mechanisms, such as internal reviews by hospitals, industry-led standard setting, or ex-post tort liability, may be important. This multifaceted governance landscape emphasizes the need for robust institutional processes, where hospitals and other healthcare organizations institute their own health AI-assurance laboratories or committees to systematically evaluate AI tools (192).

Clinician–AI interaction

Alongside high-level normative frameworks, practice-specific dynamics in clinician–AI interactions must also be considered. Clinicians, as primary users, are not passive receivers but active interpreters of AI systems (193), with their engagement shaped by personal expertise, institutional incentives, and local infrastructure. The same tool, for example, may be regarded as less authoritative by an experienced clinician at a tertiary hospital but more authoritative by a junior clinician in a community clinic. In the former setting, heightened scrutiny can lead to alert fatigue and reduced efficiency, while in the latter, risks include over-reliance, over-diagnosis, and de-skilling (194). Improving clinicians’ literacy regarding the mechanisms, strengths, and limitations of AI tools is therefore essential to prevent misconceptions (195, 196). At the same time, sociological studies of clinician–AI interactions (197, 198) can provide important insights to refine normative frameworks so they are better aligned with clinical realities.

Patient perspectives are equally central to the success of AI in clinical settings. Trust in AI-driven tools may be shaped by prior experiences with healthcare systems, cultural attitudes toward technology, perceptions of data privacy and surveillance, and the degree to which patients feel informed about how AI is being used in their care (142). These factors are likely to vary across populations and cannot be assumed to be uniform. Engaging patients as active participants in the governance and evaluation of clinical AI will be important for ensuring that deployment reflects the values and expectations of those it is intended to serve.

In the near future, clinicians are expected to maintain control over clinical decision-making and primarily use AI-based technologies to inform their practice. However, given the rapid advancement of AI technology, it is important to consider what the future landscape of medicine might entail, as we potentially head toward AGI and the accompanying ethical considerations (64). Eventually, AI may be capable of handling more complex responsibilities such as human interaction, clinical care, treatment planning, and drug discovery (199). This evolution reinforces the concept of agentic AI, i.e., multimodal agents that can act independently and respond to their environment without human intervention.

Accountability

However, the emergence of autonomous agents raises significant ethical and legal questions. If an autonomous AI system makes a mistake that results in patient harm, can it be held accountable? While lawsuits against physicians using AI are possible, they have thus far faced limited success and there are complicated questions when trying to allocate liability between physicians, AI developers, and hospital systems (200–202). Existing principles on robotics and AI primarily focus on human–AI interactions, popularized by Isaac Asimov’s Three Laws of Robotics (203). Future frameworks may need to address interactions between multiple AI entities to ensure human rights are upheld, especially as AI systems begin to collaborate or make collective decisions, such as the proposed “AIonAI” law (204, 205). These are only some examples of important ethical dilemmas that may eventually be important to consider.

Any technological advancement in AI must be accompanied by robust and adaptable ethical, regulatory, and legal frameworks and essentially viewed as a socio-technical process. This perspective recognizes that AI operates within complex systems comprising both technical infrastructure and social, organizational, and cultural factors that shape how it is designed, implemented, and trusted (206). Developing frameworks will require oversight by clinicians, researchers, computer scientists, bioethicists, legal experts, policymakers, and industry leaders. Such multidisciplinary efforts are essential to ensure that the integration of agentic AI into healthcare enhances patient care while addressing the profound ethical and legal challenges it presents. The goal is to harness the full potential of AI technologies responsibly, maintain trust in the healthcare system, and safeguard the well-being of patients. This will be especially important as the impact of AI expands beyond healthcare into wider societal frameworks (178).

Conclusion

AI has the potential to meaningfully contribute to health system transformation through diverse applications: real-time monitoring, at-home and community-based diagnosis and risk prediction, and remote clinical reviews. These applications span conventional biopsy to liquid biopsy and emerging “video-biopsy” approaches. Deep learning in medical imaging and general-purpose foundation models, including LLMs based on the transformer architecture, are already showing promise in enhancing diagnostics, personalizing treatments, and improving healthcare efficiency. However, a persistent gap remains between technical capability and demonstrated clinical benefit, and realizing the potential of AI will require navigating substantial challenges, including computational demands and energy consumption, integration with legacy health systems, implementation barriers across diverse clinical settings, and critical ethical issues such as fairness and accountability. Balancing innovation with caution and responsibility is essential. With rigorous evaluation, robust governance, and sustained multidisciplinary collaboration, AI may help shift care upstream toward prediction and prevention, support more equitable outcomes, and contribute to reducing disease burden, though the pace and extent of this transformation will depend on whether technical advances are matched by the practical conditions required for safe and effective clinical adoption.

References

• 1

  MatsubayashiT. Computer-aided pattern recognition of bone tumor roentgenograms [Japanese]. Nihon Igaku Hoshasen Gakkai Zasshi (1971) 31(6):604–6.

  • Google Scholar
  • View reference in article

• 2

  VirtamaPKatevuoKMäkeläPMäkinenE. Computer aided diagnosis of bone tumors. Acta Radiol Diagn (Stockh) (1979) 20(1A):70–4. doi: 10.1177/028418517902001A10

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  TaddeoMFloridiL. Solving the symbol grounding problem: a critical review of fifteen years of research. J Exp Theor Artif Intell (2005) 17(4):419–45. doi: 10.1080/09528130500284053

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  RajpurkarPChenEBanerjeeOTopolEJ. AI in health and medicine. Nat Med (2022) 28(1):31–8. doi: 10.1038/s41591-021-01614-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  BeamALDrazenJMKohaneISLeongT-YManraiAKRubinEJ. Artificial intelligence in medicine. N Engl J Med (2023) 388(13):1220–21. doi: 10.1056/NEJMe2206291

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  TopolEJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med (2019) 25(1):44–56. doi: 10.1038/s41591-018-0300-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  RajkomarADeanJKohaneI. Machine learning in medicine. N Engl J Med (2019) 380(14):1347–58. doi: 10.1056/NEJMra1814259

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  National Health Service England. Evolving to meet a changing world [online] (2024). Available at: https://www.england.nhs.uk/future-of-human-resources-and-organisational-development/the-future-of-nhs-human-resources-and-organisational-development-report/evolving-to-meet-a-changing-world

  • Google Scholar
  • View reference in article

• 9

  Organisation for Economic Co-operation and Development. The future of health systems [online] (2024). Available at: https://www.oecd.org/en/topics/policy-issues/the-future-of-health-systems.html

  • Google Scholar
  • View reference in article

• 10

  World Health Organization. Health workforce [online] (2024). Available at: https://www.who.int/health-topics/health-workforce

  • Google Scholar
  • View reference in article

• 11

  BerwickDMHackbarthAD. Eliminating waste in US health care. JAMA (2012) 307(14):1513–16. doi: 10.1001/jama.2012.362

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  WeingartNSWilsonRMGibberdRWHarrisonB. Epidemiology of medical error. BMJ (2000) 320(7237):774–7. doi: 10.1136/bmj.320.7237.774

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  LidströmerNAshrafianH. Artificial intelligence in medicine. Cham: Springer (2022).

