In the European Union (EU), healthcare is a cornerstone of the social and economic framework. Yet, EU health systems face mounting pressures: ageing populations and rising chronic disease rates are driving up demand just as labour shortages constrain service delivery. Meeting these challenges will require both investment in the health workforce and the adoption of innovative tools to enhance system resilience and operational efficiency. This chapter assesses uptake of artificial intelligence in the EU healthcare sector with a focus on diagnostics and medical imaging, healthcare operations and drug discovery. It is based on a literature review and interviews with EU business associations and enterprises between December 2024 and April 2025.
Progress in Implementing the European Union Coordinated Plan on Artificial Intelligence (Volume 2)
Uptake in High-Impact Sectors
Abstract
The healthcare sector is a foundational pillar of the European Union (EU), contributing significantly to both societal well-being and economic resilience. In 2022, healthcare accounted for 10.4% of gross domestic product (GDP) in the European Union and employed more than 10% of the workforce. Yet, EU health systems face intensifying pressures, including an ageing population, increasing prevalence of chronic diseases and persistent healthcare workforce shortages. These structural challenges are straining system capacity, driving up costs and threatening equitable access to care, particularly in underserved regions.
Artificial Intelligence (AI) is rapidly emerging as a key enabler to help address these challenges. AI can alleviate workforce shortages by automating administrative tasks, optimising resource allocation and supporting clinical decision making. Additionally, AI could improve care quality by supporting personalised medicine, enabling earlier detection of diseases and facilitating real-time patient monitoring.
This chapter provides an overview of how the EU healthcare sector is using AI and outlines the key conditions necessary for the responsible, effective and scalable adoption of AI solutions. It begins by mapping the major application areas in which AI is already delivering value or holds significant future promise: medical imaging, hospital operations and drug discovery.
The chapter is structured in two parts:
Overview of the EU healthcare sector and the strategic role of AI: a profile of the EU healthcare system identifies key pressures driving AI adoption and presents the core AI applications shaping the sector. It provides an overview of the state of AI deployment, key enablers and barriers to adoption. This covers issues related to data governance, workforce readiness and regulatory alignment.
Spotlight on selected AI use-cases in healthcare: three high-impact AI use-cases are examined in detail – diagnostics and medical imaging; healthcare operations; and drug discovery. Each section draws on targeted literature reviews and expert interviews, highlighting practical examples of AI deployment. The analysis identifies common barriers, assesses competitiveness potential and provides targeted policy recommendations to support the safe, efficient and equitable scaling of AI technologies across the EU healthcare sector.
Methodological considerations are discussed in Chapter 1. A summary of the key recommendations is provided next.
Data availability and access
Accelerate implementation of the EHDS: fast-track the EHDS roll-out with a clear governance model that ensures seamless, responsible access to health data for research and innovation. Build on federated architecture principles to preserve privacy, maintain intellectual property protections and enable statistical power through distributed access, while encouraging use of shared semantic standards for full interoperability.
Improve health data quality and representativeness: address critical issues related to the availability of high-quality, comprehensive health datasets to support AI development. This includes adopting common data quality standards, documenting data sources rigorously, ensuring broad population coverage and implementing systematic statistical bias identification and mitigation practices.
Strengthen comprehensive health data governance frameworks within and across Member States: develop robust governance structures that ensure timely data availability, quality and interoperability, while facilitating secure cross-sector exchange among public institutions, private entities and research organisations. These frameworks should combine technical infrastructures (such as the EHDS) with clear access conditions, standardised privacy and security protocols, interoperable data-sharing rules and co-ordinated oversight across Member States.
Support open access knowledge bases: support the creation of public probabilistic medical knowledge bases to foster innovation and ensure cross-border AI compatibility. These could include structured, open medical repositories that underpin tools like probabilistic symptom checkers or AI-powered triage systems. Encourage contributors to document data provenance, population coverage and mitigation steps for any identified biases, thereby enhancing trust and usability across diverse clinical contexts.
Infrastructure
Expand AI factories: strengthen dedicated healthcare-focused AI compute and development centres with secure access to specialised datasets and high-performance computing resources.
Invest in cross-border collaboration: support collaborative research through dedicated funding.
AI tools and software
Foster public-private foundations for AI models: establish partnerships to develop foundation models in healthcare that are open, secure and designed for European contexts.
Establish an EU registry of AI approved solutions: create an EU-wide platform listing CE-marked healthcare AI systems, modelled on the Food and Drug Administration device listing in the United States. This should include training data summaries, clinical validation and post-market surveillance to improve transparency, market confidence and clinical uptake.
Institutionalise benchmarking protocols: develop EU-wide “gold standard” model evaluation tests, such as generalisability to external laboratory data. Ensure that models can perform reliably across diverse clinical settings and are not limited to their original development environment. Publish results via public leaderboards to promote transparency and reproducibility.
Regulatory frameworks
Clarify the applicability of the research exemption in the AI Act, particularly regarding pharmaceutical research and development, and internal use of AI models.
Ensure clear pathways for compliance: develop consolidated guidance to streamline requirements across the GDPR, Medical Device Regulation and AI Act, and simplify navigation for developers and healthcare institutions. Share notified‑body expertise, harmonised guidance and, where feasible, a single technical dossier.
Ensure consistent interpretation across the European Union: promote uniform application of AI-related rules through centralised guidance and regulatory co-ordination mechanisms.
Enable regulatory sandboxes: support experimental spaces where start-ups and researchers can test AI systems under regulatory supervision, including sandboxes under the joint oversight of medical device and AI regulators, while maintaining robust risk management.
Skills, trust and collaboration
Develop and retain AI talent in the healthcare sector: scale EU-wide AI education and training programmes, talent attraction schemes and reskilling initiatives to strengthen joint health and AI competences.
Support eHealth training across Member States: create transferable digital skillsets and develop European-level quality indicators to support continuing medical education and cross-border competency recognition.
Institutionalise AI literacy in healthcare quality standards: integrate operational AI literacy into hospital accreditation frameworks to ensure sustained demand for skills development and safe adoption.
Support trust building: invest in public communication campaigns and success stories that highlight responsible AI use to build societal confidence in AI in healthcare and promote responsible adoption.
Promote co-development and knowledge-sharing: encourage early engagement of stakeholders, including healthcare professionals and patients, in the design and deployment of AI technologies.
Health systems in the European Union are at a pivotal moment, shaped by a confluence of challenges including digitalisation, demographic transformation and the ongoing recovery from the COVID‑19 pandemic. These pressures are placing unprecedented demands on healthcare infrastructure, financing and human resources, while also accelerating innovation and reform (OECD/European Commission, 2024[1]).
Key characteristics of the healthcare sector
Sectoral make-up
The healthcare sector is a cornerstone of the EU social and economic landscape, playing a critical role in ensuring public well-being, driving innovation, and contributing significantly to employment and GDP. As a fundamental component of the social model, healthcare systems across Member States are designed to provide universal or near-universal access to medical services, ensuring high standards of patient care and health outcomes.
In 2022, healthcare expenditure accounted for 10.4% of EU GDP (Figure 3.1) and over 10% of total employment, making it the largest employment sector in the region. The sector encompasses hospitals, outpatient care, medical research, pharmaceuticals, biotechnology and medical device industries. The diverse nature of healthcare services reflects the variations in national healthcare systems. These range from predominantly public-funded models, such as in Scandinavian countries, to mixed public-private systems in Germany and France.
Expenditures in healthcare in the last decades have increased significantly. Between 2014 and 2022, the overall level of healthcare spending rose by 39.9%, with some countries – Czechia, Cyprus, Latvia, Lithuania, Malta and Romania – more than duplicating their investments in the sector over the period (Eurostat, 2024[2]). However, while health spending surged in 2020‑2021 during the COVID-19 response, it declined in 2022. Curative and rehabilitative services comprise the largest share of spending, with hospital services accounting for approximately 37% and ambulatory services for around 25%.
Note: Data for the EU, Poland and Finland are provisional.
Source: (Eurostat, 2025[3]), “Health care expenditure by financing scheme”, https://ec.europa.eu/eurostat/databrowser/product/page/HLTH_SHA11_HF
While the European Union has extensive healthcare infrastructure, disparities exist among Member States. The European Union had more than 2.3 million hospital beds (516 per 100 000 people) in 2022 (Eurostat, 2024[4]) and an estimated 1.83 million practising physicians (Eurostat, 2024[5]). However, countries like Bulgaria (823) and Germany (766) maintain over 700 hospital beds per 100 000 inhabitants compared to fewer than 300 in countries like Sweden (190 beds), the Netherlands (245), Denmark (248), Finland (261), Ireland (291) and Spain (294) (Figure 3.2).
Notes:
1. No data for long-term care beds; 2. Estimate; 3. Provisional; 4. Deviation from the definition; 5. Break in time series.
Source: (Eurostat, 2025[6]), “Beds in nursing and other residential long-term care facilities” , https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_bdltc__custom_11992439/bookmark/table?lang=en&bookmarkId=90ccb567-1ba8-41e6-a4a8-ff5edb61d0ca; Eurostat (Eurostat, 2025[7]), “Hospital beds by function and type of care”, https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_bds1__custom_12047472/bookmark/table?lang=en&bookmarkId=c758c065-8b34-4d0a-ba88-06f850d5db99&c=1720078966412.
Healthcare personnel
While the number of doctors is increasing across the European Union, there is still a shortage of healthcare workers. The number of practising doctors in the European Union rose from about 1.65 million to 1.83 million between 2010 and 2022, pushing the average density of physicians from 3.4 to 4.2 per 1 000 inhabitants (Figure 3.3). However, the European Union faces a critical health workforce shortfall, with an estimated deficit of 1.2 million doctors, nurses and midwives as of 2022. In all, 20 countries reported physician shortages and 15 reported nursing gaps (OECD/European Commission, 2024[1]).
This gap between supply and demand of healthcare workers is driven by dual demographic pressures: a rapidly ageing population and a concurrently ageing health workforce. Over one-third of doctors and one‑quarter of nurses are more than 55 years old and nearing retirement. Difficult working conditions, exacerbated by the COVID‑19 pandemic, have led to burnout; retention issues in the sector; and declining interest in health careers.
With respect to the lack of physicians, many countries have been primarily concerned about the growing shortage of general practitioners (GPs), particularly in rural and remote areas, restricting access to primary care. GPs represent only one in five practitioners EU‑wide, a share that has drifted downward over the past decade. The largest urban‑rural gaps are seen in Lithuania, Latvia, Hungary, the Slovak Republic, Slovenia and France (OECD, 2023[8]).
Governments are trying to reverse the slide in the availability of GPs. In France, since 2017, 40% of postgraduate internship positions have been allocated to general medicine. Since 2023, GP trainees are required to complete an extra year in ambulatory settings, preferably in underserved regions (European Observatory on Health Systems & Policies and OECD, 2023[9]).
Notes: The EU average is unweighted.
1. Data refer to all doctors licensed to practice, resulting in a large over-estimation of the number of practising doctors.
2. Data include not only doctors providing direct care to patients, but also those working in the health sector as managers, educators, researchers, etc. (adding another 5-10% of doctors).
3. Medical interns and residents are not included.
4. The latest data refer to 2017 only.
5. The data for Belgium start in 2013 and for Cyprus, Ireland and the Netherlands in 2014 to avoid breaks in time series (the last data point for Ireland relates to 2023 to avoid a break in 2022).
Source: OECD Health Statistics 2024; Eurostat (Eurostat, 2025[10]) “Health personnel” https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_prs2/default/table?lang=en&category=hlth.hlth_care.hlth_res.hlth_staff.
EU Member States need to manage the impact of rapidly ageing populations on healthcare systems. With declining fertility rates and longer life expectancy, the share of the population aged 65 and over has increased across all EU countries (Figure 3.4). This share reached 21% in 2023 and was projected to rise to 29% by 2050 (OECD/European Commission, 2024[1]). Major implications of rapid population ageing are the decline in the potential supply of labour in the economy and a greater demand for labour-intensive long-term care. Non-communicable diseases, such as cardiovascular disease, cancer and dementia, account for a significant share of disease burden. These demographic developments highlight the importance of ensuring that health systems can meet the changing needs of an older population. These developments are likely to include a greater need of integrated, person-centred care (OECD, 2023[11]).
The potential of AI to address challenges in the EU healthcare system
As health systems in the European Union face growing workforce shortages and rising demand for services, the integration of digital technologies and AI offers considerable promise. These technologies can support health workers by streamlining workflows, reducing administrative burdens and enhancing overall productivity.
AI holds significant potential to enhance healthcare systems by improving clinical outcomes, optimising resource use and enabling more personalised care. Emerging applications demonstrate how AI can support early detection of diseases through advanced medical imaging; assist surgical procedures with precision robotics; and accelerate drug discovery by analysing complex biomedical data. AI-powered tools can also reshape healthcare delivery through virtual clinical trials; remote monitoring; and intelligent virtual assistants that streamline administrative tasks and improve patient engagement. Moreover, AI systems that personalise treatment plans based on genetic and environmental factors offer the promise of more targeted and effective interventions.
Note: Provisional/estimated data for European Union, France, Poland, Romania.
Source: (Eurostat, 2025[12]), “Population structure indicators at national level”, https://ec.europa.eu/eurostat/databrowser/view/DEMO_PJANIND/default/table?lang=en.
