By: Grace Huff
Artificial Intelligence (AI) is revolutionizing healthcare, and at the heart of this transformation are foundation models (FMs), large-scale deep learning systems trained on massive datasets and a more general form of chatGPT beyond text. In their recent paper, A Comprehensive Survey of Foundation Models in Medicine accepted by IEEE Reviews in Biomedical Engineering (RBME), researchers Ruogu Fang, Associate Professor and Pruitt Family Endowed Faculty Fellow in the J. Crayton Pruitt Family Department of Biomedical Engineering, along with her Postdoc Associate Wasif Khan, provide an in-depth exploration of these models, their applications, and the challenges they present in modern medicine.
FMs have shifted the way researchers approach AI, evolving beyond specialized, task-specific algorithms to generalizable systems capable of performing multiple functions with minimal additional training. While large language models (LLMs) such as ChatGPT and Gemini have taken center stage, foundation models extend beyond text, spanning fields such as medical imaging, genomics, drug discovery, and clinical decision support.
“Foundation models are transforming our perception of AI and the way we interact with different modalities of data such as satellite images, neuroimaging, genomics, and social networks, just like chatGPT changed the way of interaction with text in late 2022” says Ruogu Fang.
What Are Foundation Models?
At their core, foundation models are large-scale neural networks that use self-supervised learning techniques to generalize knowledge across different tasks. This enables them to be adapted for applications in clinical natural language processing (NLP), medical imaging, biological data analysis, and more. By learning from diverse datasets, these models improve diagnostic accuracy, predictive modeling, and personalized medicine.
“Foundation models serve as a versatile base for a wide range of downstream tasks across different domains, including medicine,” explains Wasif Khan. “Their significance in healthcare stems from their ability to generalize across multiple applications, such as clinical NLP, medical image analysis, omics research, and decision support systems.”
The authors highlight several key areas where FMs are poised to revolutionize medicine:
● Clinical Decision Support – Enhancing diagnostics, treatment recommendations, and patient management.
● Medical Image Analysis – Improving radiology and pathology workflows by automating segmentation, classification, and anomaly detection.
● Natural Language Processing (NLP) in Healthcare – Enabling improved chatbot-based virtual assistants, medical documentation automation, and enhanced information retrieval from clinical texts.
● Drug Discovery and Precision Medicine – Accelerating new treatment development and personalizing patient care based on genetic data.
● Telemedicine and Digital Health – Integrating multimodal data for remote healthcare delivery and patient engagement.
“FMs have the potential to revolutionize medicine due to their ability to learn from diverse and large-scale medical datasets,” says Fang. “Their robustness enables generalization across multiple applications through zero-shot or few-shot learning, which substantially reduces the need for time-consuming and expensive human labeling of medical data.”
“Interpretability is one of the biggest concerns,” explains Khan. “Clinicians need to trust the outputs of these models before integrating them into their workflows. We are seeing progress in developing explainable AI techniques, but there is still work to be done.”
Despite these challenges, the future of FMs in medicine is promising. “I hope to see more lightweight, efficient, and multimodal foundation models integrated into clinical workflows over the next five to ten years,” says Fang. “These advancements could bridge the gap between multimodal medical data, the findings, and disease indications, the three different segments of knowledge that are quite siloed in today’s biomedical and clinical research and practice.”
The authors also emphasize the need for interdisciplinary collaboration: “AI researchers, clinicians, and policymakers should work together to ensure the ethical and responsible deployment of these models,” adds Khan. “Interdisciplinary research and regulatory oversight will be key in guiding the future of AI in healthcare.”
FMs are not just reshaping medicine—they are laying the groundwork for a new era of AI-driven healthcare. As researchers continue to refine these models, their ability to improve diagnostics, streamline workflows, and personalize treatments will only grow. However, addressing the challenges of bias, interpretability, and accessibility will be critical in determining how these models ultimately impact patients and healthcare providers worldwide.
“We provide a clear picture of the current development and taxonomy of FMs since their dawn in 2022,” concludes Fang, “and a roadmap for the larger AI and healthcare communities to see the direction forward.”
For those looking to leverage FMs in medicine, Fang advises: “Prioritize explainability. Developing transparent models will be critical for widespread clinical adoption. Additionally, focus on multi-modal integration, as combining text, images, and omics data can unlock the full potential of FMs in medicine.”
Artificial Intelligence (AI) is transforming medicine, and foundation models (FMs) are leading the way. These powerful AI systems can analyze medical images, assist with diagnoses, and even help discover new drugs.
In their paper, A Comprehensive Survey of Foundation Models in Medicine, researchers Wasif Khan and Ruogu Fang explain how FMs work and their potential impact. Unlike traditional AI, FMs can handle multiple tasks with minimal training, making them valuable in clinical decision support, medical imaging, and drug discovery.
“These models have the potential to revolutionize medicine,” says Khan. However, challenges remain, including trust, fairness, security, and cost. “Clinicians need to trust AI outputs before integrating them into their workflows,” adds Fang. For AI to truly succeed, researchers, doctors, and policymakers must collaborate to ensure ethical and responsible use.
Fang also credits her mentees, students, and collaborators – Seowung Leem (BME PhD student), Kyle B. See (recently graduated BME PhD student and now postdoc at UF), and Joshua K. Wong, M.D. (Assistant Professor and clinical collaborator at UF Neurology and Norman Fixel Institute for Neurological Diseases).
“The greatest joy comes from working with brilliant minds at the forefront of AI!” Fang said.
This paper was accepted and is available here with early access. This paper will officially appear in the yearly issue of the IEEE Reviews in Biomedical Engineering journal in January 2026.