TLDR: A groundbreaking study by Japanese researchers, led by Juntendo University, has demonstrated that retrieval-augmented generation (RAG) can completely eliminate dangerous hallucinations in clinical Large Language Models (LLMs) while significantly improving response times and safeguarding patient privacy. This breakthrough, particularly in radiology consultations, paves the way for safer and more reliable AI integration in healthcare.
TOKYO – A new study published in npj Digital Medicine on July 2, 2025, by a team of Japanese researchers has unveiled a significant advancement in the application of artificial intelligence in healthcare. The research, spearheaded by Associate Professor Akihiko Wada from Juntendo University, alongside Dr. Yuya Tanaka from The University of Tokyo and Dr. Mitsuo Nishizawa from Juntendo University, demonstrates that retrieval-augmented generation (RAG) can effectively eliminate hallucinations in clinical Large Language Models (LLMs) while ensuring robust patient data privacy.
In modern medical environments, particularly in high-stakes areas like radiology, timely and accurate decision-making is paramount. Physicians often face pressure to make critical decisions, such as those involving contrast media consultations, without immediate access to all relevant information. This challenge is compounded for institutions that must avoid cloud-based AI tools to protect sensitive patient data.
To address these issues, the research team focused on locally deployed LLMs for radiology contrast media consultations. Their findings were striking: the RAG-enhanced model achieved a 0% hallucination rate, a dramatic reduction from the 8% observed in traditional models. Furthermore, the RAG system delivered significantly faster response times, averaging 2.6 seconds compared to 4.9–7.3 seconds for leading cloud-based systems. This performance not only matched but in some aspects surpassed cloud models, all while keeping sensitive medical data securely onsite.
Associate Professor Akihiko Wada emphasized the critical safety implications of these results, stating, “For clinical use, reducing hallucinations to zero is a safety breakthrough. These hallucinations can lead to incorrect recommendations about contrast dosage or missed contraindications.” The study highlights that by grounding LLM responses in trusted, real-time data sources, RAG enhances clinical accuracy and increases safety, providing a reliable solution for real-time clinical decision support.
Also Read:
- PodGPT: AI Model Revolutionizes Scientific Understanding by Learning from Podcasts
- Microsoft’s MAI-DxO AI System Demonstrates Superiority in Complex Medical Diagnosis, Outperforming Human Physicians
Beyond radiology, the researchers envision the broad applicability of this technology across various medical fields, including emergency medicine, cardiology, internal medicine, and even medical education. It holds particular promise for rural hospitals and healthcare providers in low-resource settings, offering instant access to expert-level guidance without compromising data security. This study marks a major breakthrough, proving that expert-level AI performance in healthcare can be achieved without sacrificing patient privacy, paving the way for safer, more equitable, and immediately deployable AI solutions in the medical domain.
