TIDE 2.0 de-identification engine keeps patient links across clinical notes

Researchers have released TIDE 2.0, an open-source engine for removing protected health information from clinical notes while keeping the relationships that longitudinal studies depend on. The work appears in an arXiv paper submitted on 5 October 2026 by Jose D. Posada, Somalee Datta and Priya Desai.

The authors argue that treating de-identification purely as detection falls short. In their account, redaction strips clinical content along with identifiers, blanking dates destroys the time intervals needed for longitudinal analysis, and assigning a fresh random substitute at every occurrence breaks links between one patient's notes.

TIDE 2.0 is described as an MIT-licensed engine with two separable stages: an interchangeable recognizer that finds identifiers and a keyed anonymizer that replaces them. Both are meant to run on hardware the institution owns.

According to the authors, surrogates are generated cryptographically with no stored linkage table. Dates shift by a per-patient offset that preserves intervals, each value gets the same surrogate everywhere it appears under a given key, and a release made under a new key cannot be linked to earlier ones.

The release also includes TIDE2-Sentry, a recognizer distilled from a large language model. On two gold-annotated corpora from two institutions, the authors report span-level recall of 0.88 in-domain and 0.77 on the second institution's corpus, with precision of 0.88 and 0.87. They say they also report recall and precision per category.

The engine is open source, while the recognizer is available only under a gated research-use agreement. The authors say this lets institutions run, inspect and extend both within their own environments.

What is confirmed and what is not: everything here comes from the paper's abstract, which is the authors' own description. The performance figures and the privacy properties, including unlinkability across keys, are self-reported. We found no independent replication or outside coverage, and we did not review the full text, so the datasets, the per-category results and the terms of the research-use agreement are not covered here.

Why it matters: hospitals that want to share notes for research must balance privacy against usefulness. A recall of 0.77 on a second institution's notes suggests that performance can drop outside the setting a recognizer was tuned on, so teams would likely still need local evaluation. This is a single-source item.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.