The adoption of digital health tools is already improving clinical practice and reshaping the workforce. Greater use of digital health tools has led to new roles like telehealth co‑ordinators and telemedicine physicians to deliver care remotely (OECD/European Commission, 2024[1]). Across the care continuum, AI and digital tools are also beginning to augment frontline professionals: from streamlining triage and referral processes to improving diagnostic accuracy and enabling personalised treatments. In addition, AI‑powered administrative technologies, such as multilingual clinical notetaking and automated data entry, can potentially automate up to 30% of routine tasks (Health, EIT; McKinsey & Company, 2020[13]), freeing up clinicians to focus more on patient-facing care.
A forthcoming report analyses the impact of AI on the health workforce in the OECD, including 12 EU countries, with a focus on adoption, measurement and upskilling. It found that all responding countries are leveraging AI to reduce administrative burden (OECD, forthcoming[14]).
In several key use-cases, AI is already having a positive impact in health systems across delivery, management and efficiency (see Table 3.1 for an overview of key use-cases).
At a macro level, AI can both reduce and increase the need for human workers. On one side, machines may take over some tasks i.e. the displacement effect. On the other, by boosting output per worker, AI can spur employers to hire more staff, i.e. the productivity effect. A third dynamic, the reinstatement effect, emerges when AI generates new activities and, with them, new occupations, particularly for people whose skills complement the technology (OECD, 2023[15]). In healthcare, the displacement effect is expected to be modest and limited to a few niche functions. Productivity gains could ease, although not fully solve, staffing shortfalls. Meanwhile, the reinstatement effect is likely to open fresh career paths, including AI data analysts, model developers, telemedicine specialists and other professionals who will integrate, maintain and upgrade AI tools across healthcare systems (Green, 2024[16]). As AI becomes embedded in everyday practice, health systems will increasingly depend on a hybrid workforce that bridges clinical expertise with digital and data fluency.
Note: This classification is based on the “OECD Framework for the Classification of AI systems” (OECD, 2022[17]).
Realising the potential of AI in healthcare will depend on both technological readiness and the capacity of the workforce to adapt. Targeted investment in reskilling and upskilling is essential to equip health professionals with the competencies to work alongside AI systems effectively. These competencies include both technical skills (such as health data management, algorithm literacy and digital diagnostics) and soft skills (such as communication, adaptability and ethical reasoning). Given the sensitive nature of patient data and the complexity of clinical decisions, health workers must understand data privacy regulations and maintain the human-centric ethos of care (OECD/European Commission, 2024[1]).
While use of AI within the healthcare workforce has increased in recent years, it remains lower than in other sectors. AI talent concentration within hospitals and healthcare increased across all EU countries between 2018 and 2024, as reflected in LinkedIn workforce data. However, this growth has been uneven, and overall talent levels remain relatively low (Figure 3.5), particularly when compared with other sectors such as finance or information technology (IT). Bridging this gap will require co‑ordination between healthcare organisations, universities and industry. OECD (forthcoming[14]) highlights several leading practices to help countries build a resilient, AI-literate and ethically educated healthcare workforce.
Note: This chart shows the concentration of LinkedIn members with at least two AI engineering skills or who perform an AI occupation per country, industry and in time. Please see the methodological note for more information.
Source: OECD.AI (2025), data from LinkedIn Economic Graph, last updated 2025-04-07, https://oecd.ai/.
Level of AI uptake in the EU healthcare sector
The modernisation of healthcare infrastructure is increasingly focused on digital health technologies (Stoumpos, Kitsios and Talias, 2023[18]). Key innovations include electronic health records (EHRs); AI for diagnostics; and telemedicine services, which have expanded rapidly since the pandemic (European Commission, 2023[19]). However, while the OECD has made efforts to take stock of AI tools in healthcare, comparative evidence on the uptake of such tools in EU healthcare is lacking.
According to evidence from across OECD countries, AI adoption in healthcare remains fragmented and uneven across countries and within healthcare systems (OECD, 2024[20]). Two forthcoming OECD reports analyse the development of strategies at the intersection of AI and health. One assesses the level of readiness of OECD countries in AI in health across the four pillars of policy, governance, infrastructure and trust (OECD, forthcoming[21]). The other examines the impact of AI on the healthcare workforce (OECD, forthcoming[14]).
At the EU level, the only attempt to take stock of AI tools in healthcare was a one‑off “AI Watch” baseline study published in 2020. It concluded that uptake was still in its infancy and that fresh data would be needed to track progress over time. However, no follow‑up census has been carried out (JRC, 2020[22]).
While the Eurostat annual enterprise survey reports the share of firms that use at least one AI technology, the questionnaire covers all economic sectors in aggregate. It is also limited to private sector organisations with ten or more employees. As a result, public hospitals, primary‑care networks and research institutes are outside the sampling frame.
National heterogeneity compounds the patchiness of EU research data on AI and healthcare. Member States apply different definitions of “AI”, survey different subsets of providers, or merge digital‑health and AI indicators, making cross‑country figures impossible to reconcile. Even within countries, data are rarely broken down by hospital size, specialty or rural‑urban location. Thus, policymakers cannot see where adoption lags most.
On the supply side, the health sector is the second largest contributor to private sector spending in research and development (R&D) among EU-based firms (Nindl et al., 2024[23]). EU health companies rank behind their US counterparts in absolute investment levels. However, health companies within the top 800 EU‑based R&D investors accounted for 19.3% of total EU R&D investment in 2023.
Although the number of EU health firms has remained relatively stable over the past decade, several firms have shown exceptional growth. Notably, BioNTech (Germany) recorded a staggering 5 094% increase in R&D investment over the past ten years. This reflects the sector’s growing dynamism and its strategic importance in the EU innovation landscape.
Furthermore, the health sector accounts for 83.5% of R&D investment by the 99 SMEs within the top 800 EU-based R&D investors, and for 74.7% of the firms. Of the 74 health SMEs, 43% are in biotech and 46% in pharmaceuticals. The remaining 11% are in other health areas (Nindl et al., 2024[23]).
European health companies are increasingly investing in AI to enhance R&D capabilities. For instance, BioNTech SE (BNTX) has expanded its AI-driven drug discovery efforts through the acquisition of InstaDeep, a British AI technology company= (BioNTech, 2023[24]). Sanofi has partnered with OpenAI and Formation Bio to integrate AI into its drug development processes, aiming to accelerate clinical trials and improve efficiency (Sanofi, 2024[25]). Additionally, Merck KGaA is focusing on AI partnerships for early‑stage drug discovery, collaborating with companies like Exscientia and BenevolentAI to expedite development timelines (Merck Group, 2023[26]).
Across the European Union, health-tech start-ups are applying AI to long-standing clinical bottlenecks. In Copenhagen, Radiobotics (2025[27]) uses deep-learning models trained on images from more than 1 300 hospitals to flag fractures, dislocations and subtle trauma signs on plain X-rays within seconds. In Poland, Infermedica (2025[28]) couples a probabilistic knowledge graph with generative AI to run symptom-checker chats that hand clinicians a structured intake and probable differential in minutes, lightening workloads and shortening queues.
Upstream in R&D, European innovators are using simulation and generative models to accelerate drug‑discovery cycles. Budapest-headquartered Turbine turns mechanistic cell models into a virtual oncology lab, screening millions of genetic perturbations in silico each month to surface novel, synthetic-lethal targets (Turbine, 2024[29]). French start-up Iktos (2024[30]) feeds vast cheminformatics data into generative algorithms that design synthesisable small molecules in hours rather than months. Nuritas (2025[31]) in Dublin applies similar techniques to discover bioactive peptides for nutrition and therapeutics. In Denmark, Evaxion Biotech harnesses its AI-Immunology™ platform to propose vaccine antigens and cancer immunotherapies tailored to individual tumour profiles (Evaxion, 2025[32]).
While venture capital (VC) increasingly brings together healthcare and AI, it remains much lower than in other industries. VC backing for EU start-ups that combine AI with healthcare, drugs and biotechnology has accelerated sharply over the last few years. Annual funding leapt from almost negligible levels before 2018 to USD 331 million in 2020 and nearly USD 670 million in 2024, the strongest year on record. Although the 2025 tally sits closer to USD 470 million, the figure reflects only early-year deals and still outpaces any full year prior to 2021 (Figure 3.6). This surge indicates mounting investor confidence in the commercial promise of AI-driven life-science innovation. Nonetheless, VC investments in healthcare remain considerably lower than in other industries, such as financial and insurance services, energy and raw materials. Healthcare represented 7% of total VC investments in AI start-ups in the European Union in 2024 (OECD, 2025[33]).
Source: (OECD, 2025[33]), data from Preqin, last updated 2025-02-18, https://oecd.ai/.
The following sections explore in depth three selected use-cases that showcase the transformative potential of AI in healthcare: i) diagnostics and medical imaging; ii) healthcare operations; and iii) drug discovery.
Interviews and literature informed the content of Table 3.2, which highlights several relevant aspects for each key use-case. These include the way AI operates, the main barriers and the competitiveness potential.
AI in diagnostics and medical imaging
The use of AI in medical imaging is becoming increasingly widespread, with applications ranging from image acquisition to diagnostic support. AI algorithms, especially convolutional neural networks (CNN) and other deep-learning models, can be trained to detect patterns often indiscernible to the human eye (Pinto-Coelho, 2023[34]). Moreover, AI‑based laboratory systems analyse blood tests or biopsy data alongside patient history and genomic information, yielding more precise in‑vitro results and potentially reducing diagnostic errors.
Among healthcare use-cases, AI in medical imaging is considered one of the most advanced in the European Union, reflecting significant progress in both research and clinical deployment. While adoption is growing, a forthcoming study indicates that national-level rollouts remain the exception. Instead, implementation is limited to individual hospitals, health regions or research pilots (OECD, forthcoming[14]).
Main use-cases reported in literature
Radiology
Radiology has been a pioneering field for medical AI, benefitting from decades of digitised imaging and earlier generations of computer-aided detection. Modern AI applications in radiology overwhelmingly use deep learning – especially CNNs – to analyse images for abnormalities, assist in quantifying findings or even generate preliminary reports. For example, AI algorithms can swiftly detect signs of a stroke on a brain Magnetic Resonance Imaging (MRI), often in seconds, aiding radiologists in triaging urgent cases. These technologies typically work as clinical decision support: an AI system processes the image and highlights regions of interest or suggests a likely diagnosis, which the radiologist then confirms or reviews. The goal is not to replace radiologists but to augment their capabilities – catching what might be overlooked; reducing workload for routine tasks like organ segmentation; and standardising interpretations.
Notable European examples of AI integration into radiology are Better Medicine and Radiobotics (Box 3.1), whose CE-marked RB fracture™ solution supports trauma-imaging workflows. The software acts as a second reader, highlighting possible fractures or subtle trauma indicators such as joint effusion on conventional X-rays. It overlays annotations onto duplicate images, leaving the originals unaltered to preserve radiologist autonomy. Their system is used across emergency and radiology departments and is optimised for decision support rather than automation, ensuring clinical oversight remains central.
The vast majority of radiology AI systems rely on supervised deep learning trained on large image datasets. CNNs excel at pattern recognition in imaging, enabling tasks such as lesion detection, classification (e.g. characterising tumours on MRI as benign or malignant) and segmentation (delineating structures like organs or blood vessels). AI applications span almost every subfield of radiology. In neuroimaging, AI tools help identify acute intracranial haemorrhage or early signs of ischaemic stroke on computed tomography (CT) scans, allowing faster activation of stroke protocols. In chest imaging, AI was rapidly deployed during the COVID-19 pandemic to detect pneumonia on chest X‑rays and CT scans. Oncologic radiology has seen AI models that measure and track tumour sizes across serial scans or even predict tumour genetics from imaging (so-called radiomics). Another growing area is using AI for quality control in imaging such as algorithms that automatically check MRI sequences for motion artefacts or ensure a CT scan is taken at the proper anatomical level.
The European radiology community has been actively evaluating these innovations. In a 2024 survey of 572 radiology professionals across Europe, respondents said that “breast and oncologic imaging” (particularly in interpreting CT scans, mammography and MRI) and “the detection of abnormalities in asymptomatic subjects” were the two areas where AI had the most expected impact (Zanardo et al., 2024[35]). The same survey found that nearly half of radiologists who responded – 274 of 572 (48%) – were already using some form of AI in their practice, with an additional 25% planning to adopt it soon (Zanardo et al., 2024[35]).
Uptake of AI in medical imaging varies across countries. In the Netherlands, one study systematically tracked AI adoption in 69 radiology departments nationwide between 2020 and 2022 (van Leeuwen et al., 2023[36]). The results showed a steady rise in implementation: in 2020, only 14 departments (around 20%) were using any CE-marked AI software for image analysis; by 2022, this number had grown to 23 departments – roughly one-third of Dutch radiology centres (van Leeuwen et al., 2023[36]). Furthermore, the diversity (from 7 types in 2020 to 34 in 2022) and the number of total implementations (19 in 2020, 68 in 2022) expanded rapidly (van Leeuwen et al., 2023[36]). Common uses included AI for interpreting chest CT scans (e.g. detecting pulmonary nodules or COVID-related findings), brain CT scans (often stroke and trauma diagnostics) and musculoskeletal X-rays (van Leeuwen et al., 2023[36]).
