Summary

Dementia poses an increasing global health challenge, and the introduction of new drugs with diverse activity profiles underscores the need for the rapid development and deployment of tailored predictive models. Machine learning has shown promise in dementia research, but it remains largely untested in routine dementia health care—particularly for image-based decision support—owing to data unavailability. Thus, data drift remains a key barrier for equitable real-world translation. We propose and pilot a scalable, cloud-based infrastructure as code solution incorporating privacy-preserving federated learning. This architecture preserves patient privacy by keeping data localised and secure, while enabling the development of robust, adaptable artificial intelligence models. Although technology giants have successfully implemented such technologies in consumer applications, their potential in health-care applications remains largely underutilised. This Viewpoint outlines the key challenges and solutions in implementing cloud-based federated learning for dementia medicine and provides a well-documented codebase to support further research.

Introduction

Dementia—also referred to as major neurocognitive disorder—is a growing global health challenge driven by ageing populations. Dementia is a heterogeneous syndrome with multiple causes, including Alzheimer’s disease (which is the most common), vascular dementia, and Parkinson’s disease-associated dementia.1 The global economic burden of dementia is projected to reach $16·9 trillion by 2050, disproportionately affecting low-income and middle-income countries (LMICs).2 A key challenge in dementia care—particularly in memory clinics, which are specialist referral centres for evaluating cognitive impairment—is differential diagnosis owing to the heterogeneity of symptoms. Clinical decision support tools powered by artificial intelligence (AI) offer potential to improve diagnostic accuracy. In this Viewpoint, we focus on the role of cloud-based federated learning in enabling such tools, specifically for distinguishing Alzheimer’s disease from vascular and other atypical dementias.

Search strategy and selection criteria

We identified the background literature for this Viewpoint through searches of Google Scholar, PubMed, and arXiv conducted between February, 2023, and May, 2025, using the search terms “federated learning”, “cloud computing”, “dementia”, “infrastructure as code”, “artificial intelligence in healthcare”, and “data privacy”. No date restrictions were applied to the search results, and searches were limited to work written in English. Technical documentation and implementation guides from major cloud providers (Amazon Web Services, Microsoft Azure, and Google Cloud Platform) and federated learning frameworks were also reviewed and cited as appropriate. Additional relevant references were identified through citation tracking of key papers and via ResearchRabbit graphs. The final reference list was curated based on relevance to the technical and clinical aspects of cloud-based federated learning discussed in this Viewpoint.

The 2024 international guidelines for early diagnosis of dementia emphasise the use of biomarkers (eg, amyloid and tau) and neuropsychological profiling.3,4 Neuroimaging provides supportive diagnostic information and helps to exclude alternative causes in complex cases. However, neuroimaging is not required for dementia diagnosis; compared with MRI, CT is more commonly used. Although the standalone diagnostic utility of MRI remains questionable,5,6 multimodal AI approaches that integrate imaging with other data types remain promising for research applications across medical fields.7 Nonetheless, the implementation of AI in dementia care faces considerable barriers, including challenges in validation and clinical workflow integration, and concerns related to privacy, security, and regulatory compliance. These barriers restrict researchers’ access to data generated in well-resourced research settings. Moreover, such datasets have historically lacked representativeness of real-world patient populations and data quality. Recognising this gap, major initiatives such as the Alzheimer’s Disease Neuroimaging Initiative have initiated new recruitment efforts aimed at improving diversity and real-world representativeness.8 Furthermore, training AI models (particularly image-based ones) requires high-performance computing (HPC) resources that are often financially and logistically inaccessible to most hospitals, thereby limiting their implementation in clinical settings.

