Corporate report

UKHSA Advisory Board: Artificial Intelligence

Updated 9 May 2024

Date: 8 May 2024

Sponsor: Steven Riley

1. Purpose of the paper

The purpose of the paper is to update the advisory board of the ongoing activities supporting the adoption and utilisation of Artificial Intelligence (AI) in UKHSA and to seek input and guidance from Advisory Board members.

2. Recommendations

The Advisory Board is asked to:

• note the developments in our use of Artificial Intelligence and approach to organisational readiness for adoption and utilisation of Artificial Intelligence; and
• comment on areas where further development is required, our approach to enterprise level AI capabilities and our assessment of the risks posed by AI to health security.

3. Background

The Executive Committee established a cross-organisation Artificial Intelligence Taskforce to capture the benefits and associated risks of Generative AI for UKHSA.  This Taskforce reported its initial findings to the Advisory Board in November 2023 and provided a further update on use case development in January 2024.

4. Artificial Intelligence Steering Committee

As noted previously to the Advisory Board, we have established an AI Steering Committee to provide strategic oversight and guidance to our approach to adopting and utilising AI. This Committee, chaired by Steven Riley and comprising senior representation from across the agency, met for the first time on 25th March. Its purpose is to ensure our adoption (implementation of the technology) and utilisation (delivery and skills) of AI is aligned to UKHSA’s vision, mission, and objectives. It has a focus on both internal operational improvements (to boost productivity) as well as external interventions (to enhance the UK’s health security). The new Committee will work alongside the existing governance frameworks provided by the Digital, Data and Technology Assurance Board, Technical Review Board and Investment Board. It will use its expertise to support effective adoption of AI capabilities, management of risks, and transformation.

As part of the Committee’s remit, it will maintain oversight of the newly established register of AI activity which will capture key information on existing AI projects.  It also has accountability for the AI Assessment Framework which has been developed by the UKHSA Economics Team in collaboration with technical experts.  This framework assesses projects against 6 criteria (Annex 1).

5. Artificial Intelligence Health Security Threat Assessment

The National Health Analysis and Assessment Division in All Hazards Intelligence completed an assessment of the key risks to UK health security because of AI and the likelihood of these risks manifesting within the next 12 months.  This assessment consisted of a synthesis of factual information alongside assessments made by analysts that drew upon Subject Matter Experts, structured intelligence assessment techniques, existing literature and reporting at a range of classifications.

The assessment identified the key risk areas as:

• Bias and health inequalities
• Erosion of trust in UK government and health bodies
• Health mis/disinformation
• Malicious use of AI threat

Although this assessment focused on defining threats and is not a policy recommendation document, some potential mitigations identified to reduce the chances that these risks materialise include ensuring that AI driven tools used in the health sector are controlled under a regulatory framework, ensuring rigorous testing and high levels of transparency, closer collaboration between UKHSA and the Department for Science, Innovation and Technology (DSIT) Counter Disinformation Unit as well as the Home Office Disinformation analysis Team (to understand how threat actors may use health mis/disinformation and engaging more closely with national security partners to understand the threat posed by malicious actors).

The threat assessment will be updated on a regular cycle (~12 monthly or sooner if major developments trigger ad hoc requests for review) and the risks will be managed whilst we are developing and delivering internal AI capabilities.

6. Cross-Government Engagement

There is a significant amount of activity occurring across government related to AI discovery, safety, and applications, particularly in the realm of Generative AI. We have been proactive in aligning ourselves with our health partners in NHS and DHSC, as well as with other government departments such as Central Digital and Data Office (CDDO), Go Science, Department for Transport, and DSIT, to ensure our active involvement as AI continues to develop.

UKHSA have also responded to key cross government commissions on our approach to AI, including providing an indication of potential joint spending review bids for AI with DHSC. More broadly, UKHSA have also engaged with the Patients Association to discuss public engagement that may lead to panel work with patients in the future. We have also proactively engaged with the global conversation and developments regarding artificial intelligence use in public health, health security and health care provision.

