Geometrically-aggregated training samples: Leveraging summary statistics to enable healthcare data democratization

Healthcare data is highly sensitive and confidential, with strict regulations and laws to protect patient privacy and security. However, these regulations impede the access of healthcare data to a wider AI research community. As a result, AI healthcare research is often dominated by organisations with access to larger datasets or limited to silo-based development, where models are trained and evaluated on a limited population. Taking inspiration from the non-sensitive nature of the summary statistics (mean, variance, etc.) of healthcare data, this paper proposes geometrically-aggregated training samples (GATS) where each training sample is a convex combination of multiple patients’ characteristics. Thus, mappings from patients to any constructed sample are highly convoluted, preserving patient privacy. We demonstrate that these “summary training units” provide effective training on different tabular and time-series datasets (CURIAL, UCI Adult, and eICU), and indeed behave as a summary of the original training datasets. This approach takes important steps towards data accessibility and democratization.