Foundation Models in Medtech: A Game-Changer in Healthcare
In recent times, the medtech industry has witnessed a surge in the adoption of “foundation models,” a type of artificial intelligence that shows promise in revolutionizing various tasks. However, there is still some ambiguity surrounding the use of this technology in the field of medtech.
Noteworthy Developments in Foundation Models
Over the past year, major players in the healthcare sector such as GE Healthcare and Philips have taken significant strides in promoting the use of foundation models. GE Healthcare introduced an MRI research foundation model, while Philips collaborated with Nvidia to develop a foundation model for MRIs. Additionally, abstracts presented at the Radiological Society of North America meeting focused on evaluating and enhancing foundation models. The Food and Drug Administration (FDA) has also updated its database of AI-enabled medical devices to explore ways of identifying and tagging devices incorporating foundation models.
Defining Foundation Models
Magdalini Paschali, a postdoctoral scholar at Stanford’s Department of Radiology, outlined key characteristics of foundation models. These models are trained on extensive datasets, predominantly composed of unlabeled data, and have the capability to process diverse data types such as images, text, medical history, and genomics. Furthermore, they can perform a wide array of tasks, including detecting diseases that were not part of their training data.
The Emergence of Foundation Models
The term “foundation model” was coined at Stanford in 2021, with Google’s Med-PaLM marking one of the initial versions in healthcare. These models gained prominence at the 2023 RSNA event. In contrast to traditional deep learning models that focus on specific health conditions and rely on labeled data, foundation models are trained on millions of images, making the requirement of labeled data impractical.
Are Foundation Models More Accurate?
Some medical device developers claim that foundation models offer higher accuracy than narrow AI models. For instance, Aidoc asserts that foundation models facilitate the rapid development of more precise AI tools. However, the accuracy of these devices depends on their construction. While a foundation model might initially perform inferior to a specialized AI tool, it can excel once it gains exposure to examples and context.
Evaluating Foundation Models
Currently, FDA-authorized foundation models are tailored to address specific tasks, such as Aidoc’s rib fracture triage tool. However, guidelines for broader models encompassing language, images, or video are yet to be established. Hospitals have devised methods to evaluate AI models, but these approaches have limitations. Ensuring the accuracy of a foundation model involves defining metrics and tasks, testing performance across different patient subgroups and scanner types, and conducting stress tests to uncover potential issues.
Looking Towards the Future
The aspiration is that foundation models, trained on extensive datasets from diverse sources, will necessitate less evaluation before deployment. This could potentially alleviate the burden on radiologists, particularly amidst a shortage of radiologists in the U.S. and an escalating volume of medical images. As the medtech industry continues to explore the capabilities of foundation models, the quest for enhanced patient care and operational efficiency remains at the forefront of these technological advancements.
