ASQ-PHI: An adversarial synthetic data benchmark for clinical de-identification and search utility

Area of Science:

• Medical Informatics
• Natural Language Processing
• Data Science

Background:

• Hospitals utilize HIPAA-compliant large language models (LLMs) for clinical tasks, allowing Protected Health Information (PHI) input.
• LLMs require external data for up-to-date clinical evidence, necessitating a secure method for handling PHI during this transition.
• Existing de-identification datasets are not optimized for the short, query-style prompts used in chat-based clinical LLM interfaces.

Purpose of the Study:

• To introduce ASQ-PHI, a novel, fully synthetic benchmark dataset for de-identifying clinical queries in the context of LLM safe handoffs.
• To provide a resource for evaluating the de-identification of short, clinician-style search queries.
• To facilitate the development of robust systems for transitioning PHI-containing queries to external, non-BAA-covered tools.

Main Methods:

• Generated 1051 synthetic, single-turn clinical search queries mimicking clinician prompts for HIPAA-compliant LLMs.
• Annotated queries with PHI elements and their corresponding HIPAA Safe Harbor categories using machine-parsable delimiters and JSON.
• Employed an adversarial few-shot prompting pipeline with Azure OpenAI GPT-4o for query generation, including PHI-positive and hard-negative examples.

Main Results:

• The ASQ-PHI dataset contains 1051 queries, with 79.2% being PHI-positive and 20.8% hard negatives.
• 2973 PHI elements are labeled across 13 HIPAA Safe Harbor identifier types, suitable for measuring PHI removal and over-redaction.
• Baseline metrics for a commercial PHI detection service are provided within the dataset's repository.

Conclusions:

• ASQ-PHI addresses the scarcity of suitable datasets for de-identifying clinical queries in LLM safe handoff scenarios.
• The dataset enables the evaluation of de-identification performance on short, query-based prompts, distinct from traditional EHR narrative benchmarks.
• ASQ-PHI is released under an MIT license, promoting further research and development in secure clinical AI applications.