UKHSA are adopting a ‘fast follower’ approach, positioning ourselves to learn from multiple areas of early adoption before we invest ourselves in platforms such as Microsoft Co-Pilot. This, coupled with ongoing research and discovery on more nuanced applications will help UKHSA invest wisely and keep research and development costs lower in the future.

7. Internal Engagement

We have continued our programme of internal engagement to raise awareness of the potential of AI and to encourage colleagues to explore these capabilities in line with our acceptable usage policy (AUP).  The leadership provided by our Data Protection Officer enabled the development of a UKHSA specific acceptable use policy while we wait for more detailed cross-government guidance from CDDO.

Building AI capability is on the roadmap for inclusion in UKHSA core essentials work, which will form the foundations for the next iteration of the employee induction and has executive sponsorship through the ‘Building a learning culture’ programme.

We aim to bring all UKHSA colleagues along as we adopt AI as an Agency, taking a supportive and positive approach, providing guidance to colleagues over the deployment and use of AI in whatever form, and ensuring everyone is clear about the organisation’s ‘guard rails’ of use so that we can be confident in how we maximise the opportunities AI brings across the Agency. Since the last update we have published a series of blogs and articles on Pulse (UKHSA intranet) relating to our approach to AI.

8. Existing Internal Use Cases - update

The development of our internal use cases continues at pace, with details of these being recorded in the newly established AI Register.

Since the last update to Advisory Board we have successfully deployed one of our AI projects to the UKHSA Tuberculosis Unit. This system enhances the manual review of country-of-origin documentation to identify those born in high-risk countries who are eligible for TB screening in the UK. The system is currently being trialled alongside standard practice to test its performance and quantify the benefits that it brings.

We are establishing a project with the UKHSA Public Inquiry team to use Large Language Models (LLMs) to review written evidence submitted for upcoming modules to identify key themes and concerns. It is anticipated that this will significantly reduce the amount of time taken to review submissions and support the preparation of senior UKHSA colleagues for oral evidence sessions.

As reported previously, UKHSA colleagues have deployed open source large-language models on an in-house platform called Janus which provides secure use of GenAI models on graphics processing units within the UKHSA network. In parallel, we have started to pilot cloud-based systems for fine-tuning of open-source Large Language Models. This system, which is called Nevis, is allowing the exploration of cloud-capabilities which have certain potential advantages in terms of scalability, access and cost. It is being piloted in conjunction with use cases from the Knowledge and Library Services using publicly available information.

Our focus remains the assessment of open-source models to process text and extract key information to enhance our public health response and develop more efficient processes. The approach we are taking, which includes the use of comprehensive validation sets to assess the effectiveness of LLMs in performing critical public health surveillance tasks like disease classification and symptom extraction, and is increasing our confidence that they are sufficiently accurate to support public health use cases.

9. Update on deployment of Enterprise level AI capabilities

As reported previously, many of the outline expressions of interest in the use of GenAI could be delivered through the deployment of enterprise level solutions, such as Microsoft Copilot. Due to the rapidly evolving AI market, we continue to monitor and assess those tools and capabilities for suitable and optimal digital investment for UKHSA.  Through our existing relationship with CDDO, and the ongoing activities of colleagues in Technology Group, we continue to remain closely aligned with cross-government discussions on deployment.

Clear criteria and guidance has been provided for the implementation of M365 in Government Departments.   This includes both security and information protection guidance.

For UKHSA to implement M365 co-pilot several criteria must be met. Labelling and classification of data as well as permissions are not currently sufficiently well controlled to grant access to 3rd party LLMs. It is estimated that it will take 6 to 8 months to get the data, mark the labels, and correct data classifications before enabling co-pilot, subject to available funding.

In the interim, Bing Copilot search is available to all UKHSA users, providing some GenAI capabilities using external data when used within the published guidelines.

