32.1 Introduction

Artificial intelligence (AI) is a discipline at the intersection of computer science, medicine, and philosophy that aims to understand and replicate human intelligence using machines, typically computers. The most tangible manifestation of this field is machine learning: a field within computer science focused on the creation of computer programs that can learn associations from real-world examples—a task that has to date required human intelligence. Current examples of machine learning are all “narrow” AI: systems designed to perform a specific task such as analyze a photo, create a melody, or interpret written text. The ultimate challenge in AI is the creation of artificial general intelligence, also called “broad” AI: a machine that matches or exceeds the holistic nature of human intelligence.

32.2 Machine Learning: Making Associations from Real-Word Examples

The process of machine learning involves creating a computer program that itself makes associations based on examples from real-world data, and then uses these associations to predict future events. The task of developing associations is performed by an algorithm or process that creates a model to describe associations between elements of the data known as features. The rapid growth in both data and computational power has supported a commensurate increase in the performance and variety of machine learning algorithms. One type of algorithm that has recently gained prominence is the Artificial Neural Network or Neural Net. These algorithms were first proposed 70 years ago as a way of representing the information-processing capabilities of human neurons in silico. They have more recently enjoyed a renaissance through wins with tasks as diverse as image recognition and autonomous transportation—situations where vast quantities of structured, digitally available data is present. The increase in performance is due to more structured data, and with more data greater sophistication of the neural nets themselves: adding more layers and a control mechanism known as back propagation. These advances created deep neural networks (the depth referring to the number of layers of neurons) and a new science of deep learning.

32.3 Supervised versus Unsupervised Learning

Deep learning is unique in that the algorithm is designed to learn associations in the data without any supervision by a human as to which examples, or which elements of an example, are interesting. This form of learning is called unsupervised and is contrasted with supervised learning which requires an initial dataset labeled by a human expert that is used to train an algorithm such that it can produce useful outputs.

The difference between supervised and unsupervised learning can be better understood by extrapolating from Google’s landmark 2012 study. ¹ In the case of supervised learning, a machine is fed a series of pictures labeled “cat” and a series of pictures labelled “human”; this is the training dataset. If the machine is subsequently shown new pictures, it should be able to identify them as either “cat” or “human.” The machine will get better at identifying cats and humans as it is shown more pictures. In the case of unsupervised learning, a machine can browse through millions of randomly selected images on the Internet and teach itself to categorize these pictures into different groups corresponding with “humans” and “cats,” and also other groups such as “horses,” “hamsters,” et cetera, without the need for a human expert. As the machine scans bigger and bigger datasets, the risk of incorrectly classifying a picture is reduced and its confidence is said to increase. The quality and quantity of healthcare data is expanding with, for example, the roll out of electronic health records, and the online publication of research and outcomes. These increasingly large datasets will support both supervised and unsupervised machine learning.

32.4 Artificial Intelligence and Medicine

It can be argued that all tasks in medicine are in essence tasks of information processing, and that there are two high-level streams of information processing work: diagnosis and management. Diagnosis is the task of inference based on data (reported from patients as symptoms, elicited as signs, or received from investigations). Management is the task of implementing and monitoring a process to achieve a therapeutic endpoint. Surgery involves combinations of these tasks as the surgeon assesses the patient (akin to diagnostic inference), and takes the next action to deliver the surgical intervention (akin to management).

We propose that machine learning algorithms will provide support to clinicians in two key areas: diagnosis support and decision support. Current applications of machine learning involve automation of diagnosis. However, initiatives are under way to use machine learning to optimize clinical decision making/management and create personalized, adaptive management algorithms at scale.

32.5 Current Applications of Artificial Intelligence in Medicine

Examples of diagnosis support can be found in fields such as dermatology. Using deep neural networks and a training dataset of 129,450 clinical images, a supervised machine learning system achieved results equivalent to 21 board-certified dermatologists when classifying keratinocyte carcinomas, versus benign seborrheic keratoses, and malignant melanomas, versus benign nevi. ² In orthopaedic subspecialties other than spinal surgery, deep neural networks have also been used to classify upper and lower extremity plain radiographs and identify the presence of fractures with an accuracy of over 90% and 83%, respectively. ³

Evidence of decision support using AI exists in cardiology. A dataset including 40 patient factors such as gender, presence of cardiomyopathy, treatment delivered (angioplasty versus coronary artery bypass graft), and outcome was used to train a neural network. The network was subsequently applied to the data of two cohorts of patients. The first cohort was comprised of patients that survived for at least 5 years following treatment. The second cohort was comprised of patients that had died within five years of treatment. In those patients who died, neural networks were more likely to have proposed an alternative treatment plan to that which was actually carried out. Consequently, the authors suggested that, had a neural network been included as part of the team making the actual treatment plan, the 5-year survival of patients would likely have been better. ⁴