I’ll be back.

Top Line: We’re starting to feel the acceleration of literature on artificial intelligence (AI) in oncology. So we thought we should move from the “ahh, cool” stage to the “hmm, let’s dig deeper into the methods” stage.

The Study: Because, let’s be honest, few of us really understand what AI is all about. This article is a good place to start. From it, let’s review a few basic concepts so we can at least pretend we know what we’re talking about. First of all, AI is just an umbrella term that can mean, well, just about anything. More specifically, machine learning is a process whereby systems are developed to make better and better predictions based on more and more data. Deep learning, the typical subject of these articles, is a form of machine learning that specifically uses layers of arithmetic operations to create, or rather “learn,” highly complex systems. All deep learning starts with data that has to be “labeled” in some fashion so that the system can detect “features” with which to learn. Mathematical models are then used to predict an outcome from the data. Defining that outcome, known as a reference standard, is paramount. In many of the studies we read, that “standard” is cancer or no cancer. The process using several “layers” of increasing complexity to characterize data and make predictions is called “training.” Once the model is sufficiently trained, it is tested. And it’s crucial the model be tested on data that comes from outside the data used for training. Testing usually comes in two forms: 1) an “internal” test of data that wasn’t used for training but comes from the original source of training data, and 2) an “external” validation test of data from a completely external source.

TBL: Got it? Good because this week’s headliners will explore what's next in oncologic AI. | Liu, JAMA 2019