Hospital Data for Pharma AI - with the Privacy Risk Measured

The deal that stalls in review

The pharma partner wants to train a model that predicts complications and improves treatment decisions. The hospital has exactly the data it needs: years of admissions, blood-pressure measurements, therapy changes, and outcomes.

Then the project stops. Removing names does not make a medical history unrecognizable: a rare diagnosis, an unusual treatment sequence, or a particular combination of hospital visits can reveal more than any single record. Masking the data heavily destroys the patterns the model needs. Sharing it as is carries a risk no privacy officer can sign off. The partnership waits while reviewers ask a question no one can answer with evidence: is this data safe enough to share?

How PAMOLA gets the release decided

We ran this case end to end in PAMOLA on a synthetic cohort of 7,000 hypertension patients across five connected tables. No real patient data was used. In a real deployment, every step runs inside the hospital’s own environment - patient data never has to leave its control.

• Understand the exposure. PAMOLA maps direct identifiers, quasi-identifiers, and sensitive attributes across all five tables, including the combinations that could make a patient recognizable.
• Create protected data. It generates new, synthetic patient histories with differential privacy - keeping the sequence of admissions, treatments, and outcomes that makes the data valuable for research.
• Attack the result. Before anyone approves anything, PAMOLA tries to break the protection: inferring whether a patient was in the source data, replaying rare treatment paths, discovering hidden attributes, and reconnecting fragments of patient histories.
• Decide against the release rules. The measured results are evaluated against the hospital’s thresholds, and PAMOLA documents why the output passes or fails.

The result: one release approved, one blocked

The full cohort passed. Across the tested attacks, the combined measured privacy risk stayed below 1%. A complication-risk model trained on the synthetic records reached 83% of the accuracy of a model trained on the original data, with both evaluated on previously unseen real records. The protected data stays useful for the partner’s model - and the hospital knows exactly what that usefulness cost in privacy terms.

A rare-disease cohort was blocked. A second release covered just 500 patients with a rare disease. There was no demonstrated breach - there simply wasn’t enough evidence to measure the risk with the required confidence. PAMOLA treated that uncertainty as a reason not to release. Not finding a privacy problem is not the same as proving the risk is acceptable.

The difference is operational: the hospital has evidence to support a defined data-sharing decision, while the pharmaceutical partner can evaluate whether the protected dataset meets its research needs.

A tool that approves everything gives reviewers nothing to trust. One that blocks a release - and shows why - does.

The figures are specific to these demonstration runs. They are not universal guarantees for other datasets or applications.

Watch the walkthrough

Follow the whole release from start to finish: the five tables profiled, synthetic histories generated, four attacks run, and both decisions made. All patients shown are synthetic.

What the hospital gets

• Privacy-transformed dataset - data the partner can use for its model, with research utility preserved.
• Privacy and utility metrics - the measured trade-off between residual disclosure risk and analytical value, not a statement that the data was “anonymized.”
• Attack validation results - what was tested, how the protected data held up, and what blocked a release.
• PAMOLA Evidence Pack - the tests, methods, results, thresholds, and release decision in one record that privacy officers, data owners, and auditors can examine without rerunning the pipeline.

A rejected release is useful too: the hospital sees what prevented approval instead of an unexplained “no.”

Bring the dataset that is actually blocked

If your organization holds patient data that could power AI research but cannot get past privacy review - as a health system, a research network, or a pharma partner - start with one dataset and one intended use. In a 30-45 day Evidence Sprint inside your environment, Realm helps you reach a defensible go / no-go on sharing it, grounded in measured results your reviewers have already seen.