Fake Doctor Notes Help Computers Spot Parole Status

Doctors need to know if a patient is on parole or probation. This fact can shape the right care plan. But it is rarely in the structured parts of a health record. It usually lives in free-text doctor notes. Finding it by hand is slow and expensive. Privacy rules also make it hard to get enough real examples.

A research team built a fake dataset named Synth-PPD. It contains 7,579 made-up clinical notes. Large language models wrote these notes. Filters removed duplicates. Automated checks judged quality. Two people then reviewed 1,517 of the synthetic notes. The goal was to train systems without needing real, expert-labeled data.

The scientists tested the fake notes using several tools. They tried large language models, older machine learning methods, and transformer encoders. They looked at both unsupervised and supervised setups. To check real-world performance, they used 293 actual notes from the U.S. Department of Veterans Affairs. These were labeled by experts. The synthetic data matched human reviewers about 90 percent of the time.

In one test, a model called Llama 3.3-70B-Instruct scored 87.22 percent. It used chain-of-thought prompts and examples from the fake set. That beat a zero-shot version at 80.10 percent. It also beat the best rule-based method at 77.55 percent. All gaps were statistically significant. This shows made-up notes can teach computers to find justice-related details in real medical charts. It could help health teams give continuous care without costly manual labeling.