In other EU countries, a mix of local enterprise and collaborations with AI vendors has driven adoption of AI imaging. In Germany, radiology departments have begun integrating AI tools, although uptake has largely occurred on a hospital-by-hospital basis. In 2021, in a notable milestone, the Accident Hospital Berlin (UKB) became the first German hospital to implement a comprehensive AI platform for radiology. The UKB deployed an AI suite to automatically flag time-critical findings on scans – such as brain bleeds or pulmonary emboli – to prioritise and expedite treatment for patients with life-threatening conditions (Aidoc, 2021[37]). The chief of radiology at UKB praised the system for providing an “additional safety net for patients and radiologists” without removing the expert from the loop (Aidoc, 2021[37]).
Several barriers continue to constrain the pace of AI adoption in radiology. The first challenge for implementers is to integrate AI into the workflow and hospital IT systems. Radiology departments rely on Picture Archiving and Communication Systems (PACS) and radiology information systems; an AI algorithm must seamlessly plug into these, fetch images and return results without disrupting routine. A Dutch study found IT integration problems were among the top obstacles, alongside the costs of AI software (van Leeuwen et al., 2023[36]).
While AI tools promise long-term savings through efficiency, European hospital administrators are cautious about the upfront expense and unclear reimbursement models. Indeed, reimbursement is complex: most EU health systems do not yet have dedicated payment codes for AI-supported diagnostics, which means that hospitals absorb the cost. Collective purchasing models can help address this issue and enabled scalable, equitable access to imaging support tools. Through the Artificial Intelligence Finance Institute in the Netherlands, for example, five hospitals pooled resources to evaluate and co-deploy AI systems (Radiobotics, 2025[38]).
Healthcare institutions could also benefit from clear guidance and best practices on how to assess, test and select the most suitable AI tools for their specific settings and clinical workflows. This type of structured support would help reduce implementation costs and lower barriers to entry. In so doing, it would empower hospitals to make informed, efficient and safe choices tailored to their needs.
Companies cited a further barrier to integration: local testing. While AI systems perform reliably across institutions, many hospitals still require three-month pilots, citing the uniqueness of their patient population, data and workflows. Local testing can provide valuable reassurance and help fine-tune integration. However, it also introduces delays, increases implementation costs and complicates efforts to scale AI solutions efficiently. These pilot requirements create friction in the deployment process and can deter smaller vendors, ultimately slowing innovation and uptake at a system-wide level.
Another reported challenge is regulatory uncertainty. Under the EU Medical Device Regulation (MDR), radiology AI software is classified as a medical device that requires European conformity certification (CE marking) and evidence of safety and clinical effectiveness. Navigating this approval process can be complex, especially as the EU AI Act imposes additional requirements specific to algorithms (Busch et al., 2024[39]). Under the Act, medical devices using AI must comply with rigorous risk assessment and mitigation protocols, ensuring that potential harms are identified and managed. According to companies, the dual frameworks of MDR and the EU AI Act increase complexity, cost and time-to-market, creating potential barriers, especially for smaller innovators. The medical technology industry has emphasised the importance of aligning the legislative requirements with those established under the MDR and the In Vitro Diagnostic Medical Devices Regulation to prevent contradictions, duplications or additional unnecessary burdens.
Finally, there are human factors: some radiologists remain sceptical of AI’s reliability, especially if trained on data that might not reflect local patient populations or rare conditions.
Despite these challenges, European radiology is moving towards greater AI integration. Many radiologists now view AI as a helpful “second pair of eyes”. For example, an algorithm as a concurrent reader “acts as a second observer” to catch overlooked polyps in the colon or nodules in the lung (Kelly, 2019[40]). The consensus is that AI, once properly validated and integrated, can enhance diagnostic confidence and speed, allowing radiologists to focus more on complex cases and patient consultation.
Dermatology
The application of AI in dermatology has focused on analysis of skin lesions from images. Essentially, computer vision identifies malignant lesions (like melanoma or carcinoma) from benign ones (like moles or rashes). This field attracted attention after early studies showed deep-learning systems could match or even exceed the accuracy of dermatologists in classifying photographed skin lesions. A seminal 2017 study (Esteva et al., 2017[41]) famously trained a CNN on hundreds of thousands of skin images and achieved dermatologist-level performance in detecting melanoma and keratinocyte carcinomas. Since then, numerous groups worldwide, including in Europe, have refined AI models for dermoscopic images and even for smartphone-captured photos of skin lesions. The promise is especially high for improving early detection of skin cancer.
AI could help GPs or patients themselves identify suspicious lesions for referral. In so doing, they could address the imbalance between the high incidence of skin abnormalities and the limited number of dermatologists available. In Europe, the ratio of dermatologists to population is low – roughly 30 per 1 million people on average (Skin Analytics, 2023[42]), which contributes to long waiting times for specialist consultations. AI tools offer a way to triage lesions: AI-driven teledermatology could potentially handle routine or obviously benign cases, reserving in-person dermatologist appointments for high-risk cases.
The core technology in dermatology AI is image classification using deep CNNs. These networks can be trained on clinical photographs or dermoscopic images (the latter being images of skin magnified by a dermatoscope) to distinguish patterns associated with different skin conditions. Most AI models in this domain are supervised (trained on images labelled by experts or via biopsy-proven diagnoses). Beyond skin cancer, AI experiments aim to classify inflammatory skin diseases (psoriasis vs. eczema, for example) or to assess severity of chronic conditions like acne by analysing images. However, these are less mature than the cancer detection use-case.
Compared to radiology, AI in dermatology is less widely adopted in routine care, but momentum is building. In the most significant regulatory development, the European Union recently approved an autonomous AI diagnostic system for skin cancer. In 2023, a British-based company, Skin Analytics, announced that its AI system called DERM had achieved a Class III CE mark under the new European MDRs – the highest risk class. This made it “Europe’s first and only Class III CE-marked medical device” for autonomous skin cancer detection (Skin Analytics, 2023[42]). The approval signifies that an AI can, in a regulated way, make a clinical decision (ruling out skin cancer) without immediate clinician oversight. According to the company, the DERM system can analyse a skin lesion image and, with high accuracy, determine if it is not cancerous, safely reducing the need for unnecessary dermatology appointments for low-risk cases (Skin Analytics, 2023[42]). It reportedly achieves a 99.8% accuracy in ruling out cancer, slightly surpassing the ~98.9% performance of the average dermatologist. The Class III designation indicates regulators believed the system could influence life-and-death decisions (hence, requiring robust evidence and monitoring). DERM has been initially deployed in the National Health Service (NHS) of the United Kingdom (UK) as part of pilot programmes to shorten skin cancer referral times. Hospitals in EU countries are expected to follow suit now that EU-wide approval is in place (Skin Analytics, 2023[42]).
Better Medicine
Estonia‑based Better Medicine focuses on a multi‑organ, cancer‑centred platform that layers several software‑as‑a‑medical‑device (SaMD) modules into a single workflow.
The flagship suite, BM Vision, already contains a kidney model that became the first CE‑certified AI for kidney‑cancer detection under the EU‑MDR. The tool automatically flags, classifies and measures renal tumours on contrast‑enhanced CT scans. In this way, it acts both as a triage gatekeeper and a “second set of eyes” that catches incidental lesions outside the main field of view. A prospective study showed the software helped radiologists quantify masses 52% faster and lifted tumour‑detection sensitivity to 99.2%, while improving inter‑reader consistency. Beyond the kidney module, Better Medicine is rolling out additional BM Vision organ models: its lung nodule version reports up to 91.6% sensitivity on native or contrast CT scans.
The models plug into an ecosystem supported by Digital Imaging and Communications in Medicine (DICOM). They include BM Viewer (a radiologist‑designed DICOM viewer), BM Gateway (a DICOM-friendly solution seamlessly integrating with PACS systems and scanners) and iMeasure for automated longitudinal lesion tracking and structured reports.
Radiobotics
Copenhagen‑based Radiobotics concentrates on musculoskeletal X‑ray analytics, delivering fast, explainable results inside routine PACS.
The company’s flagship software is RB fracture,™ an AI‑as‑a‑medical‑device that reads trauma X‑rays of the entire appendicular skeleton. Trained on data from more than 1 300 hospitals around the world, the algorithm highlights fractures, dislocations, joint effusions and lipo‑haemarthrosis within a median 13 seconds. It returns bounding‑box annotations, a colour‑coded thumbnail and a DICOM‑structured report directly inside the existing PACS worklist.
The software delivers pooled accuracy, sensitivity and specificity of 94%. Retrospective audits at Kettering General Hospital showed an 86% reduction in missed fractures during the audit window following implementation.
Source: (Radiobotics, 2024[43]), “Transforming fracture detection accuracy at Kettering General Hospital”, https://radiobotics.com/case-study/transforming-fracture-detection-accuracy-at-kettering-general-hospital; (Radiobotics, 2025[27]), “RBfracture™” https://radiobotics.com/solutions/rbfracture/; (Better Medicine, 2025[44]), “Home page”, https://bettermedicine.ai/.
Dermatology faces several distinct challenges in AI uptake: data diversity and workflow integration. Skin image AI must contend with a wide variety of skin tones, lighting conditions and image acquisition methods. A model trained on dermatoscope images taken by experts may not perform well on a smartphone photo taken by a patient. Ensuring AI accuracy across the spectrum of patient populations in Europe requires extensive, representative training data, which has not always been the case. Some early AI models showed biases, performing worse on darker skin (Behara, Bhero and Agee, 2024[45]). Another challenge is integration into workflow. Unlike radiology, where the radiologist is already using a computer and can readily receive AI results, dermatologists often rely on direct visual examination. Incorporating AI might mean changing how dermatologists document and analyse lesions. For example, they may need to take standardised photos of every lesion to feed into an algorithm, which can be time-consuming.
Cardiology
Cardiology covers a broad range of diagnostics, and AI is being applied both in imaging (such as echocardiography, cardiac MRI/CT) and in other diagnostic tests like electrocardiograms (ECGs). While radiology has seen earlier adoption of AI, cardiology is catching up as digital data become more abundant in cardiovascular care.
Echocardiography, an ultrasound-based heart imaging modality, is a prime candidate for AI assistance. Interpreting an “echo” requires measuring various dimensions (like chamber sizes, wall thickness) and assessing function (ejection fraction, valve function) – tasks that AI could partially automate. In recent years, AI algorithms have become able to automatically calculate the left ventricular ejection fraction from echo images (Narang et al., 2018[46]) and detect regional wall motion abnormalities. They can even diagnose specific conditions such as hypertrophic cardiomyopathy from subtle imaging patterns (Kim et al., 2022[47]).
Beyond imaging, ECG is a widely used diagnostic tool in cardiology where AI is making a substantial impact. ECGs are essentially time-series data (waveforms). Advanced AI models (often using deep neural networks that take raw ECG signal as input) can detect atrial fibrillation from a normal sinus-rhythm ECG. Essentially, they recognise subtle patterns or predict the risk of future heart failure by finding minute abnormalities in the waveform imperceptible to humans.
European cardiologists have begun to explore using advanced AI to heighten the impact of ECGs. Research has found that “AI-powered ECG interpretation has shown promising results, improving detection of arrhythmias, ST-segment changes, QT prolongation and other ECG abnormalities” (Martínez-Sellés and Marina-Breysse, 2023[48]). One study reported that an AI-enhanced 12-lead ECG could identify acute myocardial infarction (heart attack) more accurately than conventional criteria, potentially enabling faster intervention (Sung Lee et al., 2025[49]). In 2024, the European Society of Cardiology summarised that “remarkable AI analysis by deep-learning convolutional algorithms” now enables rapid interpretation of ECGs (Androulakis and Fielder, 2024[50]).
A key challenge in cardiac imaging is the lack of standardisation across procedures. The quality and orientation of ultrasound (echocardiography) images depend heavily on how the technician holds the probe. As a result, image quality and viewing angle can change from one scan to the next. AI systems trained on high-quality, consistently captured images may struggle when applied to less optimal scans.
Another issue is too much information. Cardiology generates a wealth of data, ranging from echocardiographic videos and ECG waveforms to laboratory results, providing ample inputs for AI tools. However, this volume of data also introduces complexity. A cardiologist could end up with separate AI reports for each test (e.g. echocardiogram, ECG, blood analysis). Without a single, integrated dashboard, clinicians may face fragmented or even conflicting recommendations, which may lead to information overload and complicate decisions.
Gastroenterology
Another field where AI is gaining increasing interest is gastroenterology, particularly in the context of endoscopic imaging. Colonoscopy – the endoscopic examination of the large intestine – is a crucial tool for colorectal cancer screening and prevention. A well-known challenge in colonoscopy is that endoscopists can miss polyps (precancerous lesions), especially small or flat ones. Missed polyps can lead to interval cancers, so improving polyp detection is critical. AI offers real-time image analysis during the endoscopy procedure. The AI acts like a co-pilot: as the endoscopist navigates the colon, the AI analyses the video feed frame by frame and highlights any regions that look like polyps. This is known as computer-aided detection in endoscopy.
GI Genius was the first AI system for colonoscopy to become commercially available. GI Genius, which received CE marking in Europe and was launched in EU markets around 2019, was touted as “the first AI system for colonoscopy” (Kelly, 2019[40]). It integrates with existing endoscopy hardware – it was made “brand agnostic” to be compatible with most endoscope video systems used in hospitals (Tham et al., 2023[51]). This approach eased adoption, as an endoscopy unit can add the GI Genius module to its equipment stack without substantial upgrades. The system works by overlaying a visual marker on the video when a suspected polyp is in view. By 2021, European hospitals were using GI Genius and similar systems for routine colonoscopies, particularly in high-volume screening centres (Schöler et al., 2023[52]).