  • Google Scholar
  • View reference in article

• 14

  WarraichHJTazbazTCaliffRM. FDA perspective on the regulation of artificial intelligence in health care and biomedicine. JAMA (2024) 333(3):241–47. doi: 10.1001/jama.2024.21451

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  KossLGLinESchreiberKElgertPMangoL. Evaluation of the PAPNET™ Cytologic screening system for quality control of cervical smears. Am J Clin Pathol (1994) 101(2):220–9. doi: 10.1093/ajcp/101.2.220

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  McKinneySMSieniekMGodboleVGodwinJAntropovaNAshrafianHet al. International evaluation of an AI system for breast cancer screening. Nature (2020) 577(7788):89–94. doi: 10.1038/s41586-019-1799-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  De FauwJLedsamJRRomera-ParedesBNikolovSTomasevNBlackwellSet al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nat Med (2018) 24(9):1342–50. doi: 10.1038/s41591-018-0107-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  EstevaAKuprelBNovoaRAKoJSwetterSMBlauHMet al. Dermatologist-level classification of skin cancer with deep neural networks. Nature (2017) 542(7639):115–18. doi: 10.1038/nature21056

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  WangJYuYTanYWanHZhengNHeZet al. Artificial intelligence enables precision diagnosis of cervical cytology grades and cervical cancer. Nat Commun (2024) 15(1):4369. doi: 10.1038/s41467-024-48705-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  HowellMDCorradoGSDeSalvoKB. Three epochs of artificial intelligence in health care. JAMA (2024) 331(3):242–44. doi: 10.1001/jama.2023.25057

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  MasarieFEJrMillerRAMyersJD. INTERNIST-I properties: representing common sense and good medical practice in a computerized medical knowledge base. Comput BioMed Res (1985) 18(5):458–79. doi: 10.1016/0010-4809(85)90022-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  AggarwalRSounderajahVMartinGTingDSWKarthikesalingamAKingDet al. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis. NPJ Digit Med (2021) 4(1):65. doi: 10.1038/s41746-021-00438-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  ThirunavukarasuAJTingDSJElangovanKGutierrezLTanTFTingDSW. Large language models in medicine. Nat Med (2023) 29(8):1930–40. doi: 10.1038/s41591-023-02448-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  ClusmannJKolbingerFRMutiHSCarreroZIEckardtJ-NLalehNGet al. The future landscape of large language models in medicine. Commun Med (2023) 3(1):141. doi: 10.1038/s43856-023-00370-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  MoorMBanerjeeOAbadZSHKrumholzHMLeskovecJTopolEJet al. Foundation models for generalist medical artificial intelligence. Nature (2023) 616(7956):259–65. doi: 10.1038/s41586-023-05881-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  AcostaJNFalconeGJRajpurkarPTopolEJ. Multimodal biomedical AI. Nat Med (2022) 28(9):1773–84. doi: 10.1038/s41591-022-01981-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  ArnoldC. Inside the nascent industry of AI-designed drugs. Nat Med (2023) 29(6):1292–95. doi: 10.1038/s41591-023-02361-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  AikawaMSonawaneARChelvanambiSAsanoTHaluAMatamalasJTet al. Precision cardiovascular medicine: shifting the innovation paradigm. Front Sci (2025) 3:1474469. doi: 10.3389/fsci.2025.1474469

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  StricklandEZorpetteG. The AI apocalypse: a scorecard. How worried are top AI experts about the threat posed by large language models like GPT-4? IEEE Spectr (2023). Available at: https://spectrum.ieee.org/artificial-general-intelligence

  • Google Scholar
  • View reference in article

• 30

  FjellandR. Why general artificial intelligence will not be realized. Humanit Soc Sci Commun (2020) 7(1):10. doi: 10.1057/s41599-020-0494-4

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  World Health Organization. Self-assessment tool for the evaluation of essential public health operations in the WHO European Region. WHO Regional Office for Europe (2015). Available at: https://iris.who.int/handle/10665/344398

  • Google Scholar
  • View reference in article

• 32

  ObermeyerZPowersBVogeliCMullainathanS. Dissecting racial bias in an algorithm used to manage the health of populations. Science (2019) 366(6464):447–53. doi: 10.1126/science.aax2342

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  VayenaEBlasimmeACohenIG. Machine learning in medicine: addressing ethical challenges. PloS Med (2018) 15(11):e1002689. doi: 10.1371/journal.pmed.1002689

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  DedehayirOSteinertM. The hype cycle model: a review and future directions. Technol Forecast Soc Change (2016) 108:28–41. doi: 10.1016/j.techfore.2016.04.005

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  RoyDSrivastavaRJatMKaracaMS. A complete overview of analytics techniques: descriptive, predictive, and prescriptive. In: JeyanthiPMChoudhuryTHack-PolayDSinghTPAbujarS, editors. Decision intelligence analytics and the implementation of strategic business management. Cham: Springer (2022). 15–30. doi: 10.1007/978-3-030-82763-2_2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  HungCCTuMYChienTWLinCYChowJCChouW. The model of descriptive, diagnostic, predictive, and prescriptive analytics on 100 top-cited articles of nasopharyngeal carcinoma from 2013 to 2022: bibliometric analysis. Medicine (2023) 102(6):e32824. doi: 10.1097/MD.0000000000032824

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  ChuangHYHoSYChouWTsaiCL. Exploring the top-cited literature in telerehabilitation for joint replacement using the descriptive, diagnostic, predictive, and prescriptive analytics model: a thematic and bibliometric analysis. Medicine (2023) 102(48):e36475. doi: 10.1097/MD.0000000000036475

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  HussainAGuniAGandhewarRWarner-LevyJDavidsonAShahKet al. Is the FDA regulation of cardiology AI devices supporting cardiovascular innovation: a scoping review. Heart (2025). doi: 10.1136/heartjnl-2025-326307

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  LaursenL. Artificial intelligence: can we automate eureka moments? That's the goal of this Swiss company's AI-powered invention tech. IEEE Spectr (2024) 61(11):24–7. doi: 10.1109/MSPEC.2024.10749728

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  AdamD. The intelligence illusion: why AI isn't as smart as it is made out to be. Nature (2026) 651(8107):877–9

