Intelligent Systems In Medicine And Health The

Role Of Ai Trevor A Cohen download

https://ebookbell.com/product/intelligent-systems-in-medicine-
and-health-the-role-of-ai-trevor-a-cohen-47214010

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

Hybrid Intelligent Systems In Control Pattern Recognition And Medicine

1st Ed 2020 Oscar Castillo

https://ebookbell.com/product/hybrid-intelligent-systems-in-control-
pattern-recognition-and-medicine-1st-ed-2020-oscar-castillo-10801372

Intelligent And Adaptive Systems In Medicine 1st Edition Olivier C L

Haas

https://ebookbell.com/product/intelligent-and-adaptive-systems-in-
medicine-1st-edition-olivier-c-l-haas-1383708

Recommender Systems For Medicine And Music Studies In Computational

Intelligence 946 1st Ed 2021 Zbigniew W Ras Editor

https://ebookbell.com/product/recommender-systems-for-medicine-and-
music-studies-in-computational-intelligence-946-1st-ed-2021-zbigniew-
w-ras-editor-36373750

Intelligent Systems In Digital Transformation Theory And Applications

Cengiz Kahraman

https://ebookbell.com/product/intelligent-systems-in-digital-
transformation-theory-and-applications-cengiz-kahraman-47288488
Intelligent Systems In Production Engineering And Maintenance Iii Anna
Burduk

https://ebookbell.com/product/intelligent-systems-in-production-
engineering-and-maintenance-iii-anna-burduk-56891888

Intelligent Systems In Technical And Medical Diagnostics 1st Edition L

Travmassuys Auth

https://ebookbell.com/product/intelligent-systems-in-technical-and-
medical-diagnostics-1st-edition-l-travmassuys-auth-4405032

Intelligent Systems In Science And Information 2014 Extended And

Selected Results From The Science And Information Conference 2014 1st
Edition Kohei Arai

https://ebookbell.com/product/intelligent-systems-in-science-and-
information-2014-extended-and-selected-results-from-the-science-and-
information-conference-2014-1st-edition-kohei-arai-4974004

Intelligent Systems In Cybernetics And Automation Theory Proceedings

Of The 4th Computer Science Online Conference 2015 Csoc2015 Vol 2
Intelligent Systems In Cybernetics And Automation Theory 1st Edition
Radek Silhavy
https://ebookbell.com/product/intelligent-systems-in-cybernetics-and-
automation-theory-proceedings-of-the-4th-computer-science-online-
conference-2015-csoc2015-vol-2-intelligent-systems-in-cybernetics-and-
automation-theory-1st-edition-radek-silhavy-5141380

Intelligent Systems In Production Engineering And Maintenance Ispem

2017 Proceedings Of The First International Conference On Intelligent
Systems In Production Engineering And Maintenance Ispem 2017 Burduk

https://ebookbell.com/product/intelligent-systems-in-production-
engineering-and-maintenance-ispem-2017-proceedings-of-the-first-
international-conference-on-intelligent-systems-in-production-
engineering-and-maintenance-ispem-2017-burduk-6753418
Cognitive Informatics in Biomedicine and Healthcare

Trevor A. Cohen
Vimla L. Patel
Edward H. Shortliffe Editors

Intelligent
Systems in
Medicine and
Health
The Role of AI
Cognitive Informatics in Biomedicine
and Healthcare
Series Editor
Vimla L. Patel, Ctr Cognitive Studies in Med & PH
New York Academy of Med, Suite 454, New York, NY, USA
Enormous advances in information technology have permeated essentially all facets
of life. Although these technologies are transforming the workplace as well as
leisure time, formidable challenges remain in fostering tools that enhance
productivity, are sensitive to work practices, and are intuitive to learn and to use
effectively. Informatics is a discipline concerned with applied and basic science of
information, the practices involved in information processing, and the engineering
of information systems.
Cognitive Informatics (CI), a term that has been adopted and applied particularly
in the fields of biomedicine and health care, is the multidisciplinary study of cogni-
tion, information, and computational sciences. It investigates all facets of computer
applications in biomedicine and health care, including system design and computer-
mediated intelligent action. The basic scientific discipline of CI is strongly grounded
in methods and theories derived from cognitive science. The discipline provides a
framework for the analysis and modeling of complex human performance in tech-
nology-mediated settings and contributes to the design and development of better
information systems for biomedicine and health care.
Despite the significant growth of this discipline, there have been few systematic
published volumes for reference or instruction, intended for working professionals,
scientists, or graduate students in cognitive science and biomedical informatics,
beyond those published in this series. Although information technologies are now in
widespread use globally for promoting increased self-reliance in patients, there is
often a disparity between the scientific and technological knowledge underlying
healthcare practices and the lay beliefs, mental models, and cognitive representa-
tions of illness and disease. The topics covered in this book series address the key
research gaps in biomedical informatics related to the applicability of theories,
models, and evaluation frameworks of HCI and human factors as they apply to clini-
cians as well as to the lay public.
Trevor A. Cohen • Vimla L. Patel
Edward H. Shortliffe
Editors

