K-anonymity decay in multi-turn clinical large language model conversations

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Health Data Privacy

Background:

• Per-prompt de-identification is standard for clinical AI conversations.
• The Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor method requires removing 18 identifier categories per disclosure.
• Current methods do not assess cumulative re-identification risk across conversational turns.

Purpose of the Study:

• To quantify the re-identification risk in multi-turn clinical AI conversations.
• To evaluate the effectiveness of per-prompt de-identification against cumulative quasi-identifier disclosure.
• To highlight limitations in current privacy protection methods for clinical AI.

Main Methods:

• Simulated progressive quasi-identifier disclosure based on clinical case presentations.
• Modeled disclosure against a synthetic electronic health record cohort.
• Quantified patient re-identification risk by tracking proximity to the small-cell threshold (k < 5).

Main Results:

• 79.9% of simulated patients fell below the k-anonymity threshold by the end of disclosure sequences.
• A median of seven disclosure steps were needed to reach the threshold; this decreased to four steps when rare attributes were disclosed first.
• Cumulative quasi-identifier profiles degraded k-anonymity below safety thresholds, even without direct identifier disclosure.

Conclusions:

• Per-prompt de-identification is insufficient for multi-turn clinical AI and large language model conversations.
• Clinicians lack real-time tools to assess cumulative re-identification risk.
• Existing HIPAA Safe Harbor provisions may not adequately address the evolving privacy challenges in AI-driven healthcare.