Standardised research-scanning protocols, such as the UK Biobank protocol in the Oxford Brain Health Clinic,9 are improving the quality and consistency of clinical data relevant to dementia. However, AI models should be designed to adapt to the variability inherent in real-world health-care environments. One key limitation is data drift—a statistical mismatch between training data and real-world inference data. In health care, data drift arises from multiple sources, including demographic bias (some population groups are under-represented in training data) and research participation bias (participants enrolled in research studies differ systematically from the general population). Overcoming data drift is important in dementia care as new drugs become available for Alzheimer’s disease;10,11 nonetheless, heterogeneity in treatment response and the unpredictability of side-effects pose clinical challenges. Therefore, data drift could result in (1) misdiagnosis*,* increasing the risk of side-effects, including amyloid-related imaging abnormalities, and (2) delayed diagnosis, leading to increased burden on patients, families, and health-care systems.

To address these challenges, we propose performing research and developing algorithms using routinely collected dementia health-care data through cloud-first federated learning. This approach addresses the highlighted disparities and challenges by keeping real-world data local through federated learning and leveraging cloud infrastructure to provide global computational power and streamline model training and deployment. Such a framework can enable the development of equitable, personalised machine learning models for dementia medicine. We advocate for cloud-first federated learning over traditional data centralisation and on-premises federated learning implementations, which require every participating institution to own and maintain physical computing infrastructure.

Equitable data-driven dementia medicine: the case for cloud computing

The flexibility, scalability, and security offered by cloud-first federated learning make it a compelling approach for advancing equitable, personalised medicine in dementia care. The alternative approaches are data centralisation and traditional on-premises federated learning (using large or small computing resources in the hospital). Federated learning is a distributed machine learning approach in which models are trained collaboratively across multiple sites or nodes. Advances in computer science and cybersecurity,12–14 including privacy-preserving architectures in medical imaging as reviewed by Kaissis and colleagues,13 have made federated learning more accessible. For dementia-relevant models trained on high-dimensional neuroimaging data, every federated learning node requires HPC resources. In traditional federated learning setups, fulfilling this HPC requirement typically involves physical on-premises machines. Although such setups can address many of the privacy, governance, and legal constraints associated with data centralisation approaches, not all hospitals can support on-premises resources due to financial, logistical, or other constraints.

Secure data lakes (centralised repositories to store large amounts of data) such as Data Safe Havens or Trusted Research Environments, regardless of whether they are co-located with hospital data, are often inadequate to meet the computational demands of image-based dementia research. Centralised off-site data lakes require transferring data from hospitals or clinics, invoking considerable regulatory complexities, especially when data from multiple hospitals are involved. On-site data lakes require the purchase and maintenance of on-premises computing hardware, which introduces additional challenges, as most hospitals rarely have the required, available capital or HPC-trained personnel to support such infrastructure. Coordinating an on-premises effort across multiple hospitals further complicates implementation and might disproportionately benefit affluent regions. However, adopting privacy-preserving cloud-first federated learning addresses these concerns by allowing data to remain on locally managed hospital networks while leveraging the scalability and flexibility of global computational resources.

The shift towards cloud computing is further enhanced by the adoption of infrastructure as code, which allows computing resources, networks, and security protocols to be defined and managed through computer code rather than manual configuration. In traditional systems, configurations are applied manually—an approach that is both time-consuming and error-prone. In contrast, infrastructure as code enables these configurations to be defined once and reused or adapted as needed, ensuring a consistent, reliable, and reproducible setup. Infrastructure as code provides a theoretical framework with the potential to democratise access to dementia care. Although practical evaluation of this potential is needed, the approach conceptually addresses several barriers: (1) potential cost-effectiveness (cloud resources can be provisioned on-demand); (2) standard deployment of trained models into clinical practice through consistent and replicable infrastructure; and (3) reduced barriers for expansion into LMICs by eliminating the need for hardware purchase and maintenance.

Federated learning broadly improves access by allowing model training on distributed data. Our cloud-first federated learning approach provides additional, practical equity advantages, particularly for under-resourced or geographically isolated health-care settings. By removing the need for local high-performance infrastructure and using infrastructure as code to rapidly deploy secure, preconfigured environments, cloud-first federated learning enables meaningful participation from sites that would otherwise be excluded. This broader participation improves both the demographic diversity of training data and extends the geographical applicability of resulting models, enhancing the fairness and real-world generalisability of AI-supported dementia care. A 2025 study15 showed improved fairness when incorporating real-world data into diagnostic AI applications via targeted fine-tuning of pre-trained models on data from under-represented groups.

