Send message to Author

Paper short abstract

Scaling healthcare data for AI raises issues of quality, bias, and privacy. Wearables add skewed or missing data. Synthetic data offers solutions but raises issues around trust, ownership, and regulation. The paper explores the political-economy of using synthetic datasets for AI in healthcare.

Paper long abstract

Scaling healthcare data for AI opens several problems. Varying quality of data inputs and abstraction from contextual variables – primarily, the social determinants of health – make it harder to ensure data quality. It is more likely that datasets contain bias and exclusions, and that it is harder to ameliorate their effects. Health apps and wearables provide ever more healthcare data points on which models can be trained but also introduce the likelihood of missed data points due to faulty signals as well as skewed data. In addition, with increasingly detailed intersecting data points it becomes harder to ensure the privacy of individuals.

Synthetic data has been presented as a solution to these concerns. Firms specialised in synthetic data build on long-standing data hygiene practices (e.g. filling in missing data points, bias control, etc) but also offer fully simulated data sets that seek to emulate real data sets but with sufficient alteration to be distinct. Currently there is global disagreement over whether to trust synthetic data in medical certification, with the EU opposed but US regulators much more supportive even prior to the current administration. Whilst in many jurisdictions a patient’s health data is their property, meaning access and usage requires their consent, modifying that data to be ‘synthetic’ complicates who owns it. The paper maps these political-economic aspects, outlining how synthetic data impacts how we think of healthcare as well as its broader political consequences for regulation, states, as well as individuals.

Keywords: Synthetic data, Healthcare, AI, Political-Economy, Regulation