Regulatory approval and evidence are key challenges, but as more research is published, confidence grows. However, improvements observed in controlled trials (often done by high-level endoscopists) must translate to everyday practice, including among less experienced practitioners or during training of new endoscopists.
Insights from interviews
Most commonly adopted use-cases
Interviewees describe an EU medical-imaging AI market that has consolidated around two main categories of applications.
The first focuses on image interpretation and decision support. Interviewed companies are mainly developing instruments that present the AI-augmented findings at a second step of the process – after the radiologist has reviewed the original X‑ray image. One company detailed how its software duplicates every original X‑ray image received by the system. This approach ensures that the radiologist’s primary reading remains unbiased, maintaining a “human‑in‑the‑loop” model in which AI supports rather than dictates the diagnosis. In practice, the algorithm acts as a second pair of eyes. According to internal audits with an independent platform, the proportion of missed fractures fell consistently during months when the tool was active compared with baseline periods when clinicians worked unaided.
A second family of workflow‑automation tools targets the cumulative friction points that slow down imaging pathways. One company, for example, breaks down the full CT scan process into patient positioning, image acquisition, reconstruction, triage and reporting, applying AI at each step. This approach reportedly saved 16 minutes per scan. While each step may only gain a small amount of time, the overall impact is significant across the full care pathway. Screening programmes use a similar approach: pre-analysis algorithms, for instance, are used to re-order mammogram queues so that images with signs of possible cancer are reviewed first. This helps radiologists prioritise urgent cases, shorten reporting sessions and reduce the stress patients feel while waiting for results, according to interviews with industry representatives.
Both AI applications can reduce clinical time in a context of growing imaging volumes and staff shortages.
Key barriers and challenges
Stakeholders reported the adoption of imaging AI tools is advancing unevenly in the European Union due to five recurring barriers:
Data access and quality: effective deep-learning models require large, diverse and well-annotated imaging datasets, spanning various scanners and patient populations. However, in most EU countries, medical images remain siloed at the hospital level and are subject to varying national interpretations of the General Data Protection Regulation (GDPR). As a result, some start-ups opt to train their models on extensive archives from US-based teleradiology networks, where population variety and data accessibility are greater. This leads to European-developed algorithms being trained primarily on American data due to limited access to representative European datasets. Another developer noted a similar challenge: with few exceptions such as Estonia, where a national imaging archive links all hospitals, most researchers must negotiate with institutions individually. This process often results in fragmented and statistically narrow training cohorts.
Regulatory complexity: while AI solution providers already generate extensive clinical evidence to obtain CE certification under the MDR, they must now prepare for a parallel conformity pathway under the forthcoming AI Act. Industry associations warned that, in absence of harmonised guidance, the two regimes may lead to “duplicate audits” and to a shortage of notified‑body expertise to review deep‑learning systems. This regulatory burden is viewed as a critical threat to the survival of smaller European AI developers. One manufacturer also pointed to a gap associated with this fragmented landscape: unlike the US Food and Drug Administration (FDA), the European Union has no public database of authorised imaging algorithms. Hospitals therefore cannot verify which tools are certified, on what data they were trained or how they are performing in post‑market use. This leads to duplication of both development effort and procurement due diligence.
Digital infrastructure and readiness: many European hospitals rely on on‑premise servers to meet cybersecurity requirements, limiting the use of cloud connectivity that vendors need for monitoring and updating the model in real time. Interviewees contrasted this with consolidated US health systems that are rapidly moving entire imaging workflows to the cloud. EU users of the same tools, by comparison, must install local gateways and accept slower upgrade cycles.
Digital literacy: vendors reported investing significantly in on‑site training so that clinicians understand both the capacities and limitations of AI tools. However, small regional hospitals continue to face challenges in redesigning workflows around the new tools. Thus, even where technical capacity exists, digital literacy may be a limiting factor.
Economic concerns: procurement rules in several Member States set mandatory competitive tendering processes once contract values exceed the predetermined threshold. According to one developer, this results in pricing strategies that deliberately undervalue the clinical impact of their products to avoid lengthy procurement delays. This, in turn, “artificially puts a cap on the cost of the product” and risks encouraging market consolidation, as only larger firms can sustain operations under such margin pressures. A related issue is budget allocation within hospitals. Although AI tools often deliver benefits across several departments (e.g. emergency, radiology, hospital administration), each department typically manages its own budget, thus complicating shared investment in AI technologies.
Key recommendations
Transparency about CE-marked AI imaging algorithms emerged as a primary recommendation. Several stakeholders called for a single European portal modelled on the US FDA database for AI and machine learning (ML) tools. Such a portal would list all imaging algorithms that hold a CE mark, together with plain‑language summaries of their training data, clinical performance evidence and post-market monitoring indicators. Stakeholders noted that such a portal could help reduce duplicative development; support hospitals in benchmarking available tools quickly; and offer smaller vendors a credible and transparent way to demonstrate regulatory compliance.
Closely linked to transparency is the need for greater regulatory coherence. Interviewees recommended that policymakers ensure the AI Act and the MDR share notified‑body expertise, harmonised guidance and, where feasible, a single technical dossier. This alignment would help avoid duplicative conformity assessment procedures for the same product. Industry associations further highlighted the importance of regulatory sandboxes, enabling start-ups to test evidence-generation strategies under the joint oversight of medical device and AI regulators before committing to full clinical validation.
A third priority is improved access to data. Companies recognised the forthcoming European Health Data Space (EHDS) as a critical opportunity to standardise secure access to diverse imaging datasets. The national imaging archive in Estonia was frequently cited as an example of best practice, demonstrating the potential of a centralised system that aggregates and shares imaging data at the population level.
Finally, all participants underlined the importance of workforce readiness. Stakeholders proposed the establishment of EU‑wide funding programmes to build multidisciplinary skillsets among healthcare professionals, to ensure they can integrate and apply AI tools effectively in clinical settings. Scaling human capabilities is seen as equally vital as deploying the technology itself; without adequate training and workflow adaptation, hospitals may install AI tools but fail to realise their full efficiency gains.
AI in healthcare operations
AI is gradually becoming part of day-to-day operations in EU healthcare systems. Across EU hospitals and health services, AI tools are informing both clinical decision making, and administrative and logistical tasks. From managing beds and staffing to scheduling surgeries and transcribing clinical notes, these technologies promise to streamline workflows and alleviate long-standing inefficiencies.
Demonstrating the practical application of AI in patient-facing services, Infermedica, a Poland-based health-tech company, provides AI-driven symptom triage and intake tools that combine probabilistic knowledge graphs with large language models (LLMs) to support patients and healthcare professionals (Box 3.2).
Interest in AI for healthcare operations has increased in recent years. A forthcoming OECD report highlights that EU countries are advancing the adoption of AI for administrative uses. However, adoption is mostly limited to local or regional centres (OECD, forthcoming[21]).
Adoption has been accelerated in part by the COVID-19 pandemic, which forced hospitals to embrace data-driven planning and digital innovations. Indeed, the pandemic helped accelerate the growth, adoption and scaling of AI in EU healthcare systems by breaking down certain barriers and fostering greater openness to data-driven solutions (EIT Health, 2021[53]). Operational AI deployments are aimed not at replacing health workers, but at supporting them. They automate routine tasks and free clinicians to focus on patient care (Klumpp and Hintze, 2021[54]) or to take breaks, reducing the risk of burnout.
Infermedica is a digital health company specialising in AI-powered solutions for symptom analysis and patient triage. It adopts neuro-symbolic AI in its solutions, particularly in its Conversational Triage tool.
Neuro-symbolic AI is a hybrid approach that combines neural networks with symbolic reasoning. Neural networks are excellent at processing unstructured data such as text or speech. They enable machines to learn patterns from large datasets, making them suitable for interpreting natural language. Symbolic AI uses structured representations like rules, logic and ontologies to perform logical inference. This method offers transparency and the ability to apply predefined, human-understandable reasoning. By merging these two approaches, the result is a more robust and interpretable AI model that can handle complex tasks while maintaining clarity in its decision-making process.
In Infermedica’s case, this manifests in a solution that combines large language models with Bayesian knowledge graphs. The LLM processes patient inputs in natural language, recognising symptoms and health concerns. Meanwhile, the symbolic layer, built on a curated medical knowledge base, applies clinical reasoning to determine the urgency of the situation. This allows the system to provide patients with an accurate and explainable triage recommendation, such as self-care, primary care or emergency services. This approach enhances patient engagement through conversational interfaces and offers transparency by explaining how conclusions are reached.
Source: (Infermedica, 2025[55]), “Moving towards clinically validated neuro-symbolic AI”, https://infermedica.com/blog/articles/towards-clinically-validated-neuro-symbolic-ai.
Main use-cases reported in literature
Hospital resource allocation
One of the most critical operational challenges in hospitals is how to allocate limited resources – particularly bed capacity and healthcare staff – in the face of fluctuating patient demand. One business association underlined how AI‑powered tools can analyse bed occupancy, Intensive Care Unit (ICU) demand and patient flows in real time to help hospitals anticipate and manage spikes in admissions. In the European Union, several hospitals have piloted ML models to forecast inpatient admissions and occupancy levels. For example, predictive systems were pressed into service during the COVID-19 pandemic to manage surges.
Gemelli University Hospital in Rome, the second largest hospital in Italy, used AI-based analytics to forecast intensive care admissions and overall bed occupancy during the pandemic (SAS, 2020[56]). The system monitored real-time admissions, discharges and transfers of COVID-19 patients and visualised trends in ICU bed usage. This helped administrators plan staffing and allocate wards for COVID care versus other services (SAS, 2020[56]). The ability to accurately predict patient inflow over a given planning horizon was vital to anticipating capacity constraints and reallocating resources in time (Redondo et al., 2023[57]).
This experience echoes a broader finding in the literature that AI methods can significantly improve hospital capacity planning. Research surveys have concluded that “AI-based methods have been successfully developed to address several healthcare logistics problems such as appointment planning, patient and resource scheduling, resource utilisation and predicting demand for emergency departments or ICUs” (Klumpp and Hintze, 2021[54]). By leveraging patterns in hospital admission data, ML models can discern subtle trends (e.g. how seasonal illnesses or local events drive patient volumes) and thus forecast bed occupancy or emergency arrivals more reliably than manual heuristics.
Despite these benefits, the adoption of AI in resource allocation is still in its early stages across much of the European Union. The level of deployment varies widely by country and hospital. Outside the European Union, the UK NHS has been a frontrunner in piloting predictive tools for beds and staffing, aided by national initiatives and funding for AI in healthcare operations. Countries like Italy and Germany reported pilot projects in large university hospitals, often in collaboration with tech companies or research institutes (SAS, 2020[56]; Klumpp and Hintze, 2021[54]). However, many smaller or less resourced hospitals in the European Union have not yet implemented such systems, citing barriers such as lack of technical expertise, limited budgets or insufficient data infrastructure. Those hospitals that have adopted AI tools tend to be tertiary referral centres or part of special innovation programmes.
The challenges associated with operational AI in healthcare are significant. A primary concern is data quality: predictive accuracy depends on rich, clean historical data. In some regions, hospital data may be fragmented across facilities or not recorded in a usable form, diminishing performance of the AI model. While widespread implementation of EHRs is improving the situation, data silos often persist within hospitals, particularly between departments. Moreover, lack of understanding can hinder usage. For example, if bed managers do not understand how an AI model generates its forecast, they may be reluctant to rely on it for critical decisions. Enhancing model transparency and investing in user training are therefore essential. Another challenge is that healthcare is inherently uncertain; unexpected events (mass casualty incidents, disease outbreaks) can upend any forecast. As such, AI should be used as a decision-support tool, with planners maintaining the flexibility to adapt to unexpected developments.
Clinical scheduling
Efficient scheduling is at the heart of hospital operations, influencing how quickly patients can receive care and how well critical resources like operating theatres and clinic slots are used. Traditionally, healthcare scheduling has faced challenges, including long waiting lists, high rates of missed appointments (“no‑shows”) and suboptimal use of facilities. In recent years, AI and advanced algorithms have begun to tackle these issues in EU healthcare, introducing data-driven intelligence into what was once a purely manual scheduling process. AI systems in this domain draw on techniques such as ML prediction, optimisation algorithms and reinforcement learning to allocate appointments or surgery times in an optimal manner. EU eHealth reports highlight that AI can “automate and optimise operations”, including patient scheduling (European Commission, 2024[58]). This could address inefficiencies that contribute to waiting times and access problems.
Spain has been home to innovative projects like Preplex, an EU-funded initiative at a Madrid hospital to optimise scheduling through AI (InnoBuyer, 2025[59]). In its initial phase, Preplex compared AI-generated clinic schedules with the hospital’s usual manual schedules and evaluated metrics such as slot utilisation and waiting times (InnoBuyer, 2025[59]). The AI schedule resulted in up to 60% fewer empty slots and a 22% increase in appointment availability. It organised the schedule more efficiently, making more appointment slots open to patients without adding extra hours (InnoBuyer, 2025[59]). To overcome staff reluctance and ensure the clinicians and clerical staff embraced the new tool, the team provided training and an intuitive interface (InnoBuyer, 2025[59]). Several Dutch hospitals are already using AI tools to flag patients who are likely to miss their appointments (Aij, Knoester and Werkhoven, 2024[60]). University Medical Center Utrecht first deployed the system in the lung clinic, where it cut missed visits by about 30%. The hospital plans to roll out the model across all departments, aiming for a hospital-wide, no-show reduction of roughly 15-20% (DutchNews, 2024[61]).