Cloud-first federated learning presents some challenges. Hospitals need local information technology expertise to manage client software and navigate firewall restrictions. Additionally, depending on the machine type and usage, cloud costs can occasionally exceed those of locally maintained systems. Nevertheless, accelerated by the COVID-19 pandemic, many hospitals have already adopted cloud services to complement or replace on-premises servers,16 positioning them to be cloud-first federated learning ready. A thorough cost–benefit analysis is essential to ensure the long-term value of cloud-first federated learning for each institution. Although our cloud deployment leverages a centrally coordinated model, future implementations could explore distributed federated learning models such as peer-to-peer. These decentralised approaches might offer advantages in fault tolerance and local autonomy but typically involve more complex coordination, security management, and version control—challenges that are currently better managed in centralised cloud-first federated learning systems.

Ultimately, cloud-first federated learning holds considerable potential to improve equity in dementia health care by enabling the deployment of trained models in any setting with internet access. Rieke and colleagues17 reported that even small, remote clinics can contribute to and benefit from cloud-first federated learning, as deployment of an already trained model requires minimal computing resources. This capability of small clinics presents promising opportunities to enhance data-driven clinical decision support regardless of geographical access to specialist facilities, particularly in the diagnosis and management of rare or atypical dementias (eg, young-onset dementia and prion disease).

From promise to practice: barriers and solutions to cloud-based dementia medicine

Deploying cloud-first federated learning in health care involves both technical and logistical barriers. Our project—Piloting A Secure, Scalable Infrastructure for AI in the NHS (PASSIAN)—was a focused, time-boxed, 6-month sprint project that aimed to develop and deploy a cloud-first federated learning prototype for dementia medicine. PASSIAN provided valuable insights into overcoming the multifaceted technical and logistical aspects of deploying cloud-first federated learning within complex health-care systems such as the UK’s National Health Service (NHS). Figure 1 outlines our infrastructure as code deployment for cloud-first federated learning, for which we have provided open-source scripts and documentation. We used two real-world memory clinic datasets: 670 cases from Essex Memory Clinic and 400 cases from Cambridge University Hospital. These datasets provided valuable insights into the clinical deployment of cloud-first federated learning infrastructure.

Figure 1 Diagrammatic representation of our infrastructure as code for deploying cloud-first federated learning

The figure illustrates the implementation of the cloud federated learning system, which integrates the federated learning platform FBM with our machine learning models and pipelines on AWS using infrastructure as code. Many of the icons represent specific AWS services. This setup involves three elements: a researcher hosted zone and two local node hosted zones—one for each hospital involved in the project. Each element is deployed in a secure VPC, and communication between them occurs through VPC peering; AWS provides a secure private connection between services in the different VPCs. Model development and deployment occurs on the researcher side through Jupyter (an interactive web-based computing platform) and Tensorboard (a visualisation toolkit for machine learning experiments) services, whereas FBM communication and model parameter sharing occurs through Restful and MQTT Fargate services. On the hospital side, clinicians can see, select, tag, and upload the data on which deep learning models can be trained through a convenient graphical user interface (FBM GUI) prompt, hosted as an AWS service. Each hospital’s VPC has dedicated data storage (EFS) and on-demand GPU-enabled compute resources for model training. The actual machine learning occurs within these GPU-enabled computing instances at each hospital node, with only model parameters (not patient data) being shared between sites through the central researcher zone. The icons are AWS service icons from Amazon Web Services, used in accordance with AWS’s brand guidelines for architectural diagrams in academic publications.

AWS=Amazon Web Services. EFS=elastic file system. FBM=Fed-BioMed. GPU=graphics processing unit. GUI=graphical user interface. MQTT=message queuing telemetry transport. VPC=virtual private cloud. VPN=virtual private network.

