Artificial intelligence (AI) is transforming both biomedical research and the way we practice medicine. It is essential for hematologists to understand the principles underlying AI (or deep learning) methods to understand and evaluate the results produced by software systems and instruments that rely on these methods. We will see that simple mathematical principles, leveraging the huge increase in computational speed available on modern computers, is the underpinning of AI methods.
The core technology behind the rapid advances in AI is the neural network. In this chapter, we provide a hematology-focused overview of neural networks. The goal is to impart an understanding of how neural network techniques can be most effectively applied to hematology. This chapter surveys core neural network concepts, presents hematology-related examples, and provides references to the rapidly expanding literature. Many different methodologies can be used in the art and the science of AI. This chapter outlines several of the most important techniques currently in use. As will be described in this chapter, neural networks share many principles in common with commonly used techniques such as linear regression and logistic regression.
SUPERVISED LEARNING METHODS
Supervised learning algorithms are designed to uncover relationships between input variables (X) and output variables (Y). Supervised learning requires a training dataset with labeled examples; after learning has occurred, the algorithm is used to make predictions on input variables that it has not yet seen. Examples of supervised learning tasks are myriad. Facial recognition is a well-known example—input variables (X) are images of people’s faces, and the output is the identity of the person. In the hematology domain, a closely analogous task is to distinguish between different types of cells in blood films, marrow aspirates, or biopsies. Supervised learning problems can be subcategorized as regression problems in which the output variables are numbers or as classification problems in which output variables are categories. These examples (face and cell type predictions) are classification problems. A simple example of a regression problem might be to predict height using age, gender, and diet or to estimate house price from the size of house and neighborhood. In the hematology domain, one published example was to use regression methods to predict postradioimmunotherapy hematologic toxicity (transient thrombocytopenia and neutropenia).1
In the following, readers will get an overview highlighting selected key techniques in supervised learning algorithms and how they can be applied to different problems in hematology. A comprehensive description of these methods can be found in Goodfellow and colleagues.2 By starting with the simplest concept of linear regression, one can build on this understanding to better appreciate a nuanced approach to increasingly sophisticated machine learning techniques. This methodology provides the scaffolding to understand the science of artificial intelligence (AI); the art is how these various methods are applied to most effectively solve a given problem.