  • Google Scholar
  • View reference in article

• 41

  AbràmoffMDLavinPTBirchMShahNFolkJC. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. NPJ Digit Med (2018) 1(1):39. doi: 10.1038/s41746-018-0040-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 42

  OngAYTaribagilPSevgiMKaleAUDowERMacdonaldTet al. A scoping review of artificial intelligence as a medical device for ophthalmic image analysis in Europe, Australia and America. NPJ Digit Med (2025) 8(1):323. doi: 10.1038/s41746-025-01726-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 43

  ZhuSGilbertMChettyISiddiquiF. The 2021 landscape of FDA-approved artificial intelligence/machine learning-enabled medical devices: an analysis of the characteristics and intended use. Int J Med Inform (2022) 165:104828. doi: 10.1016/j.ijmedinf.2022.104828

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  MuehlematterUJBluethgenCVokingerKN. FDA-cleared artificial intelligence and machine learning-based medical devices and their 510(k) predicate networks. Lancet Digit Health (2023) 5(9):e618–26. doi: 10.1016/S2589-7500(23)00126-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 45

  Heads of Medicines Agencies-European Medicines Agency Big Data Steering Group. Multi-annual AI workplan 2023–2028. EMA (2023). Available at: https://www.ema.europa.eu/en/news/artificial-intelligence-workplan-guide-use-ai-medicines-regulation

  • Google Scholar
  • View reference in article

• 46

  Medicines and Healthcare products Regulatory Agency. Software and AI as a Medical Device Change Programme roadmap [online] (2023). Available at: https://www.gov.uk/government/publications/software-and-ai-as-a-medical-device-change-programme/software-and-ai-as-a-medical-device-change-programme-roadmap

  • Google Scholar
  • View reference in article

• 47

  UnsworthHWolframVDillonBSalmonMGreavesFLiuXet al. Building an evidence standards framework for artificial intelligence-enabled digital health technologies. Lancet Digit Health (2022) 4(4):e216–7. doi: 10.1016/S2589-7500(22)00030-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  United States Food and Drug Administration. Proposed regulatory framework for modifications to artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD) [online] (2019). Available at: https://www.regulations.gov/document/FDA-2019-N-1185-0001

  • Google Scholar
  • View reference in article

• 49

  United States Food and Drug Administration. Artificial intelligence-enabled medical devices [online] (2026). Available at: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices

  • Google Scholar
  • View reference in article

• 50

  United States Food and Drug Administration. 21st Century Cures Act [online] (2020). Available at: https://www.fda.gov/regulatory-information/selected-amendments-fdc-act/21st-century-cures-act

  • Google Scholar
  • View reference in article

• 51

  BenjamensSDhunnooPMeskóB. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. NPJ Digit Med (2020) 3(1):118. doi: 10.1038/s41746-020-00324-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 52

  XuWFuY-LZhuD. ResNet and its application to medical image processing: research progress and challenges. Comput Methods Programs BioMed (2023) 240:107660. doi: 10.1016/j.cmpb.2023.107660

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 53

  KoshinoKWernerRAPomperMGBundschuhRAToriumiFHiguchiTet al. Narrative review of generative adversarial networks in medical and molecular imaging. Ann Transl Med (2021) 9(9):821. doi: 10.21037/atm-20-6325

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 54

  LångKJosefssonVLarssonA-MLarssonSHögbergCSartorHet al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. Lancet Oncol (2023) 24(8):936–44. doi: 10.1016/S1470-2045(23)00298-X

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 55

  United Kingdom National Institute for Health and Care Excellence. Artificial intelligence in mammography [online] (2021). Available at: https://www.nice.org.uk/advice/mib242/chapter/The-technologies

  • Google Scholar
  • View reference in article

• 56

  Screenpoint Medical. Transpara: breast AI trusted for better workflow, higher confidence [online] (2024). Available at: https://screenpoint-medical.com

  • Google Scholar
  • View reference in article

• 57

  da SilvaLMPereiraEMSallesPGGodrichRCeballosRKunzJDet al. Independent real-world application of a clinical-grade automated prostate cancer detection system. J Pathol (2021) 254(2):147–58. doi: 10.1002/path.5662

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 58

  VenkateshKPKadakiaKTGilbertS. Learnings from the first AI-enabled skin cancer device for primary care authorized by FDA. NPJ Digit Med (2024) 7(1):156. doi: 10.1038/s41746-024-01161-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 59

  BrownJRGMansourNMWangPChuchucaMAMinchenbergSBChandnaniMet al. Deep learning computer-aided polyp detection reduces adenoma miss rate: a United States multi-center randomized tandem colonoscopy study (CADeT-CS trial). Clin Gastroenterol Hepatol (2022) 20(7):1499–1507.e4. doi: 10.1016/j.cgh.2021.09.009

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 60

  YaoXRushlowDRInselmanJWMcCoyRGThacherTDBehnkenEMet al. Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat Med (2021) 27(5):815–9. doi: 10.1038/s41591-021-01335-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61

  Cruz RiveraSLiuXHughesSEDunsterHMannaEDennistonAKet al. Embedding patient-reported outcomes at the heart of artificial intelligence health-care technologies. Lancet Digit Health (2023) 5(3):e168–73. doi: 10.1016/S2589-7500(22)00252-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 62

  PearceFJCruz RiveraSLiuXMannaEDennistonAKCalvertMJ. The role of patient-reported outcome measures in trials of artificial intelligence health technologies: a systematic evaluation of ClinicalTrials.gov records, (1997–2022). Lancet Digit Health (2023) 5(3):e160–7. doi: 10.1016/S2589-7500(22)00249-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  El ArabRAAl MoosaOA. Systematic review of cost effectiveness and budget impact of artificial intelligence in healthcare. NPJ Digit Med (2025) 8(1):548. doi: 10.1038/s41746-025-01722-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  GuniAVarmaPZhangJFehervariMAshrafianH. Artificial intelligence in surgery: the future is now. Eur Surg Res (2024) 65(1):22–39. doi: 10.1159/000536393

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65

  LiuXFaesLKaleAUWagnerSKFuDJBruynseelsAet al. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health (2019) 1(6):e271–97. doi: 10.1016/S2589-7500(19)30123-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66

  Martinez-GutierrezJCKimYSalazar-MarioniSTariqMBAbdelkhaleqRNiktabeAet al. Automated large vessel occlusion detection software and thrombectomy treatment times: a cluster randomized clinical trial. JAMA Neurol (2023) 80(11):1182–90. doi: 10.1001/jamaneurol.2023.3206

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 67

  SpinelliADavidGGidaroSCarvelloMSacchiMMontorsiMet al. First experience in colorectal surgery with a new robotic platform with haptic feedback. Colorectal Dis (2017) 20(3):228–35. doi: 10.1111/codi.13882

