Making artificial intelligence more inclusive: STANDING Together

Artificial intelligence has huge potential to improve healthcare. However, the reliability of AI algorithms is closely linked to the data it is trained upon. To be sure that algorithms work for everybody, we need to build them with datasets that represent the diverse range of people they are intended to be used in. Without this, AI systems may fail in underrepresented people and worsen health inequalities.  

In our STANDING Together (STANdards for data Diversity, INclusivity and Generalisability) project, we first mapped the problem, discovering the extent to which diversity and inclusion is reported in the types of datasets used to build AI models in health priority areas such as skin cancer, breast cancer, heart failure, eye disease and COVID-19.

Second, we explored causes, interviewing stakeholders to discover the barriers to providing this information.

Finally, we created the STANDING Together recommendations, defining best practice for reporting datasets and AI algorithms, enabling identification and mitigation of potential bias and effect on health inequalities. 

We developed these recommendations through consensus with over 350 stakeholders across 58 countries, including clinicians, academics, computer scientists, industry representatives, regulators, research funders and patients / members of the public.  

This has impacted international and UK policy (across regulation, funding, equity and inclusion), with additional benefits in systems influence, workforce development and improved service user outcomes.