De-identified data

De-identified data is health data from which the patients' names and other identifying details have been removed, so that many patients' records can be pooled and studied without anyone being able to tell whose record is whose. Researchers, drug companies and AI developers use it to learn from millions of patients at once.

Here is how it works in general. A hospital's records contain direct identifiers, which point straight to a person: the name, phone number, address, ID numbers such as Aadhaar, and the patient number. Software removes all of these. Other details do not name a person, but could point to one when they are combined. Examples are the exact date of birth, the PIN code and the date of a hospital visit. The software blurs these. For example, it keeps the age in years instead of the birth date, or the district instead of the PIN code. Often it also replaces the patient number with a token, a different number that is the same in every record of that patient. So researchers can still join one patient's lab results, prescriptions and claims, without learning who the patient is. This is a form of tokenisation.

Removing the name is not always enough. A patient with a very rare disease in a small town may be recognisable from the disease and the town alone. So experts test how easily a person could be identified again, by combining the data with other public data. This is called re-identification. Contracts also forbid the receiver from trying it. In America, the HIPAA law sets out two accepted methods of de-identification. India's Digital Personal Data Protection Act applies to any data that can still identify a person.

Take, for example, a real-world evidence platform. It pools the de-identified records of millions of patients to find out how a medicine works in older people, whom the original trial left out.

Consider a company that trains an AI model to read chest X-rays. It trains the model on tens of thousands of de-identified X-rays and their reports, with the names removed from the images as well.

Say two companies want to know how many customers they share. Outside healthcare, a data clean room lets them match their customer lists without either company seeing the other's names.

How is de-identified data different from encrypted data? Encrypted data still contains every name, and anyone with the key can read it. De-identified data no longer contains the names at all. It is also different from data masking, which hides part of a value, such as all but the last four digits of a phone number, on one screen.