Fake patient notes that doctors can often spot
A public set of synthetic Portuguese clinical notes helps research without exposing real patients, but it is not perfect.
Real patient notes are among the most private text there is, so researchers working in smaller languages often have almost nothing to practise on. One answer is to write fake ones.
Carla Teixeira Lopes and Daniel Félix built LusoClin, a public set of synthetic clinical notes in European Portuguese. They started from pre-processed real clinical data, trained an open 3-billion-parameter model on it, and generated new notes. A second model tags the medical terms so the notes can be searched.
The notes were judged on usefulness, realism, variety and privacy. The authors report that they capture clinical language patterns and can support document search. They also say human reviewers can often tell synthetic notes from real ones, and some medical inconsistencies remain.
For people who own or use data, the point is honest labelling. Synthetic data can open up work that privacy rules block, but a buyer needs to know how it was made and where it falls short. A clear description of limits makes such a dataset easier to trust and to price.


