DeidSharp
An open-source .NET toolkit for de-identifying clinical text. It combines local LLMs, rule-based detectors and clinical NER models, improves recall through voting among multiple detectors, and adopts a fail-closed strategy to ensure privacy safety.
Timeline (last 90 days)
DeidSharp is an open-source .NET toolkit that combines locally deployed large language models with rule-based detectors to achieve de-identification of clinical text.
DeidSharp runs a 26B-parameter mixture-of-experts (MoE) LLM locally, layered with deterministic pattern matching and a clinical named entity recognition (NER) model, using multi-detector voting to improve recall.
DeidSharp adopts a fail-closed strategy when detectors err, conservatively marking spans as PHI and masking them to avoid misses.
DeidSharp supports pluggable detectors, with preset recall-first (using all detectors) and precision-first (dropping the NER model) configurations, allowing a trade-off between recall and precision.
The study evaluates DeidSharp on PhysioNet nursing notes (889 PHI spans) and semi-synthetic MIMIC-IV-Note discharge summaries (29,023 PHI spans), and compares it with tools such as deid, Philter, the Stanford/Penn de-identifier, OpenAI Privacy Filter and GLiNER.
On nursing notes, recall-first stacking improved recall by 2.8 percentage points relative to Stanford/Penn (reaching 98.9%), but precision was 9.3 percentage points lower.
On discharge summaries, both stackings led Stanford/Penn by more than 20 percentage points in recall (recall-first reached 95.3%), with the gain coming almost entirely from generated untyped names; on typed labels, all three exceeded 99%.
On deid's "home turf" corpus, no configuration significantly beat deid.
In DeidSharp, the LLM is the primary detector, and NER mainly supplements place names.
DeidSharp uses paired patient-level bootstrap confidence intervals for statistical evaluation, with the tuning set and the reporting set kept separate.
All Facts (10)
DeidSharp is an open-source .NET toolkit that combines locally deployed large language models with rule-based detectors to achieve de-identification of clinical text.
50%UnverifiedDeidSharp runs a 26B-parameter mixture-of-experts (MoE) LLM locally, layered with deterministic pattern matching and a clinical named entity recognition (NER) model, using multi-detector voting to improve recall.
50%UnverifiedDeidSharp adopts a fail-closed strategy when detectors err, conservatively marking spans as PHI and masking them to avoid misses.
50%UnverifiedDeidSharp supports pluggable detectors, with preset recall-first (using all detectors) and precision-first (dropping the NER model) configurations, allowing a trade-off between recall and precision.
50%UnverifiedThe study evaluates DeidSharp on PhysioNet nursing notes (889 PHI spans) and semi-synthetic MIMIC-IV-Note discharge summaries (29,023 PHI spans), and compares it with tools such as deid, Philter, the Stanford/Penn de-identifier, OpenAI Privacy Filter and GLiNER.
50%UnverifiedOn nursing notes, recall-first stacking improved recall by 2.8 percentage points relative to Stanford/Penn (reaching 98.9%), but precision was 9.3 percentage points lower.
50%UnverifiedOn discharge summaries, both stackings led Stanford/Penn by more than 20 percentage points in recall (recall-first reached 95.3%), with the gain coming almost entirely from generated untyped names; on typed labels, all three exceeded 99%.
50%UnverifiedOn deid's "home turf" corpus, no configuration significantly beat deid.
50%UnverifiedIn DeidSharp, the LLM is the primary detector, and NER mainly supplements place names.
50%UnverifiedDeidSharp uses paired patient-level bootstrap confidence intervals for statistical evaluation, with the tuning set and the reporting set kept separate.
50%