We are meeting with the CDDO AI team, who are leading multiple initiatives to facilitate the effective delivery of AI across departments.  Members of the Taskforce will join a CDDO AI Technical Advisory Board meeting which is being established to identify a consensus on what good AI looks like for UK government. This will include discussions on how to translate ISO42001 into consumable directions for AI projects in the public sector and to provide practical support and coordination of resources. The Advisory Board includes representation for the Alan Turing Institute, British Standards Institute, Microsoft, Google, and AWS, and may include other organisations in due course.

10. Commercial Proof of Concepts & Proof of Value

Several groups within the agency are trialling AI capabilities from commercial suppliers.  We are working with these groups to ensure that any decisions regarding the implementation of these capabilities is aligned to the AI Assessment Framework.

Colleagues in Emergency Preparedness and Response Readiness are trialling First Alert which is an Artificial Intelligence (AI) platform that scans over 1,000,000 distinct sources of Publicly Available Information (PAI) in real time, producing alerts to keep users aware of emerging risks and breaking events with earlier, more comprehensive, and more accurate information. The manual collation of information relating to situational awareness is time consuming and ineffective. As part of the UKHSA 3-years Strategic Plan to improve action on public health through data and insight, and the Ready – to – Respond program, UKHSA’s National Response Centre (NRC) is trialling Situational Awareness Software.

UKHSA Knowledge and Library Services ran a trial of Scopus AI in March. This GenAI tool uses the Scopus bibliographic abstracts database by Elsevier to generate summaries, short reference lists, concept maps, suggested topic experts and related questions in response to natural language prompts; it is intended to speed up an initial exploration of a topic that is unfamiliar to the user. The trial was publicised to around 40 staff working in areas including literature searching, evidence synthesis, evaluation and behavioural science, with around 25% of staff invited sharing feedback.  Initial results are mixed regarding whether the tool would provide benefits to UKHSA staff: some found it provided a useful basic starting point on a topic, but there are also concerns around the tool accurately identifying the most relevant articles based on abstracts alone, as well as quality and accuracy issues with the generated summaries.

Technology Group are leading a Proof of Value exercise with our partners Microsoft and Accenture, looking at how GenAI could improve speed of response and outbreak management on health threats through systematic reviews; establishing the evidence base for public health interventions through the screening of thousands of papers.

11. Next Steps

We are exploring whether we can support individual groups to use Microsoft Copilot on internal data where they have well curated and managed data sets that are ring-fenced. This will support further evaluation of the potential benefits and risks associated with these GenAI capabilities.  It will be subject to available funding.

Our wider staff engagement will continue, with further articles being published on Pulse to highlight uses of GenAI.  As part of these engagement activities, we have scheduled an introduction to AI session during Learning at Work Week which is designed to provide a basic understanding and introduction to GenAI.

There is growing International interest in the use of AI to improve health security and in the coming months the International Association of National Public Health Institutes and 5 Public Health Institutes network (5PHIs) will discuss the potential of this technology.  On 3rd June, UKHSA is hosting a meeting of the 5PHIs and will lead a discussion focused on how to optimize AI in presenting publicly available information to inform early warning of health threats.

We will continue to identify internal use cases where AI and data science can be deployed to increase operational efficiency and health security.  Combining this with a maturation of the governance structures that have been established, including the AI Steering Committee, will support the development of our approach to the adoption and utilisation of AI. The development of these will be dependent upon resourcing and successful funding bids.

12. Annex 1: UKHSA criteria for assessing AI projects

Category	Explanation
Strategic priority	Whether the problem addressed through the AI is a strategic priority for UKHSA.
Appropriateness	An assessment of whether AI is an appropriate technology to apply given the technology-specific benefits and limitations.
Impact and Benefits	A qualitative assessment of the likely size of impact of the AI on the problem compared to non-AI solutions.
Model Validation	An assessment of how well the model performs technically at the task.
Risks	An assessment of the size of risk posed by AI application
Affordability and budget impact	An assessment of financial cost and proportionality to existing budgets associated with the problem area