Intelligent Systems
in Medicine and Health
The Role of AI
Editors
Trevor A. Cohen Vimla L. Patel
University of Washington New York Academy of Medicine
Seattle, WA, USA New York, NY, USA

Edward H. Shortliffe
Columbia University
New York, NY, USA

ISSN 2662-7280     ISSN 2662-7299 (electronic)

Cognitive Informatics in Biomedicine and Healthcare
ISBN 978-3-031-09107-0    ISBN 978-3-031-09108-7 (eBook)
https://doi.org/10.1007/978-3-031-09108-7

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
We wish to dedicate this volume to three
giants of science who greatly inspired and
influenced us while helping to lay the
groundwork for artificial intelligence in
medicine, even though each of them made
their primary contributions in other fields.
Furthermore, each has been a personal
friend of at least one of us and we look back
fondly on their humanity and their
contributions.
Herbert Simon (Carnegie Mellon University)
won the Nobel Prize in Economics (1978) for
his pioneering research into the decision-­
making process within economic
organizations. He is also known as an early
innovator in the field of artificial intelligence,
whose long partnership with Allen Newell
led to early work on the Logic Theory
Machine and General Problem Solver. As a
psychologist who studied human and
machine problem solving, he pioneered
notions of bounded rationality and
satisficing.
Joshua Lederberg (Stanford and Rockefeller
Universities) won the Nobel Prize in
Physiology or Medicine (1958) for his
discoveries of genetic transfer in bacteria. A
computer scientist as well as a geneticist, he
was instrumental in devising the notion of
capturing expert knowledge in computers.
His Dendral Project led to a body of work
that dealt with the interpretation of mass
spectral data to identify organic compounds,
and he created the first national resource to
support research on artificial intelligence in
medicine (SUMEX-AIM).
vi 

Walter Kintsch (University of Colorado

Boulder) is a psychologist and cognitive
scientist who used both experimental and
computational modeling techniques to study
how people understand language. He
reformulated discourse comprehension as a
constraint satisfaction process. His work on
latent semantic analysis demonstrated how
machine learning could be used to construct
a high-dimensional semantic space. His
writing has influenced research on natural
language understanding as well as the
cognitive informatics focus reflected in
this volume.
T.C., V.L.P., E.H.S.—March 2022

Photograph of Herbert Simon provided

courtesy of Carnegie Mellon University.
Photograph of Joshua Lederberg provided
courtesy of EH Shortliffe.
Photograph of Walter Kintsch provided by
Professor Kintsch.
Foreword

Delivering quality, affordable health care to large populations is one of humankind’s

greatest challenges.
There are so many facets to health care, however, that talking about “the prob-
lem” is a gross understatement. George Polya’s heuristic (from Descartes) of divid-
ing problems into subproblems1 remains one of the best tools in our cognitive
toolkit—human or artificial intelligence (AI).2 This book explores many of those
subproblems from the perspective of using computing—AI in particular—to
meet them.
The idea of automating solutions to specific human problems is rooted in Greek
mythology but actual implementations of automated problem solving with mechan-
ical or electronic computers were initially restricted to single-purpose applications.
For example, wartime uses included decoding encrypted messages and calculating
trajectories of artillery shells; business uses included payroll accounting, informa-
tion retrieval, and process automation. Medical uses included keeping medical
records, recording physiological data, and calculating drug doses.
Modern medicine is grounded in data from observations as well as measure-
ments. But the machines running initial applications of computers to medicine were
so restricted in size and scope that demonstrations beyond single-purpose problems
were severely limited.
Around 1950 Turing introduced the audacious idea that computers could be
intelligent and could be used on problems that required reasoning, understanding,

