Infrastructure for innovation: getting the NHS and social care ready for AI

Basic infrastructure challenges are holding back the potential for AI to improve health and care services. These challenges erode staff enthusiasm for AI and contribute to a demoralised workforce. Those staff who are positive about AI recognise its possibilities but in their day-to-day work they often find that poor infrastructure (such as computers crashing and unstable internet connections) adds to their workload, and that the technology is outdated and unable to support AI.  

AI tools promise huge improvements to people’s care, as well as more efficient use of finite resources. But if that huge potential is to be realised, basic technological infrastructure needs to be in place. Unfortunately, this is often not the case.   

Here we try to answer the question: what infrastructure needs to be in place to ensure that AI can be used in the health and care sectors? Informed by conversations across health and social care with staff, suppliers, researchers and patients, we consider what’s needed for not only the technical infrastructure, but also the wider environment and system capabilities.  

Technical infrastructure 

Access to suitable devices  

For the health system to be able to use AI, devices – everything from computers, tablets and smartphones to more sophisticated medical equipment – must be suitable and available. We heard from one interviewee: 

However, the quote above isn’t necessarily true for everywhere. Other providers have better digital capability are developing and using AI.  

Different parts of the NHS have varying levels of old or unsuitable technology systems; it’s estimated that between 10–50% of systems need to be modernised. Unfortunately, this doesn’t just relate to old devices but also newer devices that were purchased without considering their future functionality. Although hospitals have programmes to replace and improve devices, this still takes time.  

Old devices are unsuitable because they are unable to support AI. If the wider IT infrastructure isn’t considered then even if devices are updated, they will quickly become obsolete and unusable, and given the pace of AI without forward planning the devices will become outdated more quickly, wasting money and resources. Having infrastructure that is capable of supporting AI means both having suitable devices and having a sufficient number of devices. Unfortunately, staff often have to share IT equipment and if they are unable to access a device, they can’t make use of AI capabilities.  

Ensuring adequate connectivity 

For AI to work, connectivity is really important. A fast, high capacity and stable internet connection is essential in all workplaces, but is particularly important for AI tools in health and care. AI can work on devices (desktops), but some AI only works in the cloud. For example, clinicians can send medical images to an AI algorithm in the cloud to create a 3D simulation, which they are then able to use to better plan surgery, but getting those images to the cloud requires good connectivity. Currently, it is not clear which way AI usage is going to develop – whether it will sit on devices (desktops) or in protected cloud data centres, or a combination of both. Providers therefore need to plan for both types of AI capability, or there needs to be specific decisions at a national level as to which route AI use will take. 

Unfortunately, in 2023 current provider connectivity was determined to be inadequate and is still insufficient for current daily demands.  

Many providers have invested in wi-fi to meet present-day requirements, but this is unlikely to be sufficient to meet the increasing demand of AI applications.  

Developing a cloud strategy 

Despite the 2013 government Cloud First Strategy, which mandated that government departments and public organisations should use cloud services as their first option, the NHS continues to use a combination of both on-premises data centres and cloud-based servers for data storage and processing. Cloud is believed to be cheaper but there are different risks and benefits, such as how much budgetary control providers have and how easy it is to increase or decrease digital storage. Digital leaders have mixed opinions on the use of cloud systems based on factors including information governance risk and budgetary control. 

On-premises data centres could hold back AI development, for example if they lack physical space for more processors and cooling or have insufficient electrical supply. Cloud data centres offer more flexibility to scale up or down without having to manage the physical infrastructure, but a wholesale shift to cloud-first would need sustained funding and digital leaders who can advocate for cloud as the best option.  

However, many NHS providers leading AI development consider cloud computing to be expensive, particularly for computer-vision forms of AI that use large imaging datasets. As a result, these providers are not investing in cloud but installing more on-premises server capability. Providers are also building partnerships with academia to access more powerful computing capability, but this approach comes with its own challenge of increased complexity to navigate data-sharing agreements.  

The health and care system needs to have a comprehensive cloud and on-premises computer strategy that leverages the best of both. This should support and build on the ​​AI opportunities action plan.  

Reducing permissions constraints to allow for innovation 

The use of virtual machines (computers with limited capabilities that run in the cloud) and stringent IT permissions are substantially limiting the capabilities of devices as well as clinicians’ ability to innovate with AI tools.  

Similarly, GPs are limited in their ability to use their computers to test and innovate, which prevents experimentation and testing of AI tools.  

It’s not just the NHS but social care, too.  

Striking the right balance between centralised and individual control of technology is difficult. Limiting the functionality of devices can help improve security and avoid overloading devices’ capabilities. However, this significantly restricts the potential utility of a device, preventing the use of AI and suppressing innovation.  