PASSIAN was designed as a proof-of-concept feasibility study, prioritising infrastructure deployment over clinical algorithm development. Therefore, questions of algorithm ownership were not a primary focus of our evaluation. During the pilot, basic machine learning models were collaboratively developed across the two hospital sites for research purposes, with model parameters accessible to all participating institutions through the federated framework. Nevertheless, for clinical deployments, establishing a clear framework for algorithm ownership that accounts for the roles of participating institutions, researchers, and technology providers is essential, particularly given the collaborative nature of federated model training. Related to this, in PASSIAN, the clinical responsibility remained with participating clinicians at each site, with the infrastructure serving solely as a research tool to demonstrate technical feasibility rather than informing patient care decisions. This approach aligns with current clinical practice in which diagnostic responsibility resides with the treating medical personnel, who follow established protocols and use AI tools as decision support rather than as replacements for clinical judgement. These governance considerations, encompassing both algorithm ownership and clinical accountability, will be essential for translating cloud-first federated learning infrastructure (and AI applications more broadly) from proof-of-concept research into clinical practice.

Data drift remains a considerable additional challenge in deploying AI models in dementia care. Although cloud-first federated learning can help to mitigate data drift by allowing models to train on more diverse datasets, it is not a complete solution. As imaging protocols evolve, some degree of data drift is likely to persist. Domain adaptation techniques offer a promising approach to address data drift. A 2023 study suggests that robust modelling and training practices can minimise some biases caused by data drift in medical imaging; however, these improvements are often limited to specific imaging contexts.18 Our proposed cloud-first federated learning framework aims to improve model robustness by training on a broader spectrum of data variations.

Data processing is an essential component of cloud-based dementia medicine. Neuroimaging can support clinical decision making, particularly in distinguishing dementia subtypes, by leveraging both expert visual evaluations and advanced quantitative analyses.5 Relevant machine learning approaches—whether based on derived features or deep learning6—require HPC resources owing to the high dimensionality of neuroimaging data. Cloud-first federated learning provides the necessary flexibility to access the required HPC resources, while infrastructure as code enables automation of preprocessing steps. These steps typically include transferring scans from hospital systems (eg, in Picture Archiving and Communication Systems), converting them into machine learning-compatible formats, reorganising them into standardised structures (eg, brain imaging data structure), and performing image defacing for anonymisation. Removing scanner artifacts, correcting noise, and aligning each scan to a standard brain template can also be automated within an infrastructure as code framework. Automation reduces the need for on-site intervention at every stage, enabling faster and more efficient workflows.

Our PASSIAN cloud-first federated learning system, built on the open-source Fed-BioMed platform,19 integrates a bespoke HPC environment for preprocessing on Amazon’s Elastic Compute Cloud (EC2; figure 1). Importantly, cloud-based HPC services support composable environments, offering control and flexibility to handle different image preprocessing requirements across clinical sites. Moreover, additional services, such as custom containers or software as a service, address bottlenecks in managing large neuroimaging datasets, thereby providing substantial cost-saving advantages, especially in LMICs.

To ensure data privacy, the PASSIAN cloud-first federated learning infrastructure as code leverages virtual private network (VPN) services from the cloud provider. Amazon Web Services’ (AWS) virtual private cloud (VPC) was used to keep our entire cloud-first federated learning network secure and isolated from the open internet. The use of VPC ensured that data remained within secure hospital-managed networks or VPCs, with access strictly controlled and limited to responsible clinicians and designated machine learning engineers. Researcher or hospital nodes connect to the server via Transport Layer Security-encrypted VPN or secure hypertext-transfer-protocol.