1
Polya G. How to Solve It: A New Aspect of Mathematical Method. Princeton, NJ: Princeton
University Press, 2014 (originally published 1945).
2
Two memorable examples of identifying key subproblems are cited in McCullough’s book on
building the Panama Canal [McCullough, D. The Path between the Seas: The Creation of the
Panama Canal, 1870-1914. New York: Simon & Schuster, 1978]. First, after several years of fail-
ure, a new engineer put in charge of the project rethought the steps of digging a huge trench and
recognized that the most rate-limiting step was getting rid of the dirt. Second, the same engineer
recognized another rate-limiting step was keeping workers healthy. Neither was “the problem” per
se; solving both subproblems made the difference.

vii
viii Foreword

and learning—to work smarter, not harder. Acquiring measurements is one thing,
interpreting the data is another. Importantly, early AI research introduced a concep-
tual framework that includes two critical differences from mathematical computa-
tion: symbolic reasoning and heuristic reasoning. Reasoning with symbols, as
opposed to numeric quantities, carries with it the essential link human beings made
when they invented language between words and the world, i.e., semantics. Heuristic
reasoning is even older than humankind, being bound up with the Darwinian notion
of survival. A decision to fight or flee does not leave time to consider the conse-
quences of all possible options,3 and most decisions in everyday life have no algo-
rithm that guarantees a correct answer in a short time (or often not in any amount
of time).
The importance of human health quickly attracted early pioneers to the possibil-
ity of using computers to assist physicians, nurses, pharmacists, and other health-
care professionals.
The initial applications of AI limited their scope so they could be dealt with suc-
cessfully in single demonstration projects on the small computers of the day.
Cognitive psychologists recognized that clinicians reason about the available data
in medicine just as they do in other fields like chess. The process of diagnostic rea-
soning became the focus of work on expert systems, with early programs becoming
convincing examples that AI could replicate human reasoning in this area on nar-
rowly defined problems.
An important part of this demonstration was with Bayesian programs that used
statistics to reason in a mathematically rigorous way for clinical decision support.
However, those initial applications of Bayes’ Theorem to medicine were also lim-
ited until means were found to reduce their complexity with heuristic reasoning.
Another important set of demonstrations encoded knowledge accumulated by
human experts, along the lines suggested by cognitive science. One of the short-
comings of the expert systems approach, however, is the time and effort it takes to
acquire the knowledge from experts in the first place, and to maintain a large knowl-
edge base thereafter.
In practice, medical professionals are faced with several inconvenient truths,
which further complicate efforts to use computers in health care. The chapters of
this book address many of them. For example:
• The body of knowledge in, and relevant to, medicine is growing rapidly. E.g.,
diagnosis and treatment options for genetic diseases in the last few decades, and
of viral infections in the last few years, have come into the mainstream.
• Complete paradigm shifts in medicine require rethinking whole areas of medi-
cine. E.g., prion diseases have forced a reconceptualization of the mechanisms of
pathogen transmission.

3
Herb Simon, one of the founding fathers of AI, received a Nobel Prize in Economics in 1978 for
elucidation of this concept in the world of decision making. His term for it was “Bounded
Rationality.” See https://www.cmu.edu/simon/what-is-simon/herbert-a-simon.html (accessed
August 11, 2022).
Foreword ix

