Don’t be data poor — Anuj Iravane, Anterior
Anterior works with medical data often received as faxed documents over 300 pages long containing handwriting and tables — but they cannot store this data due to legal requirements.
Instead of attempting to train AI on data they are not permitted to retain, they reverse the process: they start with the correct answer (diagnosis or decisions), work backwards through the reasoning, and then generate the patient records that would have led there. Because they model medical guidelines as decision trees, the generated material has greater variation and is more realistic than if an AI model simply tried to create data on its own. As a result, approximately 90 percent of their training data is now synthetic, and doctors can distinguish synthetic from real data only about 60 percent of the time.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