We selected AWS owing to its compliance with NHS security standards and existing integration with NHS infrastructure—including the possibility of keeping cloud resources isolated from the open internet. Although major cloud providers implement strong isolation protocols (called sandboxing), shared infrastructure can theoretically pose risks, such as memory access vulnerabilities, if sandboxing mechanisms fail. However, such risks are minimum and generally acceptable, particularly given that cloud providers typically possess superior security expertise and resources compared with fragmented on-premises hospital systems. Additionally, federated learning inherently mitigates model inversion attacks by keeping data local, and these risks are further reduced by applying local batch normalisation updates or incorporating differential privacy.20,21

Several NHS hospitals already pay for cloud computing services; therefore, we expect cloud-first federated learning to pose considerably lower logistical and governance barriers than data centralisation. Importantly, a cloud-based federated learning system is aligned with the UK Government’s cloud-first policy (introduced in 2013),22 further facilitating potential implementation within the NHS framework. Our system complies with NHS standards; however, broader implementation requires careful consideration of global regulatory frameworks such as health insurance portability accountability act in the USA,23 general data protection regulation in the EU,24,25 and other country-specific health data regulations. Comprehensive governance frameworks addressing data access, security, and accountability will be essential for successful production deployments. The sharing of model weights across institutions, such as those from generative models, warrants careful consideration, particularly in cases in which data licensing agreements are not formalised. To mitigate these risks, additional data-sharing agreements and safeguards should be established to support the global scalability of cloud-based federated learning solutions in health care. Although our implementation used AWS, our infrastructure was designed to be adaptable across cloud platforms. The core principles and architecture outlined in this paper are vendor agnostic, and our open-source codebase can be modified for compatibility with other providers such as Microsoft Azure and Google Cloud Platform.

Best practices for cloud-based dementia research

A systematic approach integrating technical, workflow, and administrative aspects is key to effective implementation of cloud-first federated learning solutions. Based on our experiences in the PASSIAN project, we present a schematic (figure 2) outlining our recommended best practices for establishing a dementia-focused cloud-first federated learning initiative.

Figure 2 Roadmap of recommendations for implementing cloud-based federated learning in dementia research

This figure illustrates a six-stage process beginning with groundwork and planning 6–12 months before the commencement of a project, and culminating in the execution of federated learning experiments after setup completion. Each stage details key activities and stakeholder engagement necessary to facilitate the seamless integration of cloud-based federated learning within health-care settings. Figure created with BioRender.com.

IT=information technology.

The proposed cloud-first federated learning framework is inherently multimodal and can integrate neuropsychological assessments, biomarkers, and clinical features into more comprehensive predictive models, thus aligning with diagnostic trends in dementia care.3,4 The framework also holds promise for monitoring treatment response and detecting early signs of treatment-related adverse events (eg, amyloid-related imaging abnormalities) associated with anti-amyloid therapies. Thus, AI-enhanced neuroimaging analysis could considerably improve patient safety.

Finally, although our pilot focused on the UK’s NHS, the potential of cloud-first federated learning extends to LMICs. However, adapting this system to LMICs requires addressing unique challenges, including poor internet connectivity, fewer available MRI machines, and diverse regulatory environments. Future efforts should focus on optimising cloud-first federated learning for dementia research across diverse settings, including exploring low-bandwidth adaptations, enhancing compatibility with portable and low-field MRI scanners,26,27 and ensuring compliance with local regulatory requirements. Overcoming these challenges will make cloud-first federated learning an even more valuable tool for advancing dementia care globally.

Limitations and future directions

Our proof-of-concept pilot focused on demonstrating the technical feasibility of deploying a cloud-based federated learning system within NHS memory clinics, while leaving several important aspects still unaddressed. First, although our infrastructure supports secure and efficient data processing and model training, we did not evaluate clinical performance or diagnostic outcomes—an area that will be addressed in future prospective studies. Second, broader adoption will require further engagement with hospital information technology departments and institutional stakeholders, particularly to streamline VPN configurations and support long-term maintenance. Additional challenges include ensuring compatibility with diverse data processing pipelines, handling multimodal data, and quantifying the economic and environmental costs of cloud computing. Finally, robust governance frameworks will be essential to clarify accountability, data access rights, and protection against adversarial risks.