A specific application is in surgical scheduling and operating theatre management. Operating rooms (ORs) are among the costliest resources in a hospital, and efficient scheduling of surgeries can improve both financial performance and patient outcomes (by reducing delays for operations). AI has been applied to predict surgery durations more accurately and to sequence cases in an optimal way. A well-known study at Duke University in the United States showed that ML models could predict the length of surgical procedures with 13% greater accuracy than human schedulers, enabling the hospital to reduce overtime and idle OR time (Zaribafzadeh et al., 2023[62]). In the European Union, a German hospital reported that using AI for preoperative planning improved use of its operating theatres and reduced overtime work; the algorithm’s objective recommendations had removed some of the guesswork in scheduling (Bundesministerium für Bildung und Forschung, 2021[63]). Moreover, by anticipating which surgeries might require critical resources (e.g. an ICU bed post-operatively), the scheduling system can align surgical bookings with downstream resource availability. In essence, this is a move towards optimal use of operating theatres and less overtime.
Despite the promising outcomes, scaling AI scheduling across EU healthcare systems presents several challenges. One of the primary technical barriers is integration with electronic booking systems and EHRs. Many hospitals use older scheduling software, requiring custom interfaces for AI modules. Another concern is the potential for algorithmic bias. Historical scheduling data may reflect existing inequities, such as certain patient groups being less likely to receive appointments during preferred time slots. In these cases, AI models trained on such data may inadvertently reinforce or amplify those disparities. Furthermore, the variability of healthcare demand can affect model reliability. AI models perform best with stable patterns, yet healthcare can be chaotic. For that reason, most EU hospitals using AI for scheduling follow a “human-in-the-loop” approach, where staff have final say and can override the AI outputs when necessary.
Automated notetaking and transcription
Medical professionals spend a substantial portion of their time on documentation: writing patient notes, discharge summaries, referral letters and updating EHRs. This clerical burden is often cited as a contributor to physician burnout and limits the time clinicians can devote to direct patient care. AI has increasingly been deployed to ease documentation tasks through technologies like speech recognition, natural language processing, and more recently, generative AI for summarisation. In the European Union, the uptake of AI for clinical documentation has been noticeable in both hospital and primary care settings. Many such institutions adopt speech-to-text solutions in native European languages and pilot “ambient” AI assistants that automatically draft consult notes.
One of the most widespread uses of AI in documentation is medical speech recognition. Companies such as Nuance have long offered speech-to-text software that allows clinicians to dictate their notes instead of typing. European languages pose challenges for accuracy, but sustained development has led to high‑quality solutions for major languages like Spanish, French, German, Italian and Dutch, as well as the Nordic languages. For example, Norway was an early adopter of clinical speech recognition: the Sykehuset Telemark Hospital integrated a speech-to-text system (Nuance’s SpeechMagic) with its electronic medical record. The hospital reported saving approximately EUR 900 000 per year on transcription costs after rolling out the AI speech recognition system (Healthcare in Europe, 2009[64]). Moreover, document turnaround times improved markedly: 90% of all medical reports were delivered to referring physicians within seven days, exceeding the national target, because the reports were generated immediately via dictation rather than waiting for manual transcription (Healthcare in Europe, 2009[64]). Quality did not suffer; in fact, a study at that hospital found that discharge summaries produced with front-end speech recognition were rated as equal or better in content and layout compared to those transcribed by secretaries, with an average 9% improvement in document quality across several departments.
Beyond transcription, recent advances in natural language processing, particularly transformer-based models and generative AI (like GPT-like systems), are giving rise to “AI scribes” or ambient documentation assistants. These tools listen to the conversation between doctor and patient and automatically generate a draft clinical note or letter from that dialogue. Regulatory attitudes towards AI scribes vary: in the United States, they are generally regarded as low-risk medical software. Conversely, they are seen as higher-risk applications in the European Union, a difference that is slowing adoption on the continent. While this technology is not used in the European Union, the United Kingdom is launching pioneer projects. In late 2024, Great Ormond Street Hospital (GOSH) in London worked with an AI company to launch the first NHS trial of an ambient voice assistant. The assistant, named TORTUS, automates the writing of outpatient clinic notes (GOSH, 2024[65]). During a consultation, clinicians at GOSH no longer need to type or dictate into the computer; instead, the AI system “listens” to the conversation, summarising the discussion and producing a draft clinic letter by the end of the appointment. Clinicians then review and edit the AI‑generated text for accuracy before it is finalised in the patient’s record. Early results from the simulation and pilot phase were positive. Clinicians reported the AI assistant helped them give full attention to patients without having to worry about notetaking, and that the quality of the notes remained high.
A recent survey covering nine EU countries (plus the United Kingdom and Australia) found that patients are largely open to clinicians using AI for documentation if it means the doctors can pay more attention to them. Over one-third of patients surveyed said they favour doctors using AI in consultations to improve documentation processes (Beer, 2024[66]). Notably, this study reported an average of 34% of respondents believing AI documentation support is a good idea. On average, 45% of respondents felt the most compelling benefit of AI in documentation would be giving the doctor more face-to-face time with them (Beer, 2024[66]).
In the adoption of AI in clinical documentation across the European Union, accuracy and safety are paramount, as an error in a clinical note or a missed detail could affect patient safety. Thus, clinicians must carefully review AI-generated notes. Early speech recognition systems sometimes struggled with medical jargon or strong accents, leading to potentially dangerous transcription errors. While accuracy rates have improved, vigilance is needed especially when deploying in new languages or dialects. The issue of language localisation is indeed a particular challenge in Europe. English-language AI tools are often the most developed, but equivalent performance in less diffused idioms requires development and training data in those languages.
Data privacy is another critical concern. Using ambient AI assistants means potentially recording doctor-patient conversations, which are sensitive personal health data. In the EU context, strict compliance with the GDPR is essential. In most cases, explicit patient consent is required prior to recording any part of a medical consultation. The audio data may need to be processed locally or in a secure computing environment to avoid any risk of unauthorised access.
Insights from interviews
Most commonly adopted use-cases
Interviewees consistently described operational AI as a key enabler for delivering leaner and safer hospital services. Some manufacturers described systems that forecast seasonal peaks (e.g. influenza) and then dynamically rebalance beds, staff rosters and equipment across wards. AI “…starts to re‑arrange resources in the hospital more efficiently” when epidemiological signals predict a surge. In this way, radiology time, critical‑care beds and even ancillary services are re‑optimised overnight. Interviewees see similar promise for hospitals, and link AI‑driven analytics with “co‑ordination of care” and “value for money”. They argued that hospitals will need to leverage real-time analytics to reduce adverse events, smooth patient flow and maintain financial sustainability.
A second cluster of use-cases focuses on determining which patients should be seen first, where and when. One interviewee described this as “low-hanging fruit” for AI, with the potential to relieve clinicians of repetitive administrative tasks. Examples included AI tools that sequence imaging appointments; prioritise alerts from abnormal monitors; and re-order surgical lists to ensure that time is used for direct patient care rather than logistical co‑ordination. Start-ups reported similar efforts within emergency departments. One company reported that its fracture‑detection model allows emergency physicians to triage negative X‑rays immediately, thereby freeing treatment spaces and reducing admission queues. Another health-tech start-up reported the deployment of a triage and intake solution that combines LLMs with a symbolic AI engine to streamline patient interviews, support call centre nurses and prioritise cases before a visit. Hospital administrators recognised benefits, such as reduced length of stay. However, they noted challenges in determining which departmental budget should fund tools that span imaging, emergency care and administrative functions. Interviewees therefore stressed that scheduling-related AI must not only deliver accurate outputs but also integrate seamlessly with hospital information systems and EHRs.
Stakeholders converged on identifying clinical documentation load as the most immediate and culturally acceptable target for AI deployment. Automated notetaking was described as “the entry point” for generative AI because it “repositions time to spend it with the patient” without touching the diagnostic core. One business association confirmed “widespread development of voice‑recognition systems to support doctors in their administrative and clinical tasks”. One company showcased a hybrid system architecture in which an LLM transcribes free-text conversations, while a certified Bayesian engine handles clinical reasoning to mitigate the risk of hallucinations.
Finally, at the enterprise level, companies are also employing internal LLMs to assist with the drafting of clinical study reports and quality review documentation. Some have reported time savings of up to 50% for specific document modules.
Key barriers and challenges
The most frequently cited obstacle to the routine application of AI in healthcare operations is limited access to reliable, well-labelled operational data. Hospital information remains distributed across disparate systems, including EHRs, imaging repositories and administrative databases. Integrating these systems often requires customised interfaces and manual workarounds, creating significant inefficiencies. One interviewee emphasised that data access and interoperability constitute the most significant technical challenge both in clinical AI applications and also in broader domains such as manufacturing, supply chain management and commercial operations.
Regulation represents a second possible constraint, but there are two opposing views about this. On one hand, some stakeholders expressed concerns regarding the interaction of the EU AI Act with regulatory frameworks for medical devices and data protection. Medical technology firms highlighted uncertainty around how new conformity assessment pathways, notified body capacity and clinical evaluation requirements will align with the continuous software updates and cloud-based deployments that are common in hospital environments. On the other hand, several start-ups viewed regulation as a strategic opportunity to differentiate themselves rather than as a hurdle. One company noted that strict regulatory compliance is essential for building trust and ensuring patient safety. They saw the EU medical device certification process as providing a competitive advantage, assuring clients that their system’s recommendations are explainable, reproducible and clinically validated. According to this interviewee, regulation acts as both a quality safeguard and a market enabler, particularly in high-stakes environments where transparency and accountability are central to adoption.
A third recurring barrier is the shortage of professionals who combine operational know‑how with data‑science and regulatory literacy. According to one business association, a new generation of hybrid experts is needed to translate across clinicians, managers and data engineers. Otherwise, AI projects will stall through mistrust or be under‑used once deployed.
Key recommendations
A truly European health data infrastructure: companies expressed a strong desire for the EHDS to move swiftly from legislative proposal to operational, standards-based infrastructure. Specifically, they called for a system that enables authorised users to query pseudonymised operational data across borders. Interviewees emphasised the importance of building on existing interoperability standards and avoiding new proprietary formats. This, they argued, would allow AI tools to be integrated into hospital systems just once rather than requiring separate adaptation for each Member State.
Sustained investment in human capital to accompany legal and technical reforms: interviewees advocated for EU-supported master’s programmes and modular training courses. These should equip clinicians, biomedical engineers and hospital administrators with the skills to commission, evaluate and monitor AI systems effectively. One concrete proposal was to incorporate operational AI literacy into hospital quality accreditation frameworks, generating institutional demand for such training.
AI in drug discovery
From large pharmaceutical companies to biotech start-ups and academic institutions, stakeholders across the European Union are leveraging AI tools to accelerate drug discovery and improve efficiency of clinical trials. AI can be leveraged as a powerful tool throughout the entire lifecycle of a medicine – from R&D to manufacturing and post-approval activities (Figure 3.7).
Source: (EFPIA, 2024[67]), “EFPIA position on the use of artificial intelligence in the medicinal product lifecycle”, https://efpia.eu/media/tzeavw1t/efpia-position-on-the-use-of-artificial-intelligence-in-the-medicinal-product-lifecycle.pdf.
Main use-cases reported in literature
Target identification and validation
Identifying and validating novel biological targets, such as the genes or proteins that a new drug will act upon, is a critical first step in drug discovery.
Whereas traditional target discovery often relies on laborious experiments or analyses based on single data sources, AI can integrate multi-dimensional data (genomics, proteomics, clinical data, literature) to improve the precision of target identification. ML models (including deep neural networks and graph-based algorithms) are used to find patterns linking certain genes or pathways to disease outcomes. These models can both generalise from established biological knowledge and generate novel hypotheses for experimental validation.
Several EU-based initiatives further demonstrate the uptake of AI in target discovery. Sanofi entered a EUR 4.6 billion collaboration with UK-based Exscientia in 2022, which tasked Exscientia’s AI platform with leading the discovery of disease targets in oncology and immunology (Smith, 2022[68]). Smaller European biotechs are also active. Turbine (Hungary), for instance, uses an AI-driven biological simulation platform to predict cellular responses to perturbations, including drugs and genetic interventions (Box 3.3). Turbine builds in-silico models of cancer cells to understand downstream viability and phenotypic changes after a drug binds its target.
A specific limitation for this use-case is data quality and bias. Many biomedical datasets often suffer from representational imbalances. For example, genomic studies might over-represent certain populations, leading AI to miss targets relevant to under-represented groups. If an AI algorithm predominantly learns from well-studied pathways, it may bias outputs towards “popular” targets rather than truly novel targets.
Drug candidate generation (molecule design)
Once a biological target is identified, the next step is designing a molecule to modulate that target. Traditionally, this process involved laborious trial-and-error and relied heavily on human expertise and intuition. However, the integration of AI is transforming this stage into a far more efficient and scalable process. For example, Sanofi’s LLM CodonBERT, pre-trained on 10 million mRNA sequences, has reportedly halved mRNA design time.
AI systems can now autonomously generate novel chemical structures with predefined properties, guided by training on vast chemical and biological datasets. These algorithms can optimise not just efficacy and target affinity, but also factors like solubility, toxicity and synthesisability.
