Memorization in large language models in medicine prevalence characteristics and implications

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Medical Informatics

Background:

• Large Language Models (LLMs) show promise in medicine, often adapted via continued pretraining or fine-tuning on medical datasets.
• A critical concern is the extent to which these LLMs memorize their medical training data.

Purpose of the Study:

• To investigate the prevalence, characteristics, volume, and impact of LLM memorization in medical training data.
• To analyze memorization across various medical adaptation scenarios.

Main Methods:

• Systematic analysis of LLMs adapted through continued pretraining on medical corpora.
• Evaluation of LLMs fine-tuned on standard medical benchmarks.
• Assessment of LLMs fine-tuned on real-world clinical data (over 13,000 inpatient records).

Main Results:

• Memorization in medical LLMs is prevalent and significantly higher than in general-domain LLMs.
• Memorization exhibits distinct characteristics during pretraining versus fine-tuning.
• Up to 87% of memorized content from pretraining persists after fine-tuning.

Conclusions:

• Memorization is categorized as beneficial (e.g., guidelines), uninformative (e.g., templates), or harmful (e.g., sensitive data).
• Recommendations are provided to leverage beneficial memorization, reduce uninformative memorization, and mitigate harmful memorization for privacy and utility.