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 68

  Van de KelftECostaFvan der PlankenDSchilsF. A prospective multicenter registry on the accuracy of pedicle screw placement in the thoracic, lumbar, and sacral levels with the use of the O-arm imaging system and StealthStation navigation. Spine (Phila Pa 1976) (2012) 37(25):E1580–7. doi: 10.1097/BRS.0b013e318271b1fa

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 69

  KleckCJCullilmoreILaFleurMLindleyERentschlerMEBurgerELet al. A new 3-dimensional method for measuring precision in surgical navigation and methods to optimize navigation accuracy. Eur Spine J (2016) 25(6):1764–74. doi: 10.1007/s00586-015-4235-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 70

  HanRAcostaJNShakeriZIoannidisJPATopolEJRajpurkarP. Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review. Lancet Digit Health (2024) 6(5):e367–73. doi: 10.1016/S2589-7500(24)00047-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 71

  NimriRBattelinoTLaffelLMSloverRHSchatzDWeinzimerSAet al. Insulin dose optimization using an automated artificial intelligence-based decision support system in youths with type 1 diabetes. Nat Med (2020) 26(9):1380–4. doi: 10.1038/s41591-020-1045-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 72

  ManzCRParikhRBSmallDSEvansCNChiversCRegliSHet al. Effect of integrating machine learning mortality estimates with behavioral nudges to clinicians on serious illness conversations among patients with cancer: a stepped-wedge cluster randomized clinical trial. JAMA Oncol (2020) 6(12):e204759. doi: 10.1001/jamaoncol.2020.4759

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 73

  AvatiAJungKHarmanSDowningLNgAShahNH. Improving palliative care with deep learning. BMC Med Inform Decis Mak (2018) 18(Suppl 4):122. doi: 10.1186/s12911-018-0677-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 74

  PietteJDNewmanSKreinSLMarinecNChenJWilliamsDAet al. Patient-centered pain care using artificial intelligence and mobile health tools: a randomized comparative effectiveness trial. JAMA Intern Med (2022) 182(9):975–83. doi: 10.1001/jamainternmed.2022.3178

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 75

  SandalLFBachKØveråsCKSvendsenMJDalagerTStejnicher Drongstrup JensenJet al. Effectiveness of app-delivered, tailored self-management support for adults with lower back pain-related disability: a selfBACK randomized clinical trial. JAMA Intern Med (2021) 181(10):1288–96. doi: 10.1001/jamainternmed.2021.4097

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 76

  VaswaniAShazeerNParmarNUszkoreitJJonesLGomezANet al. Attention is all you need. arXiv [preprint] (2017). Available at: https://arxiv.org/abs/1706.03762

  • Google Scholar

• 77

  SinghalKAziziSTuTMahdaviSSWeiJChungHWet al. Large language models encode clinical knowledge. Nature (2023) 620(7972):172–80. doi: 10.1038/s41586-023-06291-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 78

  JiangLYLiuXCNejatianNPNasir-MoinMWangDAbidinAet al. Health system-scale language models are all-purpose prediction engines. Nature (2023) 619(7969):357–62. doi: 10.1038/s41586-023-06160-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 79

  BrodeurPGBuckleyTAKanjeeZGohELingEBJainPet al. Performance of a large language model on the reasoning tasks of a physician. Science (2026) 392(6797):524–27. doi: 10.1126/science.adz4433

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 80

  WangYLiangNRenJZhangXShenYCaiAet al. Visual language model-assisted spectral CT reconstruction by diffusion and low-rank priors from limited-angle measurements. Phys Med Biol (2025) 70(23). doi: 10.1088/1361-6560/ae0974

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 81

  LiYDuZLiRGongYZhaoZFengyuG. Computer-aided diagnosis using the large-scale visual language models in screening of digestive endoscopy. J Imaging Inform Med (2025) 39(3):2801–10. doi: 10.1007/s10278-025-01560-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 82

  RajendranPYangYNiedermayrTRGensheimerMBeadleBLeQTet al. Large language model-augmented learning for auto-delineation of treatment targets in head-and-neck cancer radiotherapy. Radiother Oncol (2025) 205:110740. doi: 10.1016/j.radonc.2025.110740

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 83

  HuangLYuWMaWZhongWFengZWangHet al. A survey on hallucination in large language models: principles, taxonomy, challenges, and open questions. arXiv [preprint] (2023). Available at: https://arxiv.org/abs/2311.05232

  • Google Scholar

• 84

  United States Food and Drug Administration. Summary for the Digital Health Advisory Committee meeting. Total product lifecycle considerations for generative AI-enabled devices [online] (2024). Available at: https://www.fda.gov/media/182871/download

  • Google Scholar

• 85

  KaczmarczykRWilhelmTIMartinRRoosJ. Evaluating multimodal AI in medical diagnostics. NPJ Digit Med (2024) 7(1):205. doi: 10.1038/s41746-024-01208-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 86

  JinQChenFZhouYXuZCheungJMChenRet al. Hidden flaws behind expert-level accuracy of multimodal GPT-4 vision in medicine. NPJ Digit Med (2024) 7(1):190. doi: 10.1038/s41746-024-01185-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 87

  TuTAziziSDriessDSchaekermannMAminMChangP-Cet al. Towards generalist biomedical AI. NEJM AI (2024) 1(3):AIoa2300138. doi: 10.1056/AIoa2300138

  • CrossRef
  • Google Scholar

• 88

  SaabKTuTWengW-HTannoRStutzDWulczynEet al. Capabilities of Gemini models in medicine. arXiv [preprint] (2024). Available at: https://arxiv.org/abs/2404.18416

  • Google Scholar

• 89

  YangLXuSSellergrenAKohlbergerTZhouYKtenaIet al. Advancing multimodal medical capabilities of Gemini. arXiv [preprint] (2024). Available at: https://arxiv.org/abs/2405.03162

  • Google Scholar

• 90

  ObikaDKellyCDingNFarranceCKrauseJMittalPet al. Safety principles for medical summarization using generative AI. Nat Med (2024) (12):3417–19. doi: 10.1038/s41591-024-03313-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 91

  KatherJNFerberDWiestICGilbertSTruhnD. Large language models could make natural language again the universal interface of healthcare. Nat Med (2024) 30(10):2708–10. doi: 10.1038/s41591-024-03199-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 92

  JumperJEvansRPritzelAGreenTFigurnovMRonnebergerOet al. Highly accurate protein structure prediction with AlphaFold. Nature (2021) 596(7873):583–9. doi: 10.1038/s41586-021-03819-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 93