European start-ups are playing a prominent role in advancing AI-driven molecule design. The French company Iktos uses generative AI to create synthetically accessible molecules through multi-parametric and 3D-constrained optimisation (Iktos, 2024[30]). For its part, Nuritas in Ireland applies AI to discover and design bioactive peptides – short chains of amino acids with therapeutic or health-promoting properties (Nuritas, 2024[69]). These approaches are no longer confined to small molecules; AI is increasingly being used to design biologics, such as therapeutic proteins and vaccine antigens, as exemplified by Evaxion Biotech in Denmark (Evaxion, 2024[70]).
A persistent challenge in AI-driven molecule design is data fragmentation and confidentiality, as proprietary datasets are often siloed within pharmaceutical firms. To address this issue, the MELLODDY consortium developed a secure, federated learning platform. It allows ten EU-based pharma companies to train AI models on over a billion data points without sharing proprietary data. This collaboration improved hit prediction while preserving privacy (Heyndrickx et al., 2023[71]; OECD, 2025[72]).
Several other challenges for this specific use-case were noted. First, AI designs may fail in lab testing due to unmodeled factors. Second, some proposed compounds may also eventually be hard to synthesise. Teams are tackling this challenge by combining AI with lab validation and synthesis-predictive tools. Finally, data scarcity, particularly for novel targets, limits accuracy, highlighting the importance of ongoing efforts in data sharing, augmentation and standardisation.
Turbine AI, a Hungary-based deep‑tech company, has developed the Simulated Cell™ platform, a cloud service that converts detailed, “mechanistic” models of cancer cells into a fully-fledged virtual laboratory. Using layers of multi‑omics data (DNA, RNA, proteins, epigenetics) and published perturbation studies, the engine learns how signalling pathways behave in different tumours. A virtual lab enables in‑silico experiments using a vast library of cell models, patient-derived samples and virtual patients. By computationally predicting therapy effects, drug developers can focus their resources and substantially increase the likelihood that new treatments will make it to the patients who need them the most.
Target selection is one of the main solutions for Turbine. When a partner is looking for first-in-class biology, the first task is spotting a weak point in the cancer cell, i.e. something that the tumour needs to survive but that healthy tissue can live without. Turbine points its Simulated Cell™ at hundreds of possible genes or proteins at once, “silences” them one by one in cyberspace and watches which digital tumours die. Turbine spends about a month calibrating avatars to that cancer type. In its 2024 collaboration with Ono Pharmaceutical, several novel targets that emerged from this screen moved into Turbine‑led laboratory validation within 12 months. This demonstrated that the simulation loop shortens discovery timelines without losing biological depth.
Source: (Turbine, 2024[29]), “Turbine achieves key milestone in collaboration with Ono Pharmaceutical”, https://turbine.ai/news/turbine-achieves-key-milestone-in-collaboration-with-ono-pharmaceutical/; (Turbine, 2025[73]), “Expand my pipeline”, https://turbine.ai/target-selection/.
Trial design and protocol optimisation
Traditionally, clinical trial design has relied on expert heuristics and prior experience. AI and data-driven approaches are now providing tools to optimise trial protocols, especially to enhance success rates and improve the patient experience. Applications of AI in this area range from using real-world data to inform eligibility criteria, to simulating in-silico trials, to creating “digital twins” of patients for novel trial designs.
One application is using AI to refine eligibility criteria. Often, clinical trials impose strict criteria on who can participate (e.g. based on age, lab results, concomitant diseases) to ensure a homogeneous study population. However, this can make recruitment hard and may exclude patients who might benefit from the intervention. AI can analyse large-scale patient outcome data to identify which criteria are truly necessary for clinical validity. It can also help ensure broad representation by analysing how criteria affect different subgroups. Reports underline that “AI can inform clinical trial eligibility criteria, enhance the diversity of participants and reduce sample size requirements” (Zhang et al., 2023[74]). Such optimisation means trials could be smaller, faster and cheaper, without losing the ability to detect a drug’s effect.
Practical evidence of such benefits includes one company’s initiative that uses AI and digital technologies to redesign its clinical trial processes. This initiative accelerated site selection by 60% and employed a decision-support platform to improve recruitment strategies, targeting under-represented populations at the screening stage. A major EU-based company is also leveraging AI for patient stratification, especially in trials targeting complex or rare diseases. These models segment patients based on genetic, clinical or behavioural variables to assign optimal dosage arms within clinical trials.
Another promising development is the use of AI to create synthetic control arms. In many trials, a control group (often receiving placebo or standard of care) is needed for comparison. However, enrolling patients into a control arm can be difficult – patients generally prefer not to risk getting placebo, especially in trials for serious diseases. AI offers a solution by enabling external control or digital twin models. The digital twin represents what would happen to that patient if they received standard care. By comparing the outcomes of each treated patient to their twin, one can statistically infer the treatment effect with fewer or even no real control patients (Institut Polytechnique de Paris, 2025[75]). Companies underlined the usefulness of digital twins, for example, in reducing the number of technical, trial and qualification batches required to manufacture a treatment or vaccines. Precise guidelines regarding this use-case are lacking, leading companies to proceed cautiously.
Another challenge is the heavy dependence on data. Designing a trial via simulation or AI works best with rich real-world or historical data to feed into the models. Finally, trial design using AI can raise ethical questions, such as when AI tools suggest criteria to include patients in a trial or not.
Monitoring and data analysis during trials
AI is enabling more efficient data management. Running a clinical trial generates a vast amount of data – from patient health measurements and laboratory results to compliance logs and safety reports. Traditionally, statisticians collected much of these data on paper or in basic electronic systems and analysed them retrospectively. Now, AI is increasingly being harnessed to monitor trials in real time and gain insights from complex, high-frequency data streams, improving patient safety and trial efficiency.
One prominent use of AI is in remote patient monitoring during trials. With the rise of wearables and telemedicine, trials are collecting continuous data on patients (e.g. vital signs, activity levels, glucose readings) outside of clinic visits. AI algorithms are essential to interpret this amount of data, and real-time analysis could allow interventions (e.g. adjusting medication or checking for adverse events) much sooner than waiting for the next scheduled visit.
Another area is safety monitoring and pharmacovigilance within trials. Business associations stressed how AI systems can aggregate reports of adverse events in real time and look for signals that might not be obvious on a case-by-case basis. This, in turn, highlights unexpected adverse effects sooner than manual review. For instance, if a few patients exhibit a slightly elevated liver enzyme, an AI might detect a trend that each occurred after a certain number of doses, indicating a possible drug-related liver effect emerging.
Decentralised trials are also being pioneered across the European Union. For example, the Trials@Home project, co‑ordinated by University Medical Center Utrecht and involving partners including Sanofi, explores the feasibility and effectiveness of decentralised and hybrid trial models using digital and AI tools across Europe. Participants in Trials@Home are provided with wearables, mobile health apps and teleconsultation systems to capture health data, report outcomes and interact with clinicians from their homes. The project aims to demonstrate how decentralised approaches can increase access, reduce burdens on participants, improve data quality, and maintain robust regulatory and ethical standards across different trial contexts (Innovative Health Initiative, 2019[76]).
These advancements also introduce new challenges. Data privacy becomes even more critical, given the continuous collection of sensitive health data. Another key issue is the validation of AI algorithms used for trial monitoring, i.e. ensuring they are accurate, reliable and fit for clinical use. On the operational side, clinical trial teams increasingly require new skill sets: data scientists may be needed to interpret AI outputs or staff may require upskilling to use these tools effectively.
Insights from interviews
Most commonly adopted use-cases
Business associations have emphasised the growing role of AI in drug discovery, noting that advanced algorithms enable researchers to identify promising molecular candidates early in the pipeline. By predicting molecular interactions with targets and eliminating those likely to fail later, AI tools can significantly shorten discovery timelines and lower R&D costs. Interviewed companies reported that they are increasingly approaching AI as a comprehensive, end-to-end accelerator across the early drug discovery continuum rather than as a series of isolated experiments. One of the primary areas of adoption is the systematic, data-driven nomination of disease targets. For example, one company described how its AI-powered target discovery engines integrate multi-omic datasets with laboratory validation. This identifies seven novel, first-in-class targets within a single year, compressing a process that traditionally spanned several years into a single R&D cycle.
Once a target is proposed, companies are turning to generative design engines that explore chemical and biological space at speeds unmatched by human teams. One company reported that its company’s LLM has halved the time needed to design optimised mRNA constructs. Simultaneously, its small-molecule reinforcement learning workflow iterates lead series with integrated consideration of potency, toxicity and manufacturability. Interviewees also highlighted a growing shift towards AI-driven antibody and peptide generation. They estimated that five to six years of accumulated innovation has already produced up to a 25% increase in discovery efficiency and speed. These gains are no longer confined to chemistry; several research groups are now using graph neural network representations of protein interaction maps to design degrader molecules, vaccine antigens and bispecific biologics.
A third, fast‑growing application of AI in drug discovery involves mechanistic simulation. In‑silico avatars connect genomic context, pathway topology and drug‑target binding to predict downstream viability, resistance and synthetic‑lethal interactions. An interviewed company emphasised that 19 of 20 oncology trials fail because the underlying biology is misunderstood, not because a compound misses its target. Their platform focuses therefore on what happens after binding, offering evidence‑based disease positioning and rational drug-combination proposals. Another company drew attention to knowledge‑graph initiatives that interlink thousands of heterogeneous European datasets to answer questions that traditional statistics could not address – from novel‑target co‑morbidities to cross‑trial safety signals.
Key challenges
Data access and quality remain significant challenges in the application of AI to drug discovery. Several companies reported that many foundational datasets are heavily skewed towards US populations, with European samples representing only a small fraction of major reference collections. Moreover, privacy regulations often preclude access to raw, patient-level data. This constraint compels firms to work with partially de-identified or harmonised datasets that frequently lack critical covariates. One executive estimated that only 10% of candidate targets reach clinical trials, and of those, merely 10% achieve regulatory approval. Broader, cleaner and more interoperable data streams are required to address this high attrition rate.
Even when data are available, benchmarking AI performance presents additional challenges. Distinguishing genuine predictive power from spurious correlations is not straightforward. One company reported developing a neural network that initially appeared to predict drug-target activity with over 90% accuracy. Upon further investigation, the model was found to be exploiting subtle “fingerprints” unique to individual laboratories, such as variations in equipment or technique, to identify the origin of samples rather than learning meaningful biological relationships. To mitigate such risks, the company now tests each new model against simple baseline models and applies statistical techniques to detect inter-laboratory bias. While interviewees widely agreed on the need for rigorous benchmarking, the sector has yet to converge on standardised evaluation protocols.
Technical capacity is another emerging constraint. Both start-ups and established firms reported that the rapid growth in demand for LLMs has strained access to cloud-based graphics processing unit (GPU) resources, which are essential for molecular simulations and graph-based model training. One company noted that it faces multi-hour delays in accessing multi-GPU nodes that were previously available on demand. This has led to extended project timelines and higher computing costs.
Regulatory uncertainty compounds these technical challenges. Companies expressed concerns about the interpretation and implementation of the EU AI Act, particularly in relation to its intersection with medicinal product legislation. One firm noted the ambiguity surrounding AI models used strictly as internal R&D tools, i.e. tools that do not function as stand-alone clinical decision systems but nonetheless fall within the potential scope of new regulatory frameworks.
Key recommendations
Create a data-sharing infrastructure: interviewees consistently identified data-sharing infrastructure as the highest-leverage intervention for accelerating AI in drug discovery. They advocated an EU‑wide federated architecture, building on the EHDS vision, that would allow model parameters to be sent to data rather than transferring data across borders. This infrastructure could integrate Privacy-Enhancing Technologies, including federated learning, homomorphic encryption and synthetic data generation. This, in turn, could enable model development and validation without exposing proprietary datasets or sensitive patient information (OECD, 2025[72]). Such an approach would preserve privacy and intellectual property boundaries while allowing AI systems to gain statistical power through access to diverse, distributed datasets across Member States.
Target investments in high-performance computing infrastructure: cloud GPU shortages were highlighted as a critical bottleneck. Delays of hours or even days in accessing compute resources undermine the very “time-to-insight” advantage AI is intended to deliver. To address this, interviewees recommended targeted investments in high-performance computing infrastructure. This should take the form of dedicated GPU clusters or “AI factories”, to support molecular simulations, graph-based learning and other computationally intensive applications.
Institutionalise meaningful AI benchmarking: interviewees proposed that regulators, research funders and companies agree on a few “gold‑standard” tests that every new model must pass. These could include, for example, predictive performance on laboratory experiments excluded from the training data, thereby testing model robustness across varying experimental conditions. Results from such benchmarks should be published on public leaderboards, enabling transparent, comparative assessment of AI models. This would not only help identify overfitted models but also promote best practices and provide regulators with a reliable, objective basis for scientific validation.
Improve regulatory agility: stakeholders called for dedicated regulatory sandboxes that would exempt AI systems used purely for internal R&D from full conformity assessments, while still requiring documented risk management. In parallel, the European Commission and Member States should issue interpretative guidance to ensure that the AI Act, the GDPR, clinical trials regulation and forthcoming EHDS provisions form a coherent and interoperable compliance framework (OECD, 2025[77]). Harmonised implementation would reduce administrative duplication, enable biotech firms to operate across EU markets from the outset and foster innovation without compromising safety or accountability.
