Artificial Intelligence in Medicine

Technical Basis and Clinical Applications

Lei Xing, Maryellen L. Giger, James K. Min

• Publisher: Academic Press
• Year: 2020
• Print ISBN: 9780128212592
• eBook ISBN: 9780128212585
• Topic: Medicine
• Subtopic: Pharmaceutical, Biotechnology & Healthcare Industry

Artificial Intelligence Medicine: Technical Basis and Clinical Applications presents a comprehensive overview of the field, ranging from its history and technical foundations, to specific clinical applications and finally to prospects. Artificial Intelligence (AI) is expanding across all domains at a breakneck speed. Medicine, with the availability of large multidimensional datasets, lends itself to strong potential advancement with the appropriate harnessing of AI.The integration of AI can occur throughout the continuum of medicine: from basic laboratory discovery to clinical application and healthcare delivery. Integrating AI within medicine has been met with both excitement and scepticism. By understanding how AI works, and developing an appreciation for both limitations and strengths, clinicians can harness its computational power to streamline workflow and improve patient care. It also provides the opportunity to improve upon research methodologies beyond what is currently available using traditional statistical approaches. On the other hand, computers scientists and data analysts can provide solutions, but often lack easy access to clinical insight that may help focus their efforts. This book provides vital background knowledge to help bring these two groups together, and to engage in more streamlined dialogue to yield productive collaborative solutions in the field of medicine.- Provides history and overview of artificial intelligence, as narrated by pioneers in the field- Discusses broad and deep background and updates on recent advances in both medicine and artificial intelligence that enabled the application of artificial intelligence- Addresses the ever-expanding application of this novel technology and discusses some of the unique challenges associated with such an approach

Part I

Introduction

Outline

• Chapter 1 Artificial intelligence in medicine: past, present, and future
• Chapter 2 Artificial intelligence in medicine: Technical basis and clinical applications

Chapter 1

Artificial intelligence in medicine: past, present, and future

Efstathios D. Gennatas and Jonathan H. Chen

Abstract

Artificial intelligence is a powerful technology that promises to vastly improve the efficiency and effectiveness of health-care delivery, and usher in the era of precision medicine, transforming our everyday lives. It is helping accelerate basic biomedical research, delivering insights into disease pathophysiology, and guiding new treatment discovery. It is optimizing clinical trials and translational research, bringing us closer to new treatments faster. At a time when the health-care system is under more strain than ever, artificial intelligence promises to revolutionize health-care delivery by capitalizing on the totality of health-related data in order to optimize clinical decision-making for each individual and improve access to health-care for all. To deliver on these promises, we must bring together basic and applied researchers, engineers, and clinicians to address the many outstanding challenges in a timely and responsible manner. It is all of our duty to strive for the safe, fair, and efficient delivery of this technology to all.

Keywords

Artificial intelligence; machine learning; precision medicine; medicine; health care

1.1 Introduction

Artificial intelligence (AI) has been through highs and lows to reclaim its place as one of the most exciting and promising technologies today. It is gaining increasing traction across fields, and the race is on for the widespread delivery of real-world applications that have the potential to transform our daily lives and society as a whole. Medicine is arguably one of the most promising and at the same time challenging fields for AI adoption. AI in medicine aims to optimize clinical decision-making and health-care delivery in general by capitalizing on the increasing volume and availability of health-related data in order to provide the most informed care to each individual. Medical AI applications are still at the early stages of development but are advancing rapidly. This book offers an overview of the ongoing advances in AI across medical subfields. In this introduction chapter, we shall begin with a historical overview of AI and its clinical applications and a set of definitions. We will then consider the promises and challenges of AI in medicine: What do we stand to gain from AI in medicine? What are the challenges we need to address before we can deliver on those promises? The coordinated work of an interdisciplinary team of health-care workers and providers, scientists, and engineers is required to fulfill the potential of AI in medicine in a safe, fair, and efficient way.

1.2 A brief history of artificial intelligence and its applications in medicine

A single formal definition of AI may not exist, but we commonly use the term to refer to a set of approaches “able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages” (“artificial intelligence, n.” OED Online, Oxford University Press; December 2019, www.oed.com/view/Entry/271625 [accessed 13.12.19]). The field originated in the mid-1950s largely within computer science but with important influences from philosophy, mathematics, economics, cognitive science, and neuroscience. Researchers’ early focus was on symbolic reasoning: building high-level representations of problems to mimic, to some extent, human thinking. This paradigm is known as Symbolic AI or Symbolism and is often referred to as “good old fashioned AI”.¹ Early successes were achieved using symbolic reasoning and expert systems, its main type of implementation. These systems rely largely on hard-coded rules designed by human experts to address a defined, circumscribed problem. An example of a very popular expert system widely used today is electronic tax preparation software. The designers of these systems have hardcoded a country’s or state’s entire Tax Law into their software. The program asks users a series of simple questions and follows a long list of if–then statements to calculate how much tax is owed. Such systems can be very effective in specific applications. Their main limitations are as follows:

