Good question. I imagine you would train models using the deidentified datasets cited on the page and design your prediction pipeline in the same vein as any other HIPAA-data related application.