Data, governance and AI monitoring 

Consistent and integrated data information systems 

Data information systems, for example electronic health records, digital care records, case management systems, radiology information systems, and picture archiving and communication systems, are essential infrastructure for AI development and implementation. These systems store essential medical information, including the data needed to train AI models, data for analysis by AI and the outputs generated by AI models. However, to do this the data needs to be easily accessible.  

A common frustration among many health and social care providers, as well as AI suppliers, is the struggle to access this data. In the experience of staff leading AI development there is a lack of conformity when it comes to standards, or the data information systems are not designed for data access. There is a perception that data information system suppliers are uninterested in helping innovators to access data and plug in new technologies. Not all suppliers behave the same way, but many fall into three main groups: those that don’t allow data access; those that allow data access but no support; and those that charge to enable data access. This is perceived to be for various reasons, including lack of resources, avoiding additional costs, protecting business interests, preventing competition and to generate income. This means there is a hidden cost that involves paying for support or investing in technical staff just to access data before AI development and implementation are even considered.  

This is a problem across both health and social care.  

Circumventing these challenges involves additional cost to NHS and social care providers. One option is to use a supplier who can offer data integration with electronic patient records. Another is for providers and integrated care systems to invest in the technical staff and expertise to engineer the requisite access to the data. A third option is to establish a renewed agreement with suppliers. Without the appropriate resources or leadership, providers are unable to overcome barriers to data access, hindering the development and use of AI tools.  

As the system continues to digitalise and legacy systems are modernised, data accessibility needs to be considered or there is a risk that data will not be able to be used for AI purposes. Even recent initiatives such as imaging networks face challenges.  

Data information systems are necessary to digitalise the NHS. There’s an incorrect assumption that digitalisation automatically enables more accessible and usable information. There can be significant costs incurred to the system or health and care providers remain unable to access the data for AI purposes.  

Effective AI development and use need federated access to data, the existence of a shared care record, access to national data infrastructure (for example, GP Connect), approval to share data across providers, and approval to use data for secondary purposes of service planning and research. Good data infrastructure creates protected and ethically approved environments to bring data together to develop and test AI models.  

Improving data quality 

Accessible data is essential for AI to be developed, run and monitored, but it’s not just about accessing the data – it also needs to be of sufficient quality and completeness. Real-world data in the NHS is often incomplete, mislabelled or unsuitable for a particular AI tool. This limits the NHS’s ability to develop, evaluate and use AI tools.  

Addressing the quality of data means having systems, tools and processes that allow datasets to be cleaned, labelled, shared and ​stored​. This starts with overcoming the challenges to accessing data within data information systems (see above) then developing the workforce, skills and tools that can clean, label and assess the quality of the data. A feedback and quality improvement loop is also needed to identify opportunities to improve data quality by supporting staff training and changing processes, where necessary.  

Making information governance work for AI 

Information governance is often viewed as a blocker to change but it’s important for ensuring that data is used appropriately and privacy protections are in place. Good information governance balances data sharing to enable better care with protecting privacy and maintaining trust. For example, clinicians can view hospital and GP records for the purposes of care provision, but if they wish to use that data for testing, developing or implementing AI, there needs to be a legal basis and different information governance agreements in place.  

The NHS, local authorities and social care providers have often put information governance in place on a case-by-case basis for each locality, creating increasing complexity. NHS and social care organisations trying to navigate information governance requirements are often frustrated by this lack of clarity, which then creates complex negotiations at a local level, as well as significant national variation.  

The variation in local interpretation of national guidance can make it more challenging to combine datasets, slowing the development and use of AI. Where we see progress is when staff are supported by local leaders. This helps staff to navigate information governance challenges and often involves conversations with communities and people to continue to build trust in data use.  

Post-deployment monitoring 

Where AI is potentially very different from previous software capabilities is that it needs to be constantly monitored for changes in performance – also known as drift. This is when the AI response changes over time and could result in incorrect care decisions. Drift could occur due to changes in the input data (eg, if a scanner is upgraded or replaced the quality of outputs may change), processes (eg, variation in tissue sample preparation), patients/environment (eg, demographic shifts) or technology (eg, the AI model changes). For example, changes in the medical imaging equipment may result in a change in the image quality, and hence the AI response.  

Drift means an AI response is no longer acceptably accurate, which could result in harm, so there need for processes for alerting for drift and rectifying it.  

As mentioned in the section above on data information systems, these systems have led to siloed data, which prevents a complete picture of the patient. This can make monitoring for drift across pathways a challenge. As AI is still quite unfamiliar, there needs to be concerted national and local approaches to ensure appropriate monitoring for drift is in place. 