  YangZZengXZhaoYChenR. AlphaFold2 and its applications in the fields of biology and medicine. Signal Transduct Target Ther (2023) 8(1):115. doi: 10.1038/s41392-023-01381-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 94

  BaekMDiMaioFAnishchenkoIDauparasJOvchinnikovSLeeGRet al. Accurate prediction of protein structures and interactions using a three-track neural network. Science (2021) 373(6557):871–76. doi: 10.1126/science.abj8754

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 95

  ChengJNovatiGPanJBycroftCŽemgulytėAApplebaumTet al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science (2023) 381(6664):eadg7492. doi: 10.1126/science.adg7492

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 96

  TunyasuvunakoolKAdlerJWuZGreenTZielinskiMŽídekAet al. Highly accurate protein structure prediction for the human proteome. Nature (2021) 596(7873):590–6. doi: 10.1038/s41586-021-03828-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 97

  ChenJFruhaufAFanCPonceJUeberheideBBhabhaGet al. Structure of an endogenous mycobacterial MCE lipid transporter. Nature (2023) 620(7973):445–52. doi: 10.1038/s41586-023-06366-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 98

  GanZYCallegariSCobboldSACottonTRMlodzianoskiMJSchubertAFet al. Activation mechanism of PINK1. Nature (2022) 602(7896):328–35. doi: 10.1038/s41586-021-04340-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 99

  KreitzJFriedrichMJGuruALashBSaitoMMacraeRKet al. Programmable protein delivery with a bacterial contractile injection system. Nature (2023) 616(7956):357–64. doi: 10.1038/s41586-023-05870-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 100

  Kp JayatungaMAyersMBruensLJayanthDMeierC. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discov Today (2024) 29(6):104009. doi: 10.1016/j.drudis.2024.104009

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 101

  European Molecular Biology Laboratory. Annual Report 2021: AlphaFold: a game-changer for structural biology [online] (2021). Available at: https://www.embl.org/about/info/annual-report/ar2021/alphafold-a-game-changer-for-structural-biology

  • Google Scholar

• 102

  AbramsonJAdlerJDungerJEvansRGreenTPritzelAet al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature (2024) 630(8016):493–500. doi: 10.1038/s41586-024-07487-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 103

  NguyenEPoliMDurrantMGKangBKatrekarDLiDBet al. Sequence modeling and design from molecular to genome scale with Evo. Science (2024) 386(6723):eado9336. doi: 10.1126/science.ado9336

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 104

  RoodJEWynneSRobsonLHupalowskaARandellJTeichmannSAet al. The Human Cell Atlas: from a cell census to a unified foundation model. Nature (2024) 637(8048):1065–71. doi: 10.1038/s41586-024-08338-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 105

  HeimbergGKuoTDePiantoDJSalemOHeiglTDiamantNet al. A cell atlas foundation model for scalable search of similar human cells. Nature (2024) 638:1085–94. doi: 10.1038/s41586-024-08411-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 106

  DeMeoBNesbittCMillerSABurkhardtDBLipChinaIFuDet al. Active learning framework leveraging transcriptomics identifies modulators of disease phenotypes. Science (2025) 390(6776):eadi8577. doi: 10.1126/science.adi8577

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 107

  Stanford Graduate School of Business. Andrew Ng: why AI is the new electricity [online] (2017). Available at: https://www.gsb.stanford.edu/insights/andrew-ng-why-ai-new-electricity

  • Google Scholar

• 108

  BauerMSKirchnerJ. Implementation science: what is it and why should I care? Psychiatry Res (2020) 283:112376. doi: 10.1016/j.psychres.2019.04.025

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 109

  LewisPPerezEPiktusAPetroniFKarpukhinVGoyalNet al. Retrieval-augmented generation for knowledge-intensive NLP tasks [article 793]. In: LarochelleHRanzatoMHadsellRBalcanMFLinH, editors. Advances in Neural Information Processing Systems 33: Proceedings of the Thirty-Fourth Conference on Neural Information Processing Systems (NeurIPS 2020); 2020 Dec 6–12; Vancouver, Canada. 9459–74. doi: 10.5555/3495724.3496517

  • CrossRef
  • Google Scholar

• 110

  KresevicSGiuffrèMAjcevicMAccardoACrocèLSShungDL. Optimization of hepatological clinical guidelines interpretation by large language models: a retrieval augmented generation-based framework. NPJ Digit Med (2024) 7(1):102. doi: 10.1038/s41746-024-01091-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 111

  WanZLiuC-KYangHLiCYouHFuYet al. Towards cognitive AI systems: a survey and prospective on neuro-symbolic AI. arXiv [preprint] (2024). Available at: https://arxiv.org/abs/2401.01040

  • Google Scholar

• 112

  HuSLuCCluneJ. Automated design of agentic systems. arXiv [preprint] (2024). Available at: https://arxiv.org/abs/2408.08435

  • Google Scholar

• 113

  QiuJLamKLiGAcharyaAWongTYDarziAet al. LLM-based agentic systems in medicine and healthcare. Nat Mach Intell (2024) 6(12):1418–20. doi: 10.1038/s42256-024-00944-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 114

  MicheliASperdutiAStaritaA. An introduction to recursive neural networks and kernel methods for cheminformatics. Curr Pharm Des (2007) 13(14):1469–95. doi: 10.2174/138161207780765981

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 115

  ScarselliFGoriMTsoiACHagenbuchnerMMonfardiniG. The graph neural network model. IEEE Trans Neural Netw (2009) 20(1):61–80. doi: 10.1109/TNN.2008.2005605

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 116

  CorsoGStarkHJegelkaSJaakkolaTBarzilayR. Graph neural networks. Nat Rev Methods Primers (2024) 4(1):17. doi: 10.1038/s43586-024-00294-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 117

  DreyfusHL. What computers still can't do: a critique of artificial reason. Cambridge, MA: MIT Press (1992).

