Not What the Doctor Ordered: Surveying LLM-based De-identification and Quantifying Clinical Information Loss

Kiana Aghakasiri, Noopur Zambare, JoAnn Thai, Carrie Ye, Mayur Mehta, Josh Mitchell, Mohamed Abdalla

Introduction

De-identification—the process of removing personally identifiable information (PII) from clinical notes—is a critical prerequisite for medical research and privacy compliance under regulations like HIPAA. As the field shifts from traditional rule-based systems to generative Large Language Models (LLMs), researchers often report "near-perfect" accuracy. However, this high-level performance often masks a significant issue: "over-redaction," where critical clinical information such as medication dosages and diagnoses is inadvertently deleted. This paper investigates the inadequacy of current evaluation metrics and introduces CIRE, a novel LLM-based metric designed to quantify this "clinical information loss" (CIL).

The Challenge of Clinical Information Loss

Traditional evaluation metrics, such as the F1-score or Accuracy, treat all redactions as equal. In a clinical context, however, the impact of an error depends heavily on the nature of the token. If a model correctly removes a patient's name but incorrectly removes a medication dose, standard metrics do not sufficiently penalize the latter. This over-redaction leads to Clinical Information Loss, potentially rendering the resulting data useless for secondary research or clinical care.

Examples of correct and incorrect de-identification, illustrating how over-redaction can remove clinically significant facts.

Methodology

The researchers employed a multi-pronged approach to evaluate the current state of LLM de-identification:

  1. Literature Survey: An analysis of existing research revealed a "wild west" of heterogeneous reporting standards, making direct comparisons between models difficult.
  2. Model Benchmarking: The study evaluated a diverse set of models, including Llama 3, ClinicalBERT, and off-the-shelf tools like Microsoft Presidio, across two datasets: the public MIMIC-III discharge summaries and a private, noisy AHS rheumatology dataset.
  3. Physician Validation: Clinical experts manually reviewed 500 "false positives" (tokens removed that were not PII) to determine if those removals constituted a "Clinically Relevant Change" (CRC).
  4. The CIRE Metric: The authors proposed the Clinical Information Retention Evaluation (CIRE). Unlike lexical metrics that rely on keyword overlaps, CIRE uses an LLM to compare original and de-identified text chunks, specifically judging whether the medical meaning has been altered.
The CIRE framework uses an LLM to judge if semantic medical meaning was lost during the de-identification process.

Key Results and Insights

  • Deceptive Accuracy: Models can achieve over 95% accuracy while still being prone to over-redaction. In the MIMIC-III dataset, 70% to 80% of the tokens incorrectly removed by models were deemed "clinically significant" by physicians.
  • The Length Penalty: Performance is not consistent across document sizes. As clinical notes increase in length, the F1-score of de-identification models typically drops by 5-25%, a trend observed in both LLMs and traditional models.
  • Model Performance: Llama 3 emerged as a top performer on standard metrics (F1 of 0.85 on the AHS dataset) but also exhibited the highest raw count of clinically relevant redactions in that same dataset. ClinicalBERT variants generally achieved higher precision, likely due to their domain-specific pre-training on medical text.
  • Metric Correlation: The proposed CIRE metric demonstrated a strong Pearson correlation (up to 0.86) with manual physician annotations. This far outperforms previous automated metrics like the Jaccard Similarity Coefficient (JSC), which often failed to capture nuances like changes in family history or social determinants of health.

Conclusion

While LLMs offer powerful capabilities for clinical de-identification, their tendency to over-redact poses a risk to the utility of medical data. The introduction of the CIRE metric provides a more nuanced, physician-aligned way to evaluate these models. Future work will focus on developing multi-stage de-identification pipelines that include a secondary review stage to ensure vital clinical facts are preserved in the final output.

Summary updated

Abstract

De-identification in the healthcare setting is an application of NLP where automated algorithms are used to remove personally identifying information of patients (and, sometimes, providers).With the recent rise of generative large language models (LLMs), there has been a corresponding rise in the number of papers that apply LLMs to de-identification.Although these approaches often report near-perfect results, significant challenges concerning reproducibility and utility of the research papers persist.This paper identifies three key limitations in the current literature: inconsistent reporting metrics hindering direct comparisons, the inadequacy of traditional classification metrics in capturing errors which LLMs may be more prone to (i.e., altering clinically relevant information), and lack of manual validation of automated metrics which aim to quantify these errors.To address these issues, we first present a survey of LLM-based de-identification research, highlighting the heterogeneity in reporting standards.Second, we evaluated a diverse set of models to quantify the extent of inappropriate removal of clinical information.Next, we conduct a manual validation of an existing evaluation metric to measure the removal of clinical information, employing clinical experts to assess their efficacy.We highlight poor performance and describe the inherent limitations of such metrics in identifying clinically significant changes.Lastly, we propose a novel methodology for the detection of clinically relevant information removal.
BibTeX
@misc{aghakasiri2025not,
  author = {Kiana Aghakasiri and Noopur Zambare and JoAnn Thai and Carrie Ye and Mayur Mehta and Josh Mitchell and Mohamed Abdalla},
  title = {{Not What the Doctor Ordered: Surveying LLM-based De-identification and Quantifying Clinical Information Loss}},
  year = {2025},
  eprint = {2509.14464},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.14464},
}
Figure