Monitoring AI when it’s in use should include automated quality checks and establishing the levels at which the AI is no longer performing well. This monitoring will have cost implications to the NHS to create the capacity and workforce to ensure robust monitoring. 

Leaders should be aware of the need for post-deployment monitoring and work with technical specialist staff and suppliers to develop appropriate monitoring technologies, tools and processes. Leaders should also ensure that procurement processes include requirements for post-deployment monitoring to comply with regulation and clinical safety requirements. 

​​Organisational and system capabilities

Supporting IT staff and teams  

Historically, NHS hospitals had IT support staff on site linked to each department. This meant that staff understood the IT systems and the needs of the department, but budget cuts and centralised IT have changed the configuration of these IT teams. There also tends to be a high turnover of IT staff, which erodes institutional memory and knowledge, and results in limited awareness of the processes and systems. The knock-on effect of all this is that it takes longer to do the necessary changes to enable systems to use AI.  

IT teams are essential to manage permissions and support the linking of IT systems so that new tools can be used. But too often their time is prioritised to immediate needs in order to keep services operational, and continued underinvestment means that these teams have little capacity to carry out the exploratory work needed to support and use AI systems. This means that innovators can find IT teams risk adverse and reluctant to change existing systems, as their focus is more on mitigating cyber and hacking risks.  

Securing and protecting funding  

More often than not technology budgets are raided for other uses, meaning that after buying the technology there is little left to fund the implementation. This then results in insufficient funding for staff recruitment, training and upskilling, changing processes and workflows, and patient engagement. Without the funding for implementation, the technology is unlikely to have the anticipated benefits and staff are unlikely to be ready to use the AI. Some suppliers and providers are using research grants to update legacy infrastructure in order to get the provider into a more AI-ready state.  

Technology funding announcements often give insufficient notice and have short application periods. Providers then respond hurriedly, making it more likely that there will be poorly co-ordinated decisions about how to allocate funding. Leaders should ensure that digital strategies take a proactive and collaborative approach to AI with regard to strategic aims and patient needs. When funding opportunities arise, a strategically aligned approach can then be taken, continuing to build progress towards organisational and system aims. 

Assessing and evaluating AI opportunities  

Providers can find themselves overwhelmed by AI suppliers and need to manage the large number of suppliers wanting to work with them. In response, providers are creating methods to sift and assess AI tools.  

Some providers have developed internal decision-making guidance, outlining the necessary steps and considerations.  

This internal guidance takes into consideration the provider’s resources and what is needed to evaluate the AI. The process of evaluating an AI product starts with ascertaining whether a tool is a solution for the provider’s most pressing problems and the improvements they seek to make. A multi-disciplinary group that includes patients and clinical leaders establishes the utility of a tool, ensuring that it meets a need, solves a problem or provides a clear improvement. This process ensures that tools align with organisational and system priorities.  

Next, there needs to be a process to check the infrastructure requirements, data use and flow, compliance with data protection and information governance, how much testing has already happened, how well the AI has performed to date, and how well the tool integrates into existing systems. The provider also needs to consider evaluation and monitoring approaches. This information should be shared with procurement staff to support informed procurement approaches.  

After this comes the assessment of a tool using historic data specifically from the provider, this helps staff to understand if the tool would work for their particular patient demographics and processes. Using retrospective data, the AI results can be compared and reviewed with historic clinical staff results to understand any differences. Some tools can have very specific tolerances. For example, one provider evaluated an AI tool used for medical-imaging analysis and found that it wouldn’t work for approximately a quarter of their scans because the tool was too specific about how the patient should be orientated within the scanner. The tool would therefore require significant investment in time and training to reduce the number of rejected images, but even this would not guarantee that the number of rejected images would be low enough to reduce workload, rather than increase it due to requiring repeated scans.  

Last, is the testing of the AI in the actual care pathways with staff. This phase presents the opportunity to assess the technology and develop the workflows and processes with the clinical teams to ensure the updated pathway is evaluated against key measures. 

Summary 

For AI tools to benefit patients, people who draw on services, staff and health and care providers it’s essential that the correct infrastructure be put in place. This includes technical infrastructure but also facilitating teams, permissions, governance and leadership. The journey the NHS and the social care sector needs to go on is a complex development of operational infrastructure, workforce shift and culture shift. Digital transformation needs the right technology to be available but also depends upon changes to the environment including infrastructure, skills, roles, workflows and processes and leadership. 

AI progress is racing ahead but the lack of investment and focus on infrastructure is an anchor on the ability of the health and care system to use AI to benefit patients and staff. Everyday struggles of crashing computers and poor connectivity suck away at the belief that technology can help to solve problems. AI recasts what infrastructure is needed for these tools to be effective but what is still important is infrastructure remains an essential necessity.