Ensure talent development keeps pace with data and technological scale: stakeholder associations, particularly those representing hospitals and medical device manufacturers, emphasised that AI literacy should be recognised as a core component of clinical quality. They called on policymakers to invest in continuous professional development to equip healthcare professionals, regulators and trial monitors with the skills to engage with AI tools effectively and responsibly.
AI-related skills
The health workforce is entering a decade in which skill shortages risk becoming more limiting than budgetary constraints. AI tools can already triage symptoms, read X‑rays and optimise hospital logistics, yet their safe, effective deployment still hinges on the abilities of the people who design, procure, validate and use them.
Among a sample of EU countries considered in a forthcoming study, only two have embedded AI training into medical school curricula. Others are reforming their medical education and continuing professional development frameworks to integrate AI competencies (OECD, forthcoming[14]).
Drawing on recent evidence, as well as experiences reported by hospitals, start-ups and manufacturers, a coherent picture emerges of the three families of competences – technical, operational and organisational – that must mature together if AI is to be deployed both effectively and responsibly.
Technical competences
Technical capability is the foundation upon which all other activities depend. Securing high‑quality, interoperable data is the first priority. Clinicians, health informaticians and data engineers must be able to map local terminologies to international standards, such as SNOMED CT for clinical concepts and Fast Healthcare Interoperability Resources for data exchange. Medical associations surveyed by the OECD cited a lack of interoperability as a moderate-to-major barrier to AI adoption (Green, 2024[16]).
Once data are available, specialists in data quality management must clean, deduplicate and normalise records. Cleansing and cataloguing remain indispensable for reliable outputs. In parallel, data scientists and ML engineers need advanced knowledge of supervised, unsupervised and reinforcement learning techniques. This will allow them to design, train and fine‑tune models for diagnostics, triage and operational optimisation.
Clinical validation is another technical function. Together, engineers and clinician-scientists must create reference datasets, establish performance thresholds and run prospective studies that demonstrate safety and efficacy. Medical associations identified the absence of robust validation processes as a critical concern (Green, 2024[16]). This was especially true for algorithmic bias and the difficulty of monitoring systems that learn continuously. Skills in statistical hypothesis testing, prospective trial design and post‑market performance monitoring are therefore essential.
Cyber‑security and privacy protection belong in the technical domain as well. Ensuring compliance with the GDPR and emerging EHDS rules requires expertise in encryption, access control and data‑minimisation strategies. Surveyed associations linked weak privacy safeguards to lower patient trust and slower uptake (Green, 2024[16]). This underscores the need for security engineers who understand both clinical data flows and regulatory requirements.
Operational competences
Operational skills translate technical potential into day‑to‑day clinical benefit. First among these is workflow integration. AI systems must return results at a point in the care pathway where they can influence decisions without adding cognitive or administrative burden. Experience from radiology shows that algorithms performing well in stand‑alone tests are often disregarded when they disrupt established reading patterns or offer explanations that clinicians find unintuitive. Consequently, healthcare professionals require competence in human-technology interaction, usability assessment and the redesign of clinical processes.
Change management is equally important. OECD (2024[16]) highlights that health providers fear insufficient involvement in design and implementation of AI rather than being replaced by it. Nearly three‑quarters (74%) of respondents pointed to deficiencies in operational infrastructure and staff readiness as major obstacles to adoption. Managers must therefore be able to orchestrate pilot projects, gather feedback, refine workflows and document measurable gains such as reduced waiting times or shorter diagnostic cycles. They must also plan for continuous professional development. Digital systems evolve rapidly, and health workers will need modular, career‑long training pathways to keep pace with updates to data standards and AI management norms.
Operational competence also encompasses data governance in practice. Clinicians interacting with AI tools must know how to record overrides, flag unexpected outputs and escalate safety issues. This calls for clear standard‑operating procedures and audit trails that satisfy both clinical quality boards and external regulators.
Finally, patient communication is an operational task. New conversational agents and decision‑support systems alter the nature of clinician‑patient interactions, raising novel consent and explanation requirements. Staff therefore need well‑honed communication skills to describe probabilistic outputs and to reassure patients that humans retain ultimate responsibility.
Organisational competences
Organisational competences provide the strategic and governance framework within which technical and operational tasks can flourish. At the highest level, boards and senior executives must establish clear accountability for AI. The Sanofi RAISE framework, built around principles of fairness, transparency, robustness and environmental responsibility, exemplifies how corporate governance can embed ethical requirements into every project stage (Sanofi, 2024[78]). Health‑service providers will similarly need steering committees or dedicated AI offices with authority to drive responsibility across the entire AI lifecycle, from design through use.
Alignment with external regulation constitutes a further organisational responsibility. For European organisations, this means integrating the AI Act with existing MDR, in-vitro diagnostics rules and data‑protection law. Stakeholders warn that overlapping frameworks create compliance uncertainty and inconsistent national interpretations, making regulatory literacy a strategic asset. Legal teams, regulatory‑affairs officers and compliance managers must collaborate to develop harmonised documentation templates, maintain conformity assessment files and liaise with notified bodies.
Finally, organisational competence includes cultivating a culture of evaluation and improvement. Hospitals that have embedded AI successfully tend to treat every deployment as a living programme subject to outcome measurement, periodic audit and recalibration. Leadership must therefore mandate metrics, invest in analytic infrastructure and empower multidisciplinary quality‑improvement teams to act on the findings.
Key recommendations to enhance AI uptake in healthcare in the European Union
Data availability and access
Accelerate implementation of the European Health Data Space (EHDS): fast-track the EHDS roll-out with a clear governance model that ensures seamless, responsible access to health data for research and innovation. Build on federated architecture principles to preserve privacy, maintain intellectual property protections and enable statistical power through distributed access, while encouraging use of shared semantic standards for full interoperability.
Improve health data quality and representativeness: address critical issues related to the availability of high-quality, comprehensive health datasets to support AI development. This includes adopting common data quality standards, documenting data sources rigorously, ensuring broad population coverage and implementing systematic statistical bias identification and mitigation practices.
Strengthen comprehensive health data governance frameworks within and across Member States: develop robust governance structures that ensure timely data availability, quality and interoperability, while facilitating secure cross-sector exchange among public institutions, private entities and research organisations. These frameworks should combine technical infrastructures (such as the EHDS) with clear access conditions, standardised privacy and security protocols, interoperable data-sharing rules and co‑ordinated oversight across Member States.
Support open access knowledge bases: support the creation of public probabilistic medical knowledge bases to foster innovation and ensure cross-border AI compatibility. These could include structured, open medical repositories that underpin tools like probabilistic symptom checkers or AI-powered triage systems. Encourage contributors to document data provenance, population coverage and mitigation steps for any identified biases, thereby enhancing trust and usability across diverse clinical contexts.
Infrastructure
Expand AI factories: strengthen dedicated healthcare-focused AI compute and development centres with secure access to specialised datasets and high-performance computing resources.
Invest in cross-border collaboration: support collaborative research through dedicated funding.
AI tools and software
Foster public-private foundations for AI models: establish partnerships to develop foundation models in healthcare that are open, secure and designed for European contexts.
Establish an EU registry of AI approved solutions: create an EU-wide platform listing CE‑marked healthcare AI systems, modelled on the Food and Drug Administration device listing in the United States. This should include training data summaries, clinical validation and post-market surveillance to improve transparency, market confidence and clinical uptake.
Institutionalise benchmarking protocols: develop EU-wide “gold standard” model evaluation tests, such as generalisability to external laboratory data. Ensure that models can perform reliably across diverse clinical settings and are not limited to their original development environment. Publish results via public leaderboards to promote transparency and reproducibility.
Regulatory frameworks
Clarify the applicability of the research exemption in the AI Act, particularly regarding pharmaceutical research and development, and internal use of AI models.
Ensure clear pathways for compliance: develop consolidated guidance to streamline requirements across the General Data Protection Regulation, Medical Device Regulation and AI Act, and simplify navigation for developers and healthcare institutions. Share notified‑body expertise, harmonised guidance and, where feasible, a single technical dossier.
Consistent interpretation across the European Union: promote uniform application of AI‑related rules through centralised guidance and regulatory co‑ordination mechanisms.
Enable regulatory sandboxes: support experimental spaces where start-ups and researchers can test AI systems under regulatory supervision, including sandboxes under the joint oversight of medical device and AI regulators, while maintaining robust risk management.
Skills, trust and collaboration
Develop and retain AI talent in the healthcare sector: scale EU-wide AI education and training programmes, talent attraction schemes and reskilling initiatives to strengthen joint health and AI competences.
Support eHealth training across Member States: create transferable digital skillsets and develop European-level quality indicators to support continuing medical education and cross-border competency recognition.
Institutionalise AI literacy in healthcare quality standards: integrate operational AI literacy into hospital accreditation frameworks to ensure sustained demand for skill development and safe adoption.
Support trust building: invest in public communication campaigns and success stories that highlight responsible AI use to build societal confidence in AI in healthcare and promote responsible adoption.
Promote co-development and knowledge-sharing: encourage early engagement of stakeholders, including healthcare professionals and patients, in the design and deployment of AI technologies.
References
[37] Aidoc (2021), “Unfallkrankenhaus Berlin brings AI to Germany – Aidoc expands presence in Europe”, https://www.aidoc.com/about/news/ukb-adopts-aidoc-ai-solutions/ (accessed on 12 January 2025).
[60] Aij, K., J. Knoester and B. Werkhoven (2024), “An artificial intelligence-based model to reduce the no-show rate in outpatient clinics of an academic hospital”, Research Square, https://doi.org/10.21203/RS.3.RS-3743388/V1.
[50] Androulakis, E. and C. Fielder (2024), “Artificial intelligence in ECG diagnostics – where are we now?”, https://www.escardio.org/Councils/Council-for-Cardiology-Practice-(CCP)/Cardiopractice/artificial-intelligence-in-ecg-diagnostics-where-are-we-now#:~:text=The%20impressive%20recent%20advances%20in,as%20an%20ideal%20substrate%20for (accessed on 12 January 2025).
[66] Beer, H. (2024), “Third of patients in favour of AI-supported consultations and clinical documents, study finds”, https://hospitalhealthcare.com/news/third-of-patients-in-favour-of-ai-supported-consultations-and-clinical-documents-study-finds/#:~:text=Over%20a%20third%20of%20patients,from%20the%20Microsoft%20company%20Nuance (accessed on 11 January 2025).
[45] Behara, K., E. Bhero and J. Agee (2024), “AI in dermatology: A comprehensive review into skin cancer detection”, PeerJ Computer Science, Vol. 10:e2530, https://doi.org/10.7717/peerj-cs.2530.
[44] Better Medicine (2025), “Homepage”, https://bettermedicine.ai/ (accessed on 12 January 2025).
[24] BioNTech (2023), “BioNTech to acquire InstaDeep to strengthen pioneering position in the field of AI-powered drug discovery, design and development”, https://investors.biontech.de/news-releases/news-release-details/biontech-acquire-instadeep-strengthen-pioneering-position-field (accessed on 12 January 2025).
[63] Bundesministerium für Bildung und Forschung (2021), “KIOM”, https://www.interaktive-technologien.de/projekte/kiom (accessed on 12 January 2025).
[39] Busch, F. et al. (2024), “Navigating the European Union artificial intelligence act for healthcare”, npj Digital Medicine, Vol. 7/10, https://doi.org/10.1038/s41746-024-01213-6.
[61] DutchNews (2024), “Hospitals turn to AI to identify potential patient no-shows”, DutchNews.nl, https://www.dutchnews.nl/2024/05/hospitals-turn-to-ai-to-identify-potential-patient-no-shows/? (accessed on 31 March 2025).
[67] EFPIA (2024), “EFPIA position on the use of artificial intelligence in the medicinal product lifecycle”, European Federation of Pharmaceutical Industries and Associations, https://efpia.eu/media/tzeavw1t/efpia-position-on-the-use-of-artificial-intelligence-in-the-medicinal-product-lifecycle.pdf (accessed on 11 January 2025).
[53] EIT Health (2021), “Healthcare workforce and organisational transformation with AI – enacting change”, Round Table Series 2020 Summary Report March 2021, EIT Health, https://thinktank.eithealth.eu/wp-content/uploads/2021/04/EIT-Health-Think-Tank-Round-Table-Series-2020-Summary-Report.pdf.
[41] Esteva, A. et al. (2017), “Dermatologist-level classification of skin cancer with deep neural networks”, Nature, Vol. 542, pp. 115-118, https://doi.org/10.1038/nature21056.
[58] European Commission (2024), Artificial intelligence in healthcare, https://health.ec.europa.eu/ehealth-digital-health-and-care/artificial-intelligence-healthcare_en#:~:text=,effective%20treatment%20options (accessed on 11 January 2025).
[19] European Commission (2023), Case Study on the Digitalisation of Health (eHealth), European Commission, Brussels, https://commission.europa.eu/document/download/652a3175-d410-4608-8f0e-642049433c35_en?filename=case-study-on-the-digitalisation-of-health-ehealth.pdf.
[9] European Observatory on Health Systems & Policies and OECD (2023), “State of health in the EU: France”, https://eurohealthobservatory.who.int/publications/m/france-country-health-profile-2023 (accessed on 12 January 2025).
[6] Eurostat (2025), Beds in nursing and other residential long-term care facilities, https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_bdltc__custom_11992439/bookmark/table?lang=en&bookmarkId=90ccb567-1ba8-41e6-a4a8-ff5edb61d0ca.