• • They are labor-intensive: A team of experts needs to manually enter, and subsequently maintain, a long, up-to-date list of rules and their relationships.
• • They are generally only possible when a comprehensive set of stable rules guiding a system is known. This is particularly limiting in medicine, where knowledge uncertainties abound in the setting of constantly evolving systems.^2,3

Early examples of expert systems in medicine included the MYCIN system, designed to recommend appropriate antibiotic treatment for bacterial infections based on user-entered patient symptoms and information,⁴ the causal-associational network CASNET applied to the management of glaucoma,⁵ and INTERNIST-I, a general medicine consultation system.⁶ While exciting, these systems failed to achieve widespread adoption.⁷ Expert systems, in general, largely fell out of fashion in AI research, but successful applications, such as the tax preparation software, remain in use today. Instead, interest grew for purely data-driven learning procedures that eschewed the laborious manual hardcoding of rules. Machine learning (ML), also known as statistical learning, developed out of the fields of statistics and computer science, often in parallel/independently, to precisely allow a machine to learn from data without explicit programing. ML refers to a large and growing collection of algorithms that have proven highly successful in a wide range of applications. Within ML, artificial neural networks (ANN) represent a versatile learning framework created as an attempt, however crude, to mimic the network architecture of the brain. Research in ANNs, starting with seminal work on parallel distributed processing,⁸ gave rise to what was later named connectionism and connectionist AI. Connectionist and symbolic AI has been largely seen as opposing views in AI.⁹ The increasing predominance of ML methods in AI today has led to the two terms often being used interchangeably even though they are not equivalent (Fig. 1.1).

Figure 1.1 The relationship between AI, machine learning, and deep learning. Machine learning refers to a large collection of algorithms and is the main approach used in AI today. Deep learning refers to a specific class of algorithms within machine learning, that are particularly effective at handling “unstructured data”: images, text, etc. AI, Artificial intelligence.

1.3 How intelligent is artificial intelligence?

Currently popular forms of AI/ML algorithms largely operate on a single circumscribed task at a time. For example, a model can be trained to estimate cardiovascular disease risk from demographic and clinical examination data, a different model can be trained to diagnose heart disease from electrocardiograms, yet another one could be trained on cardiac MRIs to select from a list of possible diagnoses. This task-focused prediction is called weak or narrow AI. In contrast, artificial general intelligence (AGI), also known as hard AI, is defined as an AI system that is able to perform any number of intelligent tasks. This remains the ultimate goal for many AI researchers, but while it has been promised or predicted multiple times already, its realization remains out of immediate reach by most estimates. There is currently no way to train a “cardiologist AI,” or a “general medicine AI.” Researchers are focusing instead on augmented intelligence, a paradigm that aims to use AI to assist humans in tackling difficult and important tasks. This is largely where current AI approaches fit in medicine: not as a technology to replace clinicians but as a powerful tool that can process vast amounts of information and to assist clinicians in making decisions, while possibly also automating some simpler tasks.

Many argue that existing machine learning algorithms (weak AI) do little more than a type of “curve fitting” or “pattern recognition” on multidimensional data¹⁰ and are, therefore, not worthy of the term “artificial intelligence,” which should be reserved for a system that possesses higher level abilities, if not general intelligence. Regardless of individual views on the matter, the term AI is widely used and recognized, and the important distinction between weak or narrow and hard AI should be clear to the reader. At the same time, there is increasing interest in bridging the gap between symbolism and connectionism. Such an approach may be the key in paving the way toward AGI. This work can also improve the interpretability, intelligibility, or “explainability” of AI and, therefore, boost its trustworthiness in critical applications such as medicine.

1.4 Artificial intelligence, machine learning, and precision medicine

Advances in ML algorithms along with increases in computational power allow biomedical and clinical researchers to easily analyze large and complex datasets. AI’s benefits in medicine extend across the spectrum of basic biomedical research to translational research and clinical practice. ML in basic research is used to extract insights on disease pathophysiology and guide new treatment discovery. Currently, the majority of applications are in basic research, while clinical applications are slowly being developed and tested.

Precision medicine, sometimes called personalized or individualized medicine, is defined by the NIH as “an emerging approach for disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle for each person” (https://ghr.nlm.nih.gov/primer/precisionmedicine/definition). This approach recognizes that each person may have a unique (1) risk of developing a disease, (2) presentation when they develop disease, and (3) response to treatment and progression of the disease. The central premise of precision medicine is, therefore, to treat individuals, not diseases. This requires the integration of all available health-related data sources to offer individualized estimates of disease risk, prevention strategies, and treatment planning.

1.5 Algorithms and models

ML includes a broad range of methods that can address many types of tasks, Hastie et al.¹¹ provide a comprehensive overview of machine learning methods, and Koller et al.¹² offer a more accessible introduction...

Table of contents

• Cover image
• Title page
• Table of Contents
• Copyright
• Dedication
• List of contributors
• Foreword
• Preface
• Acknowledgments
• Part I: Introduction
• Part II: Technical basis
• Part III: Clinical applications
• Part IV: Future outlook
• Index