  • Google Scholar

• 118

  DuanJYuSTanHLZhuHTanC. A survey of embodied AI: from simulators to research tasks. IEEE Trans Emerg Top Comput Intell (2022) 6(2):230–44. doi: 10.1109/TETCI.2022.3141105

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 119

  ZweigerAPariJGuoHAkyürekEKimYAgrawalP. Self-adapting language models. arXiv [preprint] (2025). Available at: https://arxiv.org/abs/2506.10943

  • Google Scholar

• 120

  ZhangJHuSLuCLangeRCluneJ. Darwin Gödel Machine: open-ended evolution of self-improving agents. arXiv [preprint] (2025). Available at: https://arxiv.org/abs/2505.22954

  • Google Scholar

• 121

  Seyyed-KalantariLZhangHMcDermottMBAChenIYGhassemiM. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med (2021) 27(12):2176–82. doi: 10.1038/s41591-021-01595-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 122

  AroraAAldermanJEPalmerJGanapathiSLawsEMcCraddenMDet al. The value of standards for health datasets in artificial intelligence-based applications. Nat Med (2023) 29(11):2929–38. doi: 10.1038/s41591-023-02608-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 123

  SaenzAChenEMarklundHRajpurkarP. The MAIDA initiative: establishing a framework for global medical-imaging data sharing. Lancet Digit Health (2024) 6(1):e6–8. doi: 10.1016/S2589-7500(23)00222-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 124

  BaxterRNindTSutherlandJMcAllisterGHardyDHumeAet al. The Scottish Medical Imaging Archive: 57.3 million radiology studies linked to their medical records. Radiol Artif Intell (2023) 6(1):e220266. doi: 10.1148/ryai.220266

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 125

  RiekeNHancoxJLiWMilletarìFRothHRAlbarqouniSet al. The future of digital health with federated learning. NPJ Digit Med (2020) 3(1):119. doi: 10.1038/s41746-020-00323-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 126

  ThompsonNCGreenewaldKLeeKMansoGF. The computational limits of deep learning. arXiv [preprint] (2020). Available at: https://arxiv.org/abs/2007.05558

  • Google Scholar

• 127

  Duman KelesFWijewardenaPMHegdeC. On the computational complexity of self-attention. In: AgrawalSOrabonaF, editors. Proceedings of the 34th International Conference on Algorithmic Learning Theory. Proceedings of Machine Learning Research, Vol. 201. New York, NY: PMLR (2023). 597–619. Available at: https://proceedings.mlr.press/v201/duman-keles23a.html

  • Google Scholar

• 128

  JouppiNKurianGLiSMaPNagarajanRNaiLet al. TPU v4: an optically reconfigurable supercomputer for machine learning with hardware support for embeddings. In: Proceedings of the 50th Annual International Symposium on Computer Architecture; 2023 Jun 17–21; Orlando, FL, USA. New York, NY: Association for Computing Machinery (2023). 1–14. doi: 10.1145/3579371.3589350

  • CrossRef
  • Google Scholar

• 129

  ChauhanVNegiSJainDSinghPSagarAKSharmaAK. Quantum computers: a review on how quantum computing can boom AI. In: Proceedings of the 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE);Apr 28–29 2022 (2022). 559–63. doi: 10.1109/ICACITE53722.2022.9823619

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 130

  SculleyDHoltGGolovinDDavydovEPhillipsTEbnerDet al. Hidden technical debt in machine learning systems. In: Advances in Neural Information Processing Systems 28: Proceedings of the Twenty-Ninth Annual Conference on Neural Information Processing Systems 2015; Montreal, Canada; 7-12 December 2015. Red Hook, NY: Curran Associates, Inc. (2015). Available at: https://proceedings.neurips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html

  • Google Scholar

• 131

  United Kingdom National Health Service. NHS England Secure Data Environment Service [online] (2024). Available at: https://digital.nhs.uk/services/secure-data-environment-service

  • Google Scholar

• 132

  International Energy Agency. Executive summary [online] (2024). Available at: https://www.iea.org/reports/electricity-2024/executive-summary

  • Google Scholar

• 133

  Epoch AI. How much energy does ChatGPT use? [online] (2025). Available at: https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use

  • Google Scholar

• 134

  MyttonD. Data centre water consumption. NPJ Clean Water (2021) 4(1):11. doi: 10.1038/s41545-021-00101-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 135

  Vafaei SadrABülowRvon StillfriedSSchmitzNEJPilvaPHölscherDLet al. Operational greenhouse-gas emissions of deep learning in digital pathology: a modelling study. Lancet Digit Health (2024) 6(1):e58–69. doi: 10.1016/S2589-7500(23)00219-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 136

  de VriesA. The growing energy footprint of artificial intelligence. Joule (2023) 7(10):2191–94. doi: 10.1016/j.joule.2023.09.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 137

  BengioY. Personal and psychological dimensions of AI researchers confronting AI catastrophic risks [online] (2023). Available at: https://yoshuabengio.org/en/blog/personal-and-psychological-dimensions-ai-researchers-confronting-ai-catastrophic-risks

  • Google Scholar

• 138

  DoniaJShawJA. Co-design and ethical artificial intelligence for health: an agenda for critical research and practice. Big Data Soc (2021) 8(2):20539517211065248. doi: 10.1177/20539517211065248

  • CrossRef
  • Google Scholar

• 139

  AbsolomKHolchPWoroncowBWrightEPVelikovaG. Beyond lip service and box ticking: how effective patient engagement is integral to the development and delivery of patient-reported outcomes. Qual Life Res (2015) 24(5):1077–85. doi: 10.1007/s11136-014-0909-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 140

  AdjekumABlasimmeAVayenaE. Elements of trust in digital health systems: scoping review. J Med Internet Res (2018) 20(12):e11254. doi: 10.2196/11254

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 141

  AmannJVayenaEOrmondKEFreyDMadaiVIBlasimmeA. Expectations and attitudes towards medical artificial intelligence: A qualitative study in the field of stroke. PloS One (2023) 18(1):e0279088. doi: 10.1371/journal.pone.0279088

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 142

  SagonaMDaiTMacisMDardenM. Trust in AI-assisted health systems and AI’s trust in humans. NPJ Health Syst (2025) 2(1):10. doi: 10.1038/s44401-025-00016-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 143

  BarwiseAKCurtisSDiedrichDAPickeringBW. Using artificial intelligence to promote equitable care for inpatients with language barriers and complex medical needs: clinical stakeholder perspectives. J Am Med Inform Assoc (2024) 31(3):611–21. doi: 10.1093/jamia/ocad224

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 144

  ChaddadAPengJXuJBouridaneA. Survey of explainable AI techniques in healthcare. Sensors (2023) 23(2):634. doi: 10.3390/s23020634

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 145

  BabicBGerkeSEvgeniouTCohenIG. Beware explanations from AI in health care. Science (2021) 373(6552):284–86. doi: 10.1126/science.abg1834

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 146

  BabicBCohenIG. The algorithmic explainability “bait and switch”. Minn Law Rev (2023) 108:857–926. doi: 10.24926/265535.4262

  • CrossRef
  • Google Scholar

• 147

  HuangKChandakPWangQHavaldarSVaidALeskovecJet al. A foundation model for clinician-centered drug repurposing. Nat Med (2024). 30:3601–13. doi: 10.1038/s41591-024-03233-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 148

  WuXPengSLiJZhangJSunQLiWet al. Causal inference in the medical domain: a survey. Appl Intell (2024) 54(6):4911–34. doi: 10.1007/s10489-024-05338-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 149

