UK researchers create method to single out people at highest risk of obesity-related diseases

A new artificial intelligence tool capable of predicting a person’s 10-year risk of developing 18 different obesity-related conditions could help clinicians decide who should be prioritised for weight-loss medications on the NHS, researchers have said.

The tool, named Obscore, was developed by a team from Queen Mary University of London and the Berlin Institute of Health at Charité. By applying a form of AI known as interpretable machine learning to health data from almost 200,000 participants of the UK Biobank project, all of whom had a body mass index (BMI) of 27 or greater — meaning they are overweight or obese — the researchers were able to identify 20 key health, lifestyle and demographic features that together predict the 10-year risk of 18 separate complications. These range from gout and type 2 diabetes to stroke and cardiovascular disease.

How the tool moves beyond BMI

Current NHS eligibility for weight-loss medications such as semaglutide (Wegovy) and tirzepatide (Mounjaro) relies heavily on BMI thresholds and the presence of specific obesity-related health problems. Typically, a patient must have a BMI of 35 or above, or 30 or above with a weight-related condition such as type 2 diabetes or high blood pressure. For people from certain ethnic backgrounds — including South Asian, Chinese, Middle Eastern, Black African or African-Caribbean — a lower BMI threshold applies.

The researchers behind Obscore argue that this one-size-fits-all approach overlooks important differences in individual risk. Their analysis showed that people of the same age, sex and BMI can have dramatically different chances of developing conditions linked to excess weight. More strikingly, for some conditions such as type 2 diabetes, a considerable proportion of people classified as overweight — rather than obese — fell into the highest risk category.

“These constitute a population of individuals who may be overlooked if we only look at BMI and not other risk factors,” said Kamil Demircan, a co-author of the study from Queen Mary University of London. He is a DFG Walter Benjamin Fellow at Queen Mary’s Precision Healthcare University Research Institute and at the Berlin Institute of Health.

The tool evaluates 20 features in total, including age, sex, total cholesterol and creatinine levels. The researchers initially assessed more than 2,000 health indicators drawn from UK Biobank — such as general health, lifestyle factors, clinical data, blood tests, body measurements and molecular markers — before narrowing these down to the most predictive set. By using interpretable machine learning, a subset of AI that makes its decision-making process understandable to humans, the team aimed to build trust with clinicians and patients.

Validation and application to weight-loss drugs

Obscore’s validity was tested not only on UK Biobank data but also on two independent health studies: the European Investigation into Cancer-Norfolk and the Genes and Health studies. For each of the 18 conditions, the tool placed participants into five equal-sized categories ranging from low to high risk, and the team calculated the proportion of people in each category who actually developed the condition over a 10-year period.

The researchers also applied a version of the tool to data from a randomised controlled trial of the weight-loss drug tirzepatide. They confirmed that people who would be predicted to be at highest risk of obesity-related conditions experienced a similar degree of weight loss to others on the drug, suggesting the tool could help identify those who would benefit most from such treatments without any loss of efficacy.

Professor Nick Wareham, a co-author of the study from the University of Cambridge, said the measure was not about extending the use of particular therapies. “It’s about developing and validating a score that can help with more rational resource allocation,” he said. “So, can we prescribe therapy to those people who are most likely to need it and most likely to benefit from it – which is what we should do within the NHS.”

The findings are published in the journal Nature Medicine.

Expert caution: limitations and next steps

Professor Naveed Sattar, a professor of cardiometabolic medicine at the University of Glasgow who was not involved in the work, offered a mixed assessment. He described the research as “a thoughtful attempt to move towards more holistic risk prediction across multiple obesity‑related conditions.”

However, he pointed out that many of the obesity-related conditions are closely interrelated, and that robust and more easily implemented risk scores already exist for some of them. He also noted that several of the metrics used in the Obscore tool are not routinely available within the NHS. “Substantial further development and validation will be required before such an approach can be translated into routine clinical practice,” Sattar concluded.

The context of obesity in England underlines the importance of better targeting. Data shows about two-thirds of adults in England are overweight or obese, with rates highest in the North East and lowest in London, and significantly higher in more deprived areas. Childhood obesity rates also rose notably between 2019-20 and 2020-21. Access to weight-loss injections on the NHS is currently limited; for example, the rollout of tirzepatide is being phased, starting with around 220,000 people with the highest clinical need over three years. From April 2026, prescribing may be incorporated into the GP contract, though participation by practices is optional.

The lead author of the study is Professor Claudia Langenberg, director of Queen Mary University of London’s Precision Healthcare University Research Institute and head of the computational medicine group at the Berlin Institute of Health. She and her colleagues stress that Obscore is a tool for more personalised risk assessment, not a replacement for clinical judgment. The UK Biobank, which provided the foundational data, is a long-term prospective biobank containing de-identified biological samples and health data of half a million people aged 40-69 at recruitment. It has already supported more than 9,000 peer-reviewed publications.

Sattar added that, as effective weight-loss interventions become more evaluated and their costs decline, the reliance on such risk scores might diminish over the next decade. For now, he said, “substantial further development and validation will be required before such an approach can be translated into routine clinical practice.”