Artificial intelligence health technologies can cause harm if dataset biases go unrecognised

Without careful dissection of the ways bias can be encoded into artificial intelligence (AI) health technologies, there is a risk of perpetuating existing health inequalities at scale. One major source of bias is the data that underpins these technologies. The STANDING Together recommendations aim to reduce and mitigate the harmful impact of these biases by encouraging transparency around the limitations of health datasets and proactive evaluation of  their impact on AI performance across population groups.

We have created recommendations for the documentation and use of health datasets

Recommendations for Documentation of Health Datasets provides guidance to those creating datasets to enable transparent communication about their composition and any limitations. Recommendations for Use of Health Datasets aims to minimise the risk of algorithmic bias when datasets are chosen and used, and to enable identification and mitigation of any bias which does occur.

Help others to adopt these recommendations by sharing your experience

Development of the STANDING Together recommendations involved representatives from 58 countries. We recognise that not every country and culture is included in this list, and that the recommendations may need contextualising or adaptation in some cases. You can help us by sharing your experience of using these recommendations for a dataset or AI health technology you are responsible for - what works and what doesn't?