Conclusion

Cloud-first federated learning—combining federated learning with infrastructure as code—represents a practical step towards translating research into equitable clinical dementia care. We have released well-documented open-source code for this infrastructure, along with best practice guidelines for implementation. As dementia diagnostics continue to evolve towards multimodal approaches, the flexibility of our cloud-first federated learning framework positions it as a valuable research platform that can adapt to changing clinical and scientific priorities. We hope this work contributes meaningfully to bridging the gap between research and health care, paving the way for more equitable, data-driven dementia medicine.

Contributors

MM and BR contributed to the drafting of the original manuscript, including data processing, coding, discussions, and study design. MM performed revisions and edited the final manuscript version. MM and TD created the figures. NPO designed and led the study, provided overall guidance and study funding, led discussions, co-wrote both the original draft and the final version of the manuscript, and contributed to coding. TD performed most of the coding for our cloud-based federated learning software infrastructure, wrote the software documentation, and contributed to discussions and diagrams. TR, ZW, and DL contributed to collaboration and discussion, and participated in the review and editing of the manuscript. All authors approved the final version of the manuscript.

Declaration of interests

NPO is a paid consultant for Queen Square Analytics Limited (UK) on projects unrelated to the content of this paper. TR declares grant funding unrelated to this project from Alzheimer’s Society UK (AS-PhD-21-027) and the UK Research and Innovation Medical Research Council (G655128). TR is a Charitable Trustee of Peterhouse, University of Cambridge. DL declares grant funding unrelated to this project from Alzheimer’s Research UK (DEMON Network). All other authors declare no competing interests.

Acknowledgments

All authors acknowledge research funding from the UK Research and Innovation Medical Research Council (MR/X005674/1) received as a grant to their respective institutions (Principal Investigator: NPO). TR acknowledges research funding from Alzheimer’s Research UK (G127979) received as a grant to his institution. NPO is a UKRI Future Leaders Fellow (MR/S03546X/1, MR/X024288/1).

References

1.

Arvanitakis, Z ∙ Shah, RC ∙ Bennett, DA

Diagnosis and management of dementia: review

JAMA. 2019; 322:1589-1599

2.

Nandi, A ∙ Counts, N ∙ Chen, S ∙ et al.

Global and regional projections of the economic burden of Alzheimer’s disease and related dementias from 2019 to 2050: a value of statistical life approach

EClinicalMedicine. 2022; 51, 101580

3.

Dubois, B ∙ Villain, N ∙ Schneider, L ∙ et al.

Alzheimer disease as a clinical-biological construct—an international working group recommendation

JAMA Neurol. 2024; 81:1304-1311

4.

Frisoni, GB ∙ Festari, C ∙ Massa, F ∙ et al.

European intersocietal recommendations for the biomarker-based diagnosis of neurocognitive disorders

Lancet Neurol. 2024; 23:302-312

5.

Pemberton, HG ∙ Zaki, LAM ∙ Goodkin, O ∙ et al.

Technical and clinical validation of commercial automated volumetric MRI tools for dementia diagnosis—a systematic review

Neuroradiology. 2021; 63:1773-1789

6.

Ansart, M ∙ Epelbaum, S ∙ Bassignana, G ∙ et al.

Predicting the progression of mild cognitive impairment using machine learning: a systematic, quantitative and critical review

Med Image Anal. 2021; 67, 101848

7.

Chen, X ∙ Wang, X ∙ Zhang, K ∙ et al.

Recent advances and clinical applications of deep learning in medical image analysis

Med Image Anal. 2022; 79, 102444

8.

Weiner, MW ∙ Veitch, DP ∙ Miller, MJ ∙ et al.

Increasing participant diversity in AD research: plans for digital screening, blood testing, and a community-engaged approach in the Alzheimer’s disease neuroimaging initiative 4

Alzheimers Dement. 2023; 19:307-317

9.

Griffanti, L ∙ Gillis, G ∙ O’Donoghue, MC ∙ et al.

Adapting UK Biobank imaging for use in a routine memory clinic setting: the Oxford Brain Health Clinic

Neuroimage Clin. 2022; 36, 103273

10.

van Dyck, CH ∙ Swanson, CJ ∙ Aisen, P ∙ et al.