[3] Eurostat (2025), Health care expenditure by financing scheme, https://ec.europa.eu/eurostat/databrowser/product/page/HLTH_SHA11_HF.
[10] Eurostat (2025), Health personnel, https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_prs2/default/table?lang=en&category=hlth.hlth_care.hlth_res.hlth_staff.
[7] Eurostat (2025), Hospital beds by function and type of care, https://ec.europa.eu/eurostat/databrowser/view/hlth_rs_bds1__custom_12047472/bookmark/table?lang=en&bookmarkId=c758c065-8b34-4d0a-ba88-06f850d5db99&c=1720078966412.
[12] Eurostat (2025), Population structure indicators at national level, https://ec.europa.eu/eurostat/databrowser/view/DEMO_PJANIND/default/table?lang=en.
[4] Eurostat (2024), “EU counted 2.3 million hospital beds in 2022”, https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20240711-2 (accessed on 11 January 2025).
[2] Eurostat (2024), “Healthcare expenditure statistics – overview”, https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Healthcare_expenditure_statistics_-_overview#Developments_over_time (accessed on 12 January 2025).
[5] Eurostat (2024), “Healthcare personnel statistics – physicians”, https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Healthcare_personnel_statistics_-_physicians (accessed on 12 January 2025).
[32] Evaxion (2025), “Homepage”, https://evaxion.ai/ (accessed on 11 January 2025).
[70] Evaxion (2024), “The power of AI-Immunology™ in vaccine target discovery, design and development”, https://evaxion.ai/ai-immunology (accessed on 13 January 2025).
[65] GOSH (2024), “GOSH pilots AI tool to give clinicians more quality-time with patients”, Great Ormond Street Hospital for Children, https://www.gosh.nhs.uk/news/gosh-pilots-ai-tool-to-give-clinicians-more-quality-time-with-patients/#:~:text=Typically%2C%20during%20outpatient%20consultations%20with,to%20patients%20and%20their%20families (accessed on 12 January 2025).
[16] Green, A. (2024), “Artificial intelligence and the changing demand for skills in the labour market”, OECD Artificial Intelligence Papers, No. 14, OECD Publishing, Paris, https://www.oecd.org/en/publications/artificial-intelligence-and-the-changing-demand-for-skills-in-the-labour-market_88684e36-en.html.
[13] Health, EIT; McKinsey & Company (2020), “Transforming healthcare with AI: The impact on the workforce and organizations”, https://eithealth.eu/ (accessed on 12 January 2025).
[64] Healthcare in Europe (2009), “Norway readily adopts speech recognition”, https://healthcare-in-europe.com/en/news/norway-readily-adopts-speech-recognition.html#:~:text=The%20Sykehuset%20Telemark%20HF%20hospital,the%20quality%20of%20discharge%20summaries (accessed on 12 January 2025).
[71] Heyndrickx, W. et al. (2023), “MELLODDY: Cross-pharma federated learning at unprecedented scale unlocks benefits in QSAR without compromising proprietary information”, Journal of Chemical Information and Modeling, Vol. 64/7, https://doi.org/10.1021/acs.jcim.3c00799.
[30] Iktos (2024), “Discovery platform”, https://iktos.ai/discovery-platform (accessed on 11 January 2025).
[55] Infermedica (2025), “Moving towards clinically validated neuro-symbolic AI”, https://infermedica.com/blog/articles/towards-clinically-validated-neuro-symbolic-ai (accessed on 15 January 2025).
[28] Infermedica (2025), “Provide the most accurate and efficient care with Infermedica’s symptom checker and virtual triage”, https://infermedica.com/ (accessed on 15 January 2025).
[59] InnoBuyer (2025), “PREPLEX: A data-driven approach optimizing healthcare scheduling with AI”, https://innobuyer.eu/preplex-a-data-driven-approach-optimizing-healthcare-scheduling-with-ai/#:~:text=Image (accessed on 13 January 2025).
[76] Innovative Health Initiative (2019), “Trials@Home”, https://www.ihi.europa.eu/projects-results/project-factsheets/trialshome (accessed on 12 January 2025).
[75] Institut Polytechnique de Paris (2025), “The digital twin, an ally for patients and doctors”, https://www.ip-paris.fr/en/news/digital-twin-ally-patients-and-doctors (accessed on 13 January 2025).
[22] JRC (2020), “AI Watch: AI uptake in health and healthcare, 2020”, https://ai-watch.ec.europa.eu/publications/ai-watch-ai-uptake-health-and-healthcare-2020_en (accessed on 12 January 2025).
[40] Kelly, S. (2019), “Medtronic rolls out AI-assisted colonoscopy system in Europe”, MedTech Dive, https://www.medtechdive.com/news/medtronic-rolls-out-ai-assisted-colonoscopy-system-in-europe/565431/#:~:text=FDA%20to%20date%20has%20approved,same%20accuracy%20as%20healthcare%20professionals (accessed on 12 January 2025).
[47] Kim, S. et al. (2022), “Fully automated quantification of cardiac chamber and function assessment in 2-D echocardiography: Clinical feasibility of deep learning-based algorithms”, The International Journal of Cardiovascular Imaging, Vol. 38/5, pp. 1047-1059, https://doi.org/10.1007/s10554-021-02482-y.
[54] Klumpp, M. and M. Hintze (2021), “Artificial intelligence for hospital health care: Application cases and answers to challenges in European hospitals”, Healthcare (Basel), Vol. 9/8, p. 961, https://doi.org/10.3390/healthcare9080961.
[48] Martínez-Sellés, M. and M. Marina-Breysse (2023), “Current and future use of artificial intelligence in electrocardiography”, Journal of Cardiovascular Development and Disease, Vol. 10/4, p. 175, https://doi.org/10.3390/jcdd10040175.
[26] Merck Group (2023), “Merck enters two strategic collaborations to strengthen AI-driven drug discovery”, https://www.merckgroup.com/en/news/two-ai-partnerships-in-healthcare-20-09-2023.html (accessed on 20 January 2025).
[46] Narang, A. et al. (2018), “Machine learning based automated dynamic quantification of left heart chamber volumes”, European Heart Journal – Cardiovascular Imaging, Vol. 20/1, pp. 541-549, https://doi.org/10.1093/ehjci/jey137.
[23] Nindl, E. et al. (2024), The 2024 EU industrial R&D investment scoreboard, Publications Office of the European Union, Luxembourg, https://doi.org/10.2760/0775231.
[31] Nuritas (2025), “Nuritas, See nature differently”, https://www.nuritas.com/ (accessed on 12 January 2025).
[69] Nuritas (2024), “The era of consumer centric intelligent ingredients”, https://www.nuritas.com/technology/bioactive/ (accessed on 12 January 2025).
[77] OECD (2025), “Facilitating the secondary use of health data for public interest purposes across borders”, OECD Digital Economy Papers, No. 376, OECD Publishing, Paris, https://doi.org/10.1787/d7b90d15-en.
[72] OECD (2025), “Sharing trustworthy AI models with privacy-enhancing technologies”, OECD Artificial Intelligence Papers, No. 38, OECD Publishing, Paris, https://doi.org/10.1787/a266160b-en.
[33] OECD (2025), “Total VC investments in AI by country and industry, data from Prequin”, OECD.AI Policy Observatory, (database), https://oecd.ai/en/data?selectedArea=investments-in-ai-and-data&selectedVisualization=total-vc-investments-in-ai-by-country-and-industry (accessed on 16 May 2025).
[20] OECD (2024), AI in Health: Huge Potential, Huge Risks, OECD, Paris, https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/01/ai-in-health-huge-potential-huge-risks_ff823a24/2f709270-en.pdf.
[11] OECD (2023), Health at a Glance 2023: OECD Indicators, OECD Publishing, Paris, https://doi.org/10.1787/7a7afb35-en.
[15] OECD (2023), OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market, OECD Publishing, Paris, https://doi.org/10.1787/08785bba-en.
[8] OECD (2023), Policies for the Future of Farming and Food in the European Union, OECD Agriculture and Food Policy Reviews, OECD Publishing, Paris, https://doi.org/10.1787/32810cf6-en.
[17] OECD (2022), “OECD Framework for the Classification of AI systems”, OECD Digital Economy Papers, No. 323, OECD Publishing, Paris, https://doi.org/10.1787/cb6d9eca-en.
[21] OECD (forthcoming), State of AI in Health across the OECD, OECD Publishing, Paris.
[14] OECD (forthcoming), When Hippocrates meets AI – Readying the Health Workforce for the Responsible Adoption of AI, OECD Publishing, Paris.
[1] OECD/European Commission (2024), Health at a Glance: Europe 2024: State of Health in the EU Cycle, OECD Publishing, Paris, https://doi.org/10.1787/b3704e14-en.
[34] Pinto-Coelho, L. (2023), “How artificial intelligence is shaping medical imaging technology: A survey of innovations and applications”, bioengineering, Vol. 10/12, p. 1435, https://doi.org/10.3390/bioengineering10121435.
[38] Radiobotics (2025), “AI in practice: AIFI pilot started in five Dutch hospitals”, https://radiobotics.com/press-release/ai-in-practice-aifi-pilot-started-in-five-dutch-hospitals/ (accessed on 15 January 2025).
[27] Radiobotics (2025), “RBfracture™”, https://radiobotics.com/solutions/rbfracture/ (accessed on 14 January 2025).
[43] Radiobotics (2024), “Transforming fracture detection accuracy at Kettering General Hospital”, https://radiobotics.com/case-study/transforming-fracture-detection-accuracy-at-kettering-general-hospital/ (accessed on 11 January 2025).
[57] Redondo, E. et al. (2023), “A simulation model for predicting hospital occupancy for COVID-19 using archetype analysis”, Healthcare analytics (New York) 100197, https://doi.org/10.1016/j.health.2023.100197.
[78] Sanofi (2024), “Artificial intelligence across Sanofi”, https://www.sanofi.com/en/our-science/digital/ai-across-sanofi (accessed on 11 January 2025).
[25] Sanofi (2024), “Press release: Sanofi, Formation Bio and OpenAI announce first-in-class AI collaboration”, https://www.sanofi.com/en/media-room/press-releases/2024/2024-05-21-05-30-00-2885244 (accessed on 13 January 2025).
[56] SAS (2020), “Italy’s second-largest hospital uses advanced analytics for effective pandemic response”, https://www.sas.com/en_us/customers/gemelli-university-hospital.html (accessed on 11 January 2025).
[52] Schöler, J. et al. (2023), “Impact of AI-aided colonoscopy in clinical practice: A prospective randomised controlled trial”, BMJ Open Gastroentrology, Vol. 11/1, p. e001247, https://doi.org/10.1136/bmjgast-2023-001247.
[42] Skin Analytics (2023), “DERM makes medical history as world’s first autonomous skin cancer detection system is approved for clinical decisions in Europe”, https://skin-analytics.com/news/regulatory-certification/derm-class-iii-ce-mark/#:~:text=This%20approval%20comes%20as%20healthcare,month%20just%20to%20be%20seen (accessed on 12 January 2025).
[68] Smith, J. (2022), “Sanofi joins AI gold rush in €4.6B drug discovery pact with Exscientia”, Labiotech, https://www.labiotech.eu/trends-news/exscientia-sanofi-ai/#:~:text=the%20UK%20AI%20player%20Exscientia,develop%20oncology%20and%20immunology%20treatments (accessed on 10 January 2025).
[18] Stoumpos, A., F. Kitsios and M. Talias (2023), “Digital transformation in healthcare: Technology acceptance and its applications”, International Journal of Environmental Research and Public Health, Vol. 20/4, p. 3407, https://doi.org/10.3390/ijerph20043407.
[49] Sung Lee, M. et al. (2025), “Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: The ROMIAE multicentre study”, European Heart Journal, Vol. 46/20, pp. 1917-1929, https://doi.org/10.1093/eurheartj/ehaf004.
[51] Tham, S. et al. (2023), “The imitation game: A review of the use of artificial intelligence in colonoscopy, and endoscopists’ perceptions thereof”, Annals of Colonprotcology, Vol. 11/1, p. e001247, https://doi.org/10.3393/ac.2022.00878.0125.
[73] Turbine (2025), “Expand my pipeline”, https://turbine.ai/target-selection/ (accessed on 12 January 2025).
[29] Turbine (2024), “Turbine achieves key milestone in collaboration with Ono Pharmaceutical”, https://turbine.ai/news/turbine-achieves-key-milestone-in-collaboration-with-ono-pharmaceutical/ (accessed on 10 January 2025).
[36] van Leeuwen, K. et al. (2023), “Clinical use of artificial intelligence products for radiology in the Netherlands between 2020 and 2022”, European Radiology, Vol. 34, pp. 348-354, https://doi.org/10.1007/s00330-023-09991-5.
[35] Zanardo, M. et al. (2024), “Impact of AI on radiology: A EuroAIM/EuSoMII 2024 survey among members of the European Society of Radiology”, Insights into Imaging, Vol. 15/250, https://doi.org/10.1186/s13244-024-01801-w.
[62] Zaribafzadeh, H. et al. (2023), “Development, deployment, and implementation of a machine learning surgical case length prediction model and prospective evaluation”, Annals of Surgery, Vol. 278/6, pp. 890-895, https://doi.org/10.1097/SLA.0000000000005936.
[74] Zhang, B. et al. (2023), “Harnessing artificial intelligence to improve clinical trial design”, Communications Medicine, Vol. 3/191, https://doi.org/10.1038/s43856-023-00425-3.