Memorization in large language models in medicine prevalence characteristics and implications

2026-06-19 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Health & Medical Research, Artificial Intelligence & Machine Learning · Depth: Expert, short

Summary

A study investigates memorization in Large Language Models (LLMs) adapted for medical applications, analyzing its prevalence, characteristics, volume, and implications. Researchers systematically examined three adaptation scenarios: continued pretraining on medical corpora, fine-tuning on standard medical benchmarks, and fine-tuning on over 13,000 real-world inpatient records from Yale New Haven Health System. The findings reveal that memorization is significantly more prevalent in medical LLMs compared to general domain models. It exhibits distinct characteristics during pretraining and fine-tuning, with up to 87% of memorized content persisting after fine-tuning. The study categorizes memorization into three types: beneficial (e.g., accurate recall of clinical guidelines), uninformative (e.g., templated language), and harmful (e.g., sensitive clinical content). Practical recommendations are provided to manage these different forms of memorization.

Key takeaway

For AI Scientists and Research Scientists developing medical LLMs, understanding memorization is critical for ethical deployment. You must implement strategies to protect patient privacy by mitigating harmful memorization, especially given its high prevalence and persistence (up to 87%) in medical contexts. Focus on techniques that facilitate beneficial recall of clinical guidelines while minimizing uninformative content to improve model utility and safety.

Key insights

LLMs adapted for medicine exhibit high, persistent memorization of training data, requiring careful management of beneficial, uninformative, and harmful types.

Principles

• Medical LLM memorization is prevalent and persistent.
• Memorization types include beneficial, uninformative, and harmful.
• Harmful memorization risks patient privacy.

Method

The study systematically analyzed LLM memorization across three adaptation scenarios: continued pretraining on medical corpora, fine-tuning on medical benchmarks, and fine-tuning on 13,000+ real-world clinical records.

In practice

• Facilitate beneficial memorization.
• Minimize uninformative memorization.
• Mitigate harmful memorization.

Topics

• Large Language Models
• Medical AI
• Data Memorization
• Patient Privacy
• Clinical Data
• Model Fine-tuning

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• When eight popular large language models (LLMs) answer common parent questions about early maxillary expansion, how reliable are they—and how readable is what they say? — LLM performance in healthcare varies significantly across accuracy, comprehensiveness, and readability, with no single model excelling in all.
• MIT scientists investigate memorization risk in the age of clinical AI — AI models trained on EHRs can memorize patient data, necessitating rigorous testing to prevent privacy breaches.
• Real-world safety and harms from patient-facing LLMs — Real-world data on AI chatbot harms in health is limited but growing, highlighting mental health risks.
• Benchmarking GPT Models for Conversational AI Systems: Can AI Read a Doctor’s Notes? — LLMs struggle with medical abbreviation expansion, especially context-dependent ones, despite larger models performing better.
• Please don’t trust your chatbot for medical advice — LLMs are unreliable for medical advice, frequently generating misinformation with overconfidence and lacking clinical reasoning.
• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.