Lecanemab in early Alzheimer’s disease

N Engl J Med. 2023; 388:9-21

11.

Sims, JR ∙ Zimmer, JA ∙ Evans, CD ∙ et al.

Donanemab in early symptomatic Alzheimer disease: the TRAILBLAZER-ALZ 2 randomized clinical trial

JAMA. 2023; 330:512-527

12.

Karargyris, A ∙ Umeton, R ∙ Sheller, MJ ∙ et al.

Federated benchmarking of medical artificial intelligence with MedPerf

Nat Mach Intell. 2023; 5:799-810

13.

Kaissis, GA ∙ Makowski, MR ∙ Rückert, D ∙ et al.

Secure, privacy-preserving and federated machine learning in medical imaging

Nat Mach Intell. 2020; 2:305-311

14.

Tedeschini, BC ∙ Savazzi, S ∙ Stoklasa, R ∙ et al.

Decentralized federated learning for healthcare networks: a case study on tumor segmentation

IEEE Access. 2022; 10:8693-8708

15.

Wang, Y ∙ Bone, R ∙ Fleisher, W ∙ et al.

Using under-represented subgroup fine tuning to improve fairness for disease prediction, 2025

Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies. 2025;, 2025 240-253

16.

Cresswell, K ∙ Williams, R ∙ Sheikh, A

Using cloud technology in health care during the COVID-19 pandemic

Lancet Digit Health. 2021; 3:e4-e5

17.

Rieke, N ∙ Hancox, J ∙ Li, W ∙ et al.

The future of digital health with federated learning

NPJ Digit Med. 2020; 3:119

18.

Wang, R ∙ Chaudhari, P ∙ Davatzikos, C

Bias in machine learning models can be significantly mitigated by careful training: evidence from neuroimaging studies

Proc Natl Acad Sci U S A. 2023; 120, e2211613120

19.

Silva, S ∙ Altmann, A ∙ Gutman, B ∙ et al.

Fed-BioMed: a general open-source frontend framework for federated learning in healthcare

Albarqouni, S ∙ Bakas, S ∙ Kamnitsas, K ∙ et al. (Editors)

Domain adaptation and representation transfer, and distributed and collaborative learning

Springer International Publishing, Cham, 2020; 201-210

20.

Pati, S ∙ Kumar, S ∙ Varma, A ∙ et al.

Privacy preservation for federated learning in health care

Patterns (N Y). 2024; 5, 100974

21.

Hatamizadeh, A ∙ Yin, H ∙ Molchanov, P ∙ et al.

Do gradient inversion attacks make federated learning unsafe?

IEEE Trans Med Imaging. 2023; 42:2044-2056

22.

UK Government

Government cloud first policy

23.

O’Herrin, JK ∙ Fost, N ∙ Kudsk, KA

Health insurance portability accountability act (HIPAA) regulations: effect on medical record research

Ann Surg. 2004; 239:772-776

24.

Larrucea, X ∙ Moffie, M ∙ Asaf, S ∙ et al.

Towards a GDPR compliant way to secure European cross border healthcare industry 4.0

Comput Stand Interfaces. 2020; 69, 103408

25.

Amini, MM ∙ Jesus, M ∙ Sheikholeslami, DF ∙ et al.

Artificial intelligence ethics and challenges in healthcare applications: a comprehensive review in the context of the European GDPR mandate

Mach Learn Knowl Extr. 2023; 5:1023-1035

26.

Arnold, TC ∙ Freeman, CW ∙ Litt, B ∙ et al.

Low-field MRI: clinical promise and challenges

J Magn Reson Imaging. 2023; 57:25-44

27.

Chetcuti, K ∙ Chilingulo, C ∙ Goyal, MS ∙ et al.

Implementation of a low-field portable MRI scanner in a resource-constrained environment: our experience in Malawi

AJNR Am J Neuroradiol. 2022; 43:670-674