Medical Data as Commodity: Why Your Health Records Are the Most Valuable Training Data

In 2021, a little-known company quietly purchased the de-identified medical records of approximately 50 million Americans for around $1 billion. The data included diagnoses, prescriptions, lab results, and visit histories, stripped of personal identifiers like names and Social Security numbers. The hospitals and health systems that collected this data never informed the patients that their records were being sold to build AI products.

The buyers, who included insurers, pharmaceutical companies, and AI labs, were unaware that they were purchasing the most valuable training data in medical AI. Your health records are the most valuable training data because they are longitudinal, labeled, scarce, and multimodal. They provide a decade-long record with outcomes, structured codes, and physician annotations, making them a goldmine for models.

Unlike social media posts or browsing history, health data is genuinely rare and worth more. De-identification, governed by HIPAA and GDPR, does not equate to anonymization. A study published in Nature Communications found that de-identified health data is often trivially easy to re-identify. The consent patients give when they enter a hospital is not for model training, but for treatment, payment, and operations.

The transaction happens in a language patients are not fluent in, on a timeline they cannot control, and they receive no compensation. The medical AI industry is built on extracting, packaging, and selling health data, one electronic health record (EHR) at a time, without patient knowledge or consent.