  Healthcare Information and Management Systems Society. 2021 HIMSS Healthcare Cybersecurity Survey. HIMSS (2021). Available at: https://www.carahsoft.com/innovation-healthcare/innovation-resources/2021-HIMSS-healthcare-cybersecurity-survey

  • Google Scholar

• 150

  McKinsey & Company. Health systems’ digital and AI constraints [online] (2024). Available at: https://www.mckinsey.com/featured-insights/week-in-charts/health-systems-digital-and-ai-constraints

  • Google Scholar

• 151

  ShimizuHNakayamaKI. Artificial intelligence in oncology. Cancer Sci (2020) 111(5):1452–60. doi: 10.1111/cas.14377

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 152

  AmorimAMBOrzełUCaniceiroABRosário-FerreiraNMoreiraIS. Artificial intelligence in rare diseases: toward clinical impact. Trends Pharmacol Sci (2025) 46(12):1241–68. doi: 10.1016/j.tips.2025.10.010

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 153

  KellyCJKarthikesalingamASuleymanMCorradoGKingD. Key challenges for delivering clinical impact with artificial intelligence. BMC Med (2019) 17(1):195. doi: 10.1186/s12916-019-1426-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 154

  MatarasoSJEspinosaCASeongDReinckeSMBersonEReissJDet al. A machine learning approach to leveraging electronic health records for enhanced omics analysis. Nat Mach Intell (2025) 7(2):293–306. doi: 10.1038/s42256-024-00974-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 155

  BachtigerPPlymenCMPabariPAHowardJPWhinnettZIOpokuFet al. Artificial intelligence, data sensors and interconnectivity: future opportunities for heart failure. Card Fail Rev (2020) 6:e11. doi: 10.15420/cfr.2019.14

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 156

  WangCHeTZhouHZhangZLeeC. Artificial intelligence enhanced sensors—enabling technologies to next-generation healthcare and biomedical platform. Bioelectron Med (2023) 9(1):17. doi: 10.1186/s42234-023-00118-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 157

  RignyLFletcher-LloydNCapstickANilforooshanRBarnaghiP. Assessment of sleep patterns in dementia and general population cohorts using passive in-home monitoring technologies. Commun Med (Lond) (2024) 4(1):222. doi: 10.1038/s43856-024-00646-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 158

  CapstickAPalermoFZakkaKFletcher-LloydNWalshCCuiTet al. Digital remote monitoring for screening and early detection of urinary tract infections. NPJ Digit Med (2024) 7(1):11. doi: 10.1038/s41746-023-00995-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 159

  SwansonKWuWBulaongNLPakJEZouJ. The Virtual Lab: AI agents design new SARS-CoV-2 nanobodies with experimental validation. bioRxiv [preprint] (2024). doi: 10.1101/2024.11.11.623004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 160

  GargMKarpinskiMMatelskaDMiddletonLBurrenOSHuFet al. Disease prediction with multi-omics and biomarkers empowers case–control genetic discoveries in the UK Biobank. Nat Genet (2024) 56(9):1821–31. doi: 10.1038/s41588-024-01898-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 161

  ShmatkoAJungAWGauravKBrunakSMortensenLHBirneyEet al. Learning the natural history of human disease with generative transformers. Nature (2025) 647(8088):248–56. doi: 10.1038/s41586-025-09529-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 162

  WoolfSH. The power of prevention and what it requires. JAMA (2008) 299(20):2437–9. doi: 10.1001/jama.299.20.2437

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 163

  WangJLisanzaSJuergensDTischerDWatsonJLCastroKMet al. Scaffolding protein functional sites using deep learning. Science (2022) 377(6604):387–94. doi: 10.1126/science.abn2100

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 164

  WatsonJLJuergensDBennettNRTrippeBLYimJEisenachHEet al. De novo design of protein structure and function with RFdiffusion. Nature (2023) 620(7976):1089–100. doi: 10.1038/s41586-023-06415-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 165

  BoveT. CEO of Google's DeepMind says we could be ‘just a few years’ from A.I. that has human-level intelligenceFortune. (2023). Available at: https://fortune.com/2023/05/03/google-deepmind-ceo-agi-artificial-intelligence/

  • Google Scholar

• 166

  SchwabeDBeckerKSeyferthMKlaßASchaeffterT. The METRIC-framework for assessing data quality for trustworthy AI in medicine: a systematic review. NPJ Digit Med (2024) 7(1):203. doi: 10.1038/s41746-024-01196-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 167

  ShojaeePMirzadehIAlizadehKHortonMBengioSFarajtabarM. The illusion of thinking: understanding the strengths and limitations of reasoning models via the lens of problem complexity. arXiv [preprint] (2025). Available at: https://arxiv.org/abs/2506.06941

  • Google Scholar

• 168

  JiangMLiuKZZhongMSchaefferROuyangSHanJet al. Investigating data contamination for pre-training language models. arXiv [preprint] (2024). Available at: https://arxiv.org/abs/2401.06059

  • Google Scholar

• 169

  MancoridisMWeeksBVafaKMullainathanS. Potemkin understanding in large language models. arXiv [preprint] (2025). Available at: https://arxiv.org/abs/2506.21521

  • Google Scholar

• 170

  World Health Organization. Ethics and governance of artificial intelligence for health: WHO guidance. Geneva: WHO (2021). Available at: https://www.who.int/publications/i/item/9789240029200

  • Google Scholar

• 171

  Coalition for Health AI. Responsible AI Guidance [online] (2023). Available at: https://www.chai.org/workgroup/responsible-ai/blueprint-for-trustworthy-ai

  • Google Scholar

• 172

  European Parliament, Council of the European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act) [document 32024R1689]. European Union (2024). Available at: http://data.europa.eu/eli/reg/2024/1689/oj

  • Google Scholar

• 173

  van KolfschootenHvan OirschotJ. The EU Artificial Intelligence Act, (2024): implications for healthcare. Health Policy (2024) 149:105152. doi: 10.1016/j.healthpol.2024.105152

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 174

  World Health Organization. WHO/Europe launches Technical Advisory Group on Artificial Intelligence for Health [online] (2025). Available at: https://www.who.int/europe/news/item/29-09-2025-who-europe-launches-technical-advisory-group-on-artificial-intelligence-for-health

  • Google Scholar

• 175

  NaikNHameedBMZShettyDKSwainDShahMPaulRet al. Legal and ethical consideration in artificial intelligence in healthcare: who takes responsibility? Front Surg (2022) 9:862322. doi: 10.3389/fsurg.2022.862322

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 176

  HindochaSBadeaC. Moral exemplars for the virtuous machine: the clinician’s role in ethical artificial intelligence for healthcare. AI Ethics (2022) 2(1):167–75. doi: 10.1007/s43681-021-00089-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 177

