Why Data Synthesis is Useful

Imagine you are asked to build an AI model that predicts health outcomes for patients. The most obvious source of training material is the records that already exist: real patient files, collected over years of care. Yet in many organizations those records cannot simply be handed to a model-building team. The data is valuable precisely because it describes real people, and that is the reason it is difficult to use.

This article follows that problem step by step. You will trace what happens to real records when they are considered for training, see the approach the course describes for mitigating the problem, which is synthetic data, and then check your understanding by deciding which kind of data to use in a scenario.

Start with a real patient file as the state of the data at the beginning. It holds information about a real person, gathered during care. When a model-building team wants to use that file as training material, the first question is whether the data may be used at all. The file contains sensitive information, and privacy concerns arise whenever data contains sensitive information such as real patient records and real insurance claims.

So the record in its current state is valuable as a description of real people but blocked as training material. The useful question becomes: can the same kind of information be obtained in a form that does not carry the sensitive content? That question leads to synthetic data.

The definition has two halves that work together. The first half says synthetic data is generated: it is produced rather than collected from people. The second half says it mimics real data, so it resembles the original material closely enough to be useful. The final clause is the reason it exists: the generated data does not contain the sensitive information that makes the real records restricted.

Synthetic data is often the only option where human-generated data cannot be used due to privacy concerns. Synthetic patient records and synthetic claims are the two examples the course uses to make this concrete.

The movement of data can be read as a sequence of states. The real patient records begin in a restricted state. A synthetic dataset is then generated to mimic them. The synthetic dataset is in a different state: it is free of the sensitive information. Only that second state can be used to train the model without exposing the sensitive content.

Read the diagram from left to right. The data changes state at the generation step, and the change is what matters: the dataset that reaches the model is the one that no longer carries the sensitive content. The real records remain restricted, and the model is never trained on them directly in this path.

Healthcare is the clearest case in the course material. Patient records describe real people over years of care, and legislation in healthcare restricts their use for training models. Insurance is the second industry the course names: insurance claims are real records that contain sensitive information, so they raise the same privacy concern.

In both industries the data is valuable for exactly the reason it is hard to share. A claim or a patient file is useful because it records what happened to a real person, and that same fact is what makes it sensitive.

The clearest way to see the difference is to set each real record beside its synthetic counterpart. The two columns below pair real patient records with synthetic patient records, and real insurance claims with synthetic claims.

Each synthetic record is paired with the real record it mimics, so it keeps the same general shape as the original. The difference lies in the sensitive information: the real records contain it, and the synthetic records are described as not containing any. This is why the synthetic versions can be used where the originals cannot.