---
title: "AI-generated records improve cross-hospital de-identification"
description: "A study using GPT-4-generated clinical text reports better cross-dataset de-identification scores after one-shot augmentation."
url: "https://digitvibe.com/blog/privacy-safe-clinical-data-augmentation/"
---

[All blog posts](https://digitvibe.com/blog/)

Generative AI in organizations and society

# AI-generated records improve cross-hospital de-identification

A study using GPT-4-generated clinical text reports better cross-dataset de-identification scores after one-shot augmentation.

By DigitVibe blog · Published November 17, 2026

*   Peer reviewed · 2024

A clinical de-identification system must recognize names, dates and other protected health information in records it may not have seen before. That task is difficult because hospitals can use different formats and labeling rules. Kim and colleagues test whether synthetic examples generated with GPT-4 can improve performance when a model moves between datasets.

![Complete abstract from Generalizing Clinical De-identification Models by Privacy-safe Data Augmentation using GPT-4.](https://digitvibe-blog.workspace-scripts.workers.dev/media/2f9d064f0fcd4485784767e386d2198bf0719fbcbac7be7af2c720699dce085f.webp)

Complete abstract from the original paper. Woojin Kim, Sungeun Hahm, Jaejin Lee (2024), [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). [Source paper, page 1](https://doi.org/10.18653/v1/2024.emnlp-main.1181) [View full-size excerpt](https://digitvibe-blog.workspace-scripts.workers.dev/media/2f9d064f0fcd4485784767e386d2198bf0719fbcbac7be7af2c720699dce085f.webp)

### Augmenting data without sending raw identifiers

The study uses the i2b2 2006 and 2014 clinical de-identification datasets. Its approach asks GPT-4 to generate synthetic clinical-style examples, with protected information redacted before the text is sent to the model. The authors compare zero-shot generation, which does not require an example dataset, with one-shot augmentation, which uses a small amount of source data to guide generation. They then train and test de-identification models across the two datasets (method and experimental setup, PDF pp. 1–5; proceedings pp. 21204–21208).

The task is a form of sequence tagging: a model marks spans of text that belong to protected-identifier categories. The evaluation reports three F1 measures. Binary token-level F1 asks whether each token is protected or not; token-level micro F1 aggregates category-level token decisions; entity-level F1 evaluates complete identifier spans. These scores reflect different kinds of errors, so a gain in one does not necessarily mean every identifier type is handled equally well.

### Cross-dataset scores rise with one-shot examples

In tests transferring between i2b2 2006 and 2014, one-shot augmentation improves results across the reported F1 measures. The paper reports gains of up to 6.48%, 29.83% and 36.50% for binary token-level, token-level micro and entity-level F1, respectively, compared with the no-augmentation cross-dataset setting (Table 2, PDF pp. 9–10; proceedings pp. 21212–21213). The biggest improvement is for complete entity spans, which can be harder to capture than individual protected tokens.

These results support the value of extra, task-shaped training examples in this transfer setting. They do not establish that a synthetic corpus will improve every hospital’s system, or that the generated records reproduce the full variety of real documentation.

### Privacy and transfer still need review

Redacting PHI before querying an online model is a privacy measure described by the authors, not a complete privacy guarantee for every workflow. Data governance, redaction quality, provider terms and residual disclosure risks remain relevant. The study also tests two benchmark datasets; formats and policies in other institutions may differ. The authors discuss limits around dataset variation and generalization (limitations, PDF pp. 12–13; proceedings pp. 21215–21216).

### Possible implications for health data systems

The work suggests that carefully controlled synthetic augmentation could help address gaps between clinical datasets. It also shows why de-identification evaluations should measure both token recognition and complete spans. Whether the same gains hold for current records and locally defined labels is a separate empirical question.

## Bibliography & sources

1.  Woojin Kim, Sungeun Hahm, and Jaejin Lee. “Generalizing Clinical De-identification Models by Privacy-safe Data Augmentation using GPT-4.” In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 21204–21218 (2024). [https://doi.org/10.18653/v1/2024.emnlp-main.1181](https://doi.org/10.18653/v1/2024.emnlp-main.1181).

---

DigitVibe. Index of public content: https://digitvibe.com/llms.txt