  NingYTeixayavongSShangYSavulescuJNagarajVMiaoDet al. Generative artificial intelligence and ethical considerations in health care: a scoping review and ethics checklist. Lancet Digit Health (2024) 6(11):e848–56. doi: 10.1016/S2589-7500(24)00143-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 178

  AshrafianH. Engineering a social contract: Rawlsian distributive justice through algorithmic game theory and artificial intelligence. AI Ethics (2023) 3(4):1447–54. doi: 10.1007/s43681-022-00253-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 179

  YangYZhangHGichoyaJWKatabiDGhassemiM. The limits of fair medical imaging AI in real-world generalization. Nat Med (2024) 30(10):30(10):2838–48. doi: 10.1038/s41591-024-03113-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 180

  LiuXCruz RiveraSMoherDCalvertMJDennistonAKChanA-Wet al. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat Med (2020) 26(9):1364–74. doi: 10.1038/s41591-020-1034-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 181

  RiveraSCLiuXChanA-WDennistonAKCalvertMJ. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. BMJ (2020) 370:m3210. doi: 10.1136/bmj.m3210

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 182

  SounderajahVGuniALiuXCollinsGSKarthikesalingamAMarkarSRet al. The STARD-AI reporting guideline for diagnostic accuracy studies using artificial intelligence. Nat Med (2025) 31(3):3283–89. doi: 10.1038/s41591-025-03953-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 183

  CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBet al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ (2024) 385:e078378. doi: 10.1136/bmj-2023-078378

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 184

  VaseyBNagendranMCampbellBCliftonDACollinsGSDenaxasSet al. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. BMJ (2022) 377:e070904. doi: 10.1136/bmj-2022-070904

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 185

  GuniASounderajahVWhitingPBossuytPDarziAAshrafianH. Revised tool for the quality assessment of diagnostic accuracy studies using AI (QUADAS-AI): protocol for a qualitative study. JMIR Res Protoc (2024) 13:e58202. doi: 10.2196/58202

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 186

  CollinsGSDhimanPNavarroCLAMaJHooftLReitsmaJBet al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open (2021) 11(7):e048008. doi: 10.1136/bmjopen-2020-048008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 187

  GuniAWhitingPDarziAAshrafianH. The next generation of evidence synthesis for diagnostic accuracy studies in artificial intelligence. Lancet Digit Health (2024) 6(8):e541–42. doi: 10.1016/S2589-7500(24)00115-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 188

  AldermanJEPalmerJLawsEMcCraddenMDOrdishJGhassemiMet al. Tackling algorithmic bias and promoting transparency in health datasets: the STANDING Together consensus recommendations. Lancet Digit Health (2025) 7(1):e64–88. doi: 10.1016/S2589-7500(24)00224-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 189

  BabicBGlenn CohenISternADLiYOuelletM. A general framework for governing marketed AI/ML medical devices. NPJ Digit Med (2025) 8(1):328. doi: 10.1038/s41746-025-01717-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 190

  OnitiuDWachterSMittelstadtB. How AI challenges the medical device regulation: patient safety, benefits, and intended uses. J Law Biosci (2024) 11(1):lsae007. doi: 10.1093/jlb/lsae007

  • CrossRef
  • Google Scholar

• 191

  TikhomirovLSemmlerCPrizantNBhasinSKenyonGvan der VegtAet al. A scoping review of silent trials for medical artificial intelligence. Nat Health (2026) 1:532–54. doi: 10.1038/s44360-025-00048-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 192

  ShahNHHalamkaJDSariaSPencinaMTazbazTTripathiMet al. A nationwide network of health AI assurance laboratories. JAMA (2024) 331(3):245–9. doi: 10.1001/jama.2023.26930

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 193

  OudshoornNPinchT, editors. How users matter: the co-construction of users and technology. Cambridge, MA: MIT Press (2005).

  • Google Scholar

• 194

  CabitzaF. Cobra AI: exploring some unintended consequences of our most powerful technology. In: PelilloMScantamburloT, editors. Machines we trust: perspectives on dependable AI. Cambridge, MA: MIT Press (2021). 87–104. doi: 10.7551/mitpress/12186.003.0011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 195

  WartmanSACombsCD. Reimagining medical education in the age of AI. AMA J Ethics (2019) 21(2):E146–52. doi: 10.1001/amajethics.2019.146

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 196

  McCoyLGNagarajSMorgadoFHarishVDasSCeliLA. What do medical students actually need to know about artificial intelligence? NPJ Digit Med (2020) 3:86. doi: 10.1038/s41746-020-0294-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 197

  HuW.Machine learning in medical systems: toward a sociological agenda. In: BorchCPardo-GuerraJP, editors. The Oxford handbook of the sociology of machine learning. Oxford: Oxford University Press. (2023). 483–508.

  • Google Scholar

• 198

  SloaneMMossE. AI’s social sciences deficit. Nat Mach Intell (2019) 1(8):330–1. doi: 10.1038/s42256-019-0084-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 199

  AshrafianHDarziAAthanasiouT. A novel modification of the Turing test for artificial intelligence and robotics in healthcare. Int J Med Robot (2015) 11(1):38–43. doi: 10.1002/rcs.1570

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 200

  PriceWNGerkeSCohenIG. Potential liability for physicians using artificial intelligence. JAMA (2019) 322(18):1765–6. doi: 10.1001/jama.2019.15064

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 201

  MelloMMGuhaN. Understanding liability risk from using health care artificial intelligence tools. N Engl J Med (2024) 390(3):271–78. doi: 10.1056/NEJMhle2308901

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 202

  PriceWNicholsonICohenIG. Locating liability for medical AI. DePaul Law Rev (2023) 73(2):339. Available at: https://via.library.depaul.edu/law-review/vol73/iss2/8

  • Google Scholar

• 203

  McCauleyL. AI armageddon and the three laws of robotics. Ethics Inf Technol (2007) 9(2):153–64. doi: 10.1007/s10676-007-9138-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 204

  AshrafianH. AIonAI: a humanitarian law of artificial intelligence and robotics. Sci Eng Ethics (2015) 21(1):29–40. doi: 10.1007/s11948-013-9513-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 205

  AshrafianH. Intelligent robots must uphold human rights. Nature (2015) 519(7544):391. doi: 10.1038/519391a

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 206

  ViningRKarimianSMcDonaldNEbieleMDoyleBMcKennaLet al. Trustworthy artificial intelligence and organisational trust: a scoping review using socio-technical systems analysis. Ergonomics (2025), 1–23. doi: 10.1080/00140139.2025.2512426

  • Pubmed Abstract
  • CrossRef
  • Google Scholar