The ecological footprint of medical AI

Abstract

Artificial intelligence (AI) is progressively being woven into the fabric of clinical medicine.Most notably, radiology, as an image-based medical specialty, offers numerous applications of AI, including image reconstruction, image analysis, and clinical decision-making.A rich body of research shows the potential of AI in radiology workflows, and indeed, many AI-based medical devices are already available and have received regulatory approval in major markets for clinical routine use.Existing AI models can support radiologists in a wide range of tasks, reaching from image acquisition [1] to diagnosis [2] and outcome prediction [3].In the future, the development of foundation models, i.e., large multimodal transformer neural networks that can manipulate text, images, and other data types, will conceivably increase the use of AI in radiology and medicine even more [4].The advancements in large language models (LLMs), e.g., Chat-GPT, are a testament to these technical developments, as these LLMs already have the capability to reason about clinically relevant topics across a wide range of domains.Behind this effectiveness of contemporaneous AI models are two primary resources: access to large high-quality datasets and the ability to train huge neural networks with up to hundreds of billions of parameters.As these neural networks are trained and are being deployed at scale, a high computational demand is exerted, which consumes electricity and hence leaves a carbon footprint.With the alarming verdict from the Intergovernmental Panel on Climate Change (IPCC-2021) suggesting that

Ähnliche Arbeiten