Training data without patient data

Synthetic datasets generated in Python so models could be trained on healthcare data that nobody can be identified from.

Talk to an engineer Data engineering & ML

Healthcare Client

Data engineering & ML Practice

July 2023 Published

01 — The challenge

Where it started.

The client’s data was cluttered and incomplete — a poor basis for training — and it was patient data, so the usual remedies for poor data quality were constrained by privacy.

Rare clinical cases were by definition under-represented, which is exactly where model performance matters most.

02 — What we built

The work itself.

Synthetic records in Python. Clean datasets generated to match the shape of the real ones, standing in for records that could not be used directly.

Rare cases simulated. Scenarios generated deliberately rather than waiting for them to appear in the sample.

A controlled environment. Algorithm performance assessed against data whose properties are known, rather than inferred from a messy sample.

03 — Results

What the client has now.

Privacy protected by substituting synthetic records for real patient data. Rare-case detection and clinical scenario simulation improved. Model performance assessable in a controlled setting. Training and educational material expanded without exposing patient records.

The rest of this study

What we built, what it produced, and the stack behind it.

We share the whole write-up — we would just like to know who is reading. One form opens every case study on the site for the next 30 days.

If this is your problem too

Start with a fixed-price pilot.

Two weeks on a real piece of your backlog, priced before we begin. You keep the output whether or not you continue.

Related work

Consumer credit · Document automation

The manual redaction step, removed

Review management · Conversational AI