Could art help support mental health in schools and care settings?

A machine-learning models surfaced a promising body of research that had attracted little policy attention. The finding offers a useful example of how a broad scientific knowledge base, precision AI and expert judgement can work together.

When Basil Mahfouz used machine learning to look for research with unrealised policy value, one result surprised him.

The models highlighted studies on using art to support mental health in schools, hospitals and care for older people. According to Mahfouz, the work showed promising social outcomes, yet it had been barely used in conventional policymaking.

The model had not discovered a new treatment or decided what schools should do. It had brought an overlooked area of research into view, giving specialists a reason to examine it more closely.

That distinction is central to Mahfouz’s work.

Watch: Basil discusses the way art can be used to support mental health

Searching by meaning, rather than visibility

Research already cited by governments is relatively easy to identify. The harder question is what they may have missed.

During his PhD at UCL, supported by Elsevier, Mahfouz used language models to compare papers that had reached policy with similarly relevant papers that had not.

The process began with research cited by governments and policy organisations. Language models then identified uncited research covering similar ground, either by pairing each cited paper with its closest match or by scoring millions of papers on how closely they resemble cited work in the field.

This made it possible to look beyond the most visible or familiar sources. The analysis could surface neglected topics, such as art and mental health, as well as individual researchers whose work appeared relevant to policy but had received little attention.

Making a large evidence base usable

The volume of published research has grown far beyond what any policy team can read. Sir Geoff Mulgan, Professor at UCL and Mahfouz’s doctoral supervisor, describes the result as a “fire hose” of information: the knowledge exists, but methods for selecting from it have struggled to keep pace.

A broad, structured scientific knowledge base provides the coverage needed to search widely. AI helps narrow that material by comparing meaning and relevance across large numbers of papers.

The final assessment remains a human responsibility. Experts need to judge the strength of the evidence, the populations studied, the limits of the findings and whether they apply to the decision under consideration.

Mahfouz also cautions that predictive models learn from previous patterns of policy use. Those patterns may contain gaps and biases of their own, so model outputs need careful review.

The research on art and mental health may or may not lead to a particular policy choice. Its value here is that it became visible. With a wider evidence search and informed expert review, decision-makers can consider useful work that established networks and conventional discovery methods may have left outside the frame.

Watch the full interview