Dementia services face increasing demand as populations age and as new therapies with different activity profiles become available. Artificial intelligence can support differential diagnosis, but deployment is often held back by fragmented data, data drift between research and routine practice and the high computing requirements of image-based models. Privacy rules and hospital information technology constraints further reduce the feasibility of centralising data. A cloud-first federated learning model that keeps data local while coordinating training centrally offers a workable route to broaden access, increase equity between sites and prepare health services for multimodal diagnostic pathways.
Enabling Participation Beyond High-Resource Centres
Traditional approaches rely either on centralised data repositories or on-premises federated learning. Centralisation triggers complex governance when data must cross institutional or geographical boundaries. On-premises federated learning expects every site to procure and maintain high-performance computing (HPC) and specialised IT skills. For dementia diagnostics that use high-dimensional neuroimaging, these conditions often limit participation to well-resourced centres and produce models trained on narrow populations.
Must Read: AI, Cloud and Risk Readiness at a Crossroads
Cloud-first federated learning reduces these barriers by keeping patient data within hospital-managed environments and sharing only model parameters. Infrastructure as code makes deployment reproducible, defining networks, security controls and compute resources once and reusing them across sites. Hospitals can use on-demand cloud resources rather than building large local clusters. This supports cost management, streamlines model updates and enables inclusion of smaller or remote services. Broadening participation increases demographic and protocol diversity in training data, which is important where data drift over time can undermine model performance as imaging practices, populations and treatment options evolve. A local cost–benefit assessment remains necessary because some organisations may find certain cloud workloads more expensive than local compute.
Lessons from a Federated Pilot in Memory Clinics
A 6-month pilot, Piloting A Secure, Scalable Infrastructure for AI in the NHS (PASSIAN), focused on standing up the infrastructure rather than delivering a clinical model. Two real-world memory clinic datasets were used: 670 cases from Essex Memory Clinic and 400 cases from Cambridge University Hospital. Basic machine learning models were trained collaboratively for research purposes and model parameters were made available within the federated framework. Clinical responsibility remained with clinicians at each site, and the infrastructure did not inform patient care decisions. This showed that federated research can proceed while preserving local accountability.
The pilot used the open-source Fed-BioMed platform and a bespoke HPC environment on Amazon Elastic Compute Cloud, operating inside secure virtual private clouds (VPCs) with controlled connectivity, virtual private network services and encryption. Hospital nodes hosted storage, on-demand graphics processing capability and a graphical user interface so clinicians could view, select, tag and upload training data. Communication services managed parameter exchange while keeping patient data within hospital networks or their VPCs. The architecture was designed to be portable across major cloud providers and to fit existing health service security standards. It also highlighted operational needs, including client management by local IT teams, clear arrangements for algorithm ownership in future deployments and alignment with jurisdictional data licensing rules.
Automating Imaging Pipelines while Preserving Privacy
Image-based dementia support depends on consistent preprocessing. Using infrastructure as code, the pilot automated key steps: moving scans from Picture Archiving and Communication Systems into machine learning-compatible formats, reorganising them into standard structures such as the brain imaging data structure and applying defacing for anonymisation. Additional operations, including artifact removal, noise correction and alignment to standard templates, were packaged into composable environments. Automation reduced manual work at sites and increased reproducibility, which is essential when models must be retrained to address data drift.
Privacy was prioritised by keeping data local and isolating services in VPCs. Access was restricted to responsible clinicians and designated machine learning engineers. Although shared cloud infrastructure carries theoretical risks, the design aimed to benefit from the stronger isolation controls typically available from large cloud providers compared with fragmented on-premises estates. As dementia diagnostics move towards multimodal models that combine imaging, neuropsychological assessments, biomarkers and clinical features, the same cloud-first federated approach can orchestrate training across heterogeneous data sources while maintaining local stewardship. Extending deployment to lower income settings will require attention to bandwidth, scanner variability and regulatory diversity, but the approach allows compatibility extensions, including for portable and low-field MRI, without redesigning the entire infrastructure.
Cloud-first federated learning, supported by infrastructure as code, provides a practical route to make AI-enabled dementia decision support more equitable, secure and reproducible. By allowing hospitals to keep data local, orchestrating training centrally and automating image processing, it addresses core obstacles that have limited routine use of image-based models. Feasibility demonstrated in memory clinics indicates that health services can build a scalable research platform that adapts to changing diagnostic practice, provided that governance, interoperability, cost evaluation and strategies for handling data drift continue to evolve alongside the technology.
Source: The Lancet Digital Health
Image Credit: iStock
