🩺 lekhansh/modern-gliner-large-mednotes-deidentifier

🔍 Model Summary

This is a fine-tuned version of knowledgator/modern-gliner-bi-large-v1.0, tailored for fine-grained de-identification of real-world medical notes. It performs robustly on noisy, informal clinical text containing misspellings, abbreviations, and shorthand.

• Base model: modern-gliner-bi-large-v1.0
• Fine-tuned on: 1,045 human-annotated medical records; validation on 386.
• F1 Score (baseline): 0.06
• F1 Score (post fine-tuning): 0.998

---

🏷️ Entity Labels

The model recognizes the following sensitive entity types:

• person: HCW, patient or others. Does not capture salutations without names.
• address: Geogrophical entity smaller than a state: cities, districts, taluqs, landmarks and locations within them.
• address state: Indian states and UTs, including misspellings and abbreviations.
• address country
• identification number: Any numeric or alphanumeric identifier
• dates: Only fully specified dates or date ranges that can compromise privacy. 'Jan 2020' is not caught but '03/01/2020' is.
• languages
• groups: Religion, tribal, political groups that can reveal identity.
• company: Named places of occupation or health care facilities.

⚠️ Classification between address types (address, address state, address country) has sub-0.9 accuracy. Treat all address-like entities as sensitive in downstream processing.

---

💡 Intended Use

Primary use: De-identification of free-text clinical notes in electronic health records (EHRs)
Not evaluated for:

• Zero-shot performance on other domains/tasks. Since recall ratio was 0, there will be catastrophic forgetting.
• Multilingual or domain transfer scenarios

---

⚠️ Limitations and Disclaimer

⚠️ DISCLAIMER: HUMAN OVERSIGHT REQUIRED
This model must not be used as a standalone anonymization tool.
Despite high validation performance, errors may still occur—especially with rare entity forms, nested entities, or previously unseen abbreviations. Always include a human audit step before using de-identified records in research or production.

---

🔓 License

Released under a MIT. See repository for terms.

---

📚 Citation

@misc{lekhansh2025deidentifier, title = {Fine-grained Medical Note De-identifier using Modern Gliner Large}, author = {Lekhansh}, year = {2025}, note = {Fine-tuned on real-world annotated records for clinical PII removal}, url = {https://huggingface.co/lekhansh/modern-gliner-large-mednotes-deidentifier} }