TIDE 2.0 pairs clinical-note de-identification with longitudinal links

Clinical notes capture much of what is recorded about a patient's care, but research use requires removing protected health information first. The paper says finding identifiers alone does not solve the problem, because deleting them can also remove clinically useful content. Replacing dates with blanks, the authors say, erases time intervals that longitudinal analysis depends on. Using a different random substitute every time a value appears can also break connections among one patient's notes.

TIDE 2.0 is presented as an open-source, MIT-licensed engine with separate recognition and anonymization stages. Its recognizer can be exchanged, while the keyed anonymizer applies consistent surrogate values to protected information. The system is described as model-agnostic, and both stages run on hardware owned by the institution. That design retains links between notes and temporal intervals needed for longitudinal analysis, the paper says.

Surrogate values are produced cryptographically, and the engine keeps no table mapping them back to source values. Within a key, the same underlying value gets one stable surrogate wherever it recurs. Dates are moved by a patient-specific offset that preserves intervals between them. The authors say a release made with another key cannot be connected to previous releases.

The release also includes TIDE2-Sentry, a recognizer distilled from a large language model. The default setup was assessed on two gold-annotated datasets from two institutions. It achieved span-level recall of 0.88 in-domain and 0.77 on the other institution's corpus, with precision of 0.88 and 0.87, respectively. The paper also reports recall and precision separately by category, alongside these aggregate results.

Access is split between the engine and recognizer: the engine is open source, while the recognizer requires a gated research-use agreement. The paper says institutions can run, inspect and extend both within their own environments. For AI teams evaluating the system, the reported figures describe its default configuration across two institutions, while category-level results are also available in the paper.