Machine learning and artificial intelligence (AI) are very trendy right now, with coverage in the popular press, accompanied by more than 45,000 articles on PubMed, and more every day. Part of this popularity is certainly because of some major advances in machine learning’s capabilities: recent studies have achieved results that simply weren’t possible just a decade ago. But this popularity also runs the risk of overpromising what machine learning can actually deliver in healthcare, at least in the short and medium term.
How should a health system leader engage with machine learning? The goal of this chapter is to give you an understanding of some of the key points of machine learning. Specifically, we are hoping to help you develop an intuition for how machine learning works and what it can do … but we are specifically not trying to teach you how to implement machine learning algorithms. We assume that you’ve read the chapters that come before this one, focused on data in healthcare and traditional epidemiological modeling techniques—and we’ll build on that knowledge to help you better understand machine learning, AI, and what they can do in healthcare.
DEFINITIONS: WHAT IS ARTIFICIAL INTELLIGENCE? MACHINE LEARNING?
General Versus Specific Artificial Intelligence (AI)
An important distinction separates general from specific AI. Artificial general intelligence (AGI) is the stuff of science fiction—a computerized intelligence that can learn, handle any task that a human intelligence can take on, and become self-improving. AGI would be able to handle the same range of problems that a human could (or more) and do them as well as (or better than) a human. This is also sometimes called strong AI, and, importantly, it does not exist today. However, AGI is an active area of research for many groups around the world.
Artificial general intelligence (AGI)—a computerized intelligence that can learn, handle any task that a human intelligence can take on, and become self-improving—does not exist today, though it is an area of active research.
On the other hand, specific AI (also called narrow AI and weak AI) focuses on one specific, narrow task (hence the name) and gets really good at doing that one task. For example, you can now search through your home photos to find all the pictures of cats (or flowers, meals, or family members) without ever having tagged your photos with information about what’s in them. Similarly, a research team trained an algorithm to identify diabetic retinopathy from photos.1 That algorithm was then shown to be better than U.S. board-certified ophthalmologists, but just at that one, narrow problem.2 Ask the algorithm to fill out a form, have a conversation, or go grocery shopping, and the difference between general and specific AI becomes immediately apparent.
Specific, narrow AI seems likely to materially affect healthcare ...