TLDR: A study evaluated eight large language models (LLMs) on their ability to recall current medical knowledge using two new datasets, MedRevQA and MedChangeQA, derived from systematic reviews. It found that all LLMs consistently rely on outdated medical information, performing better on older knowledge. This issue stems from older findings being more prevalent in training data. The research highlights the safety risks of LLMs providing obsolete medical advice and proposes mitigation strategies like retrieval-augmented generation (RAG) and knowledge editing to improve their currency and reliability in healthcare.
Large Language Models (LLMs) are rapidly transforming various fields, and healthcare is no exception. These powerful AI systems hold immense potential to assist medical researchers and physicians, offering new avenues for democratizing health services and streamlining clinical workflows. However, a critical challenge emerges from their reliance on static training data: medical knowledge is constantly evolving, and what was considered accurate yesterday might be outdated today.
A recent study, “Facts Fade Fast: Evaluating Memorization of Outdated Medical Knowledge in Large Language Models,” by Juraj Vladika, Mahdi Dhaini, and Florian Matthes from the Technical University of Munich, delves into this significant problem. The researchers highlight that when LLMs memorize outdated medical information, they risk providing harmful advice or failing in crucial clinical reasoning tasks. This issue is particularly concerning because, even when augmented with up-to-date information, LLMs can sometimes disregard it and revert to their internally stored, obsolete knowledge.
Investigating Outdated Knowledge
To thoroughly investigate how LLMs handle evolving medical knowledge, the researchers developed two new question-answering (QA) datasets. The first, MedRevQA, comprises 16,501 QA pairs covering a broad spectrum of general biomedical knowledge. The second, MedChangeQA, is a specialized subset of 512 QA pairs where medical consensus has demonstrably shifted over time. These datasets were meticulously constructed from systematic literature reviews (SLRs) published by the Cochrane Collaboration, which are considered the highest quality evidence in medicine.
The study benchmarked eight prominent LLMs, including GPT-4o, Mistral 24B, Llama 3.3, Qwen 2.5, DeepSeek-V3, OLMo 2, PMC-LLaMa, and BioMistral. The goal was to evaluate the models’ internal knowledge and memorization without providing additional context. The models were tasked with predicting one of three labels: Supported, Refuted, or Not Enough Information, based on medical questions.
Key Findings: A Reliance on the Past
The evaluation revealed a consistent and concerning reliance on outdated medical knowledge across all tested LLMs. When comparing performance on questions with outdated versus latest labels, most models showed a struggle to adapt to current medical consensus. For instance, Llama 3.3 showed the highest degree of awareness of the latest knowledge, while models like GPT-4o, Qwen, DeepSeek, and PMC-Llama struggled more significantly. The study also observed a decline in LLM performance on questions related to more recent years, suggesting a stronger memorization of older information.
The researchers hypothesize several reasons for this phenomenon. Older scientific findings have had more time to permeate the internet, news, and subsequent studies, making them more prevalent in the vast pre-training corpora of LLMs. Additionally, the frequency of appearance in training data, learning rate, and model size can influence memorization. For example, an inspection of OLMo 2’s training data, the Dolma corpus, showed that earlier Cochrane SLRs were more prevalent, potentially leading to stronger encoding of older knowledge.
Also Read:
- The Shifting Sands of Knowledge: How Evolving Text Challenges LLM Comprehension
- Assessing LLM Capabilities: A New Framework to Counter Data Contamination
Moving Forward: Mitigating the Risk
The implications of LLMs memorizing outdated medical knowledge are significant, posing a major safety concern for their application in healthcare. Providing incorrect health advice or flawed facts in clinical reasoning could have serious consequences. The study emphasizes the need for mitigation strategies to ensure these AI systems remain current and reliable.
One promising direction is Retrieval-Augmented Generation (RAG), where LLMs are augmented with retrieved up-to-date information. Even a simple RAG strategy, like adding the abstract of the top PubMed result to the prompt, showed improvements in F1 scores. However, advanced RAG techniques are needed, including filtering and re-ranking retrieved evidence by recency and source quality, to prevent issues like hallucination or rejection of provided references. Other future directions include resolving knowledge conflicts, machine unlearning, knowledge editing, and continual learning. For more details, you can read the full research paper available at arXiv.
This research lays crucial groundwork for developing more current and reliable medical AI systems, ensuring that the benefits of LLMs in healthcare are realized without compromising patient safety due to obsolete information.