• Medical knowledge is incomplete and there are no good treatment options, and
sometimes no good diagnostic criteria, for some conditions. E.g., Parkinson’s
disease can be managed, but is still not curable after decades of research.
• Information about an individual patient is often erroneous and almost always
incomplete. E.g., false positive and false negative test results are expected for
almost every diagnostic test.
• Patients’ medical problems exist within a larger emotional, social, economic,
and cultural context. E.g., the most effective treatment options may be unafford-
able or unacceptable to an individual.
• Professionals are expected to learn from their own, and others’, experience (both
positive and negative). E.g., continuing to recommend a failed treatment modal-
ity is reason for censure.
• Professionals at the level of recognized specialists are expected to deal with
unique cases for which there are neither case studies nor established diagnostic
or treatment wisdom. E.g., primary care providers refer recalcitrant cases to spe-
cialists for just these reasons.
• Communication between patients and professionals is imperfect. E.g., language
is full of ambiguity, and we all have biases in what we want to hear or fear most.
Collectively, these issues are more than merely “inconvenient,” but are humbling
reminders that “the problem” of providing health care is overwhelming in the large.
They also represent significant barriers to harnessing the presently available power
of computers to actual healthcare delivery. The perspectives offered in this book
summarize current approaches to these issues and highlight work that remains to be
done. As such it is valuable as a textbook for biomedical informatics and as a road-
map for the possibilities of using AI for the benefit of humankind.
The book’s emphasis on reasoning provides a central focus not found in other
collections. The chapters here deal with transforming data about patients, once
acquired, to actionable information and using that information in clinical contexts.
With today’s understanding, AI offers the means to augment human intelligence by
making the accumulated knowledge available, suggesting possible options, and
considering consequences. We are betting that computers can help to overcome
human limitations of imperfect memory, reasoning biases, and sheer physical stam-
ina. We are betting on the power of knowledge over the persistent forces of random
mutations. Most of all we are betting that the synergy of human and computer intel-
ligence will succeed in the noble quest of improving the quality of human lives.

University of Pittsburgh Bruce G. Buchanan

Pittsburgh, PA, USA
August 2022
Preface

The State of AI in Medicine

Recent advances in computing power and the availability of large amounts of train-
ing data have spurred tremendous advances in the accuracy with which computers
can perform tasks that were once considered the exclusive province of human intel-
ligence. These include perceptual tasks related to medical diagnosis, where deep
neural networks have attained expert-level performance for the diagnosis of diabetic
retinopathy and the identification of biopsy-proven cases of skin cancer from der-
matological images. These accomplishments reflect an increase in activity in
Artificial Intelligence (AI), both in academia and industry. According to the 2021 AI
Index report assembled and published at Stanford University, there was a close to
twelvefold increase in the total number of annual peer-reviewed AI publications
between 2000 and 2019, and a close to fivefold increase in annual private invest-
ment in funded AI companies between 2015 and 2020, with commercial applica-
tions of AI technologies such as speech-based digital assistants and personalized
advertising and newsfeeds by now woven deeply into the threads of our everyday
lives. These broad developments throughout society have led to a resurgence of
public interest in the role of AI in medicine1 (AIM), reviving long-standing debates
about the nature of intelligence, the relative value of data-driven predictive models
and human decision makers, and the potential for technology to enhance patient
safety and to disseminate expertise broadly.
Consequently AIM is in the news, with frequent and often thoughtful accounts of
the ways in which it might influence—and hopefully improve—the practice of med-
icine appearing in high visibility media venues such as The New York Times [1, 2],
The Atlantic [3], The New Yorker [4], The Economist [5], and others [6, 7]. As bio-
medical informaticians with long-standing interests in AIM, we are encouraged to
see this level of attention and investment in the area. However, we are also aware

1
We consider the scope of AIM to include public health and clinically driven basic science research,
as well as clinical practice.

xi
xii Preface

that this is by no means the first time that the promise of AIM has emerged as a
focus of media attention. For example, a 1977 New York Times article [8] describes
the MYCIN system (discussed in Chap. 2), noting the ability of this system to make
medical diagnoses, to request missing information, and to explain how it reaches
conclusions. The article also mentions the extent of government funding for AI
research at the time at $5 million a year, which when adjusted for inflation would be
around $23 million annually. The question arises as to what has changed between
then and now, and how these changes might affect the ongoing and future prospects
for AI technologies to improve medical care. Some answers to this question are
evident in this 1977 article, which is critical of the potential of the field to accom-
plish its goals of championship-level chess performance and machine translation
(“neither has been accomplished successfully, and neither is likely to be any time
soon”) and face recognition (“they cannot begin to distinguish one face from another
as babies can”). Today all of these tasks are well within the capabilities of contem-
porary AI systems, which is one indication of the methodological advances that
have been made over the intervening four decades. However, translating methods
with strong performance on tightly constrained tasks to applications with positive
impact in health care is not (and has never been) a straightforward endeavor, on
account of the inherent complexities of the healthcare system and the prominent
role of uncertain and temporarily unavailable information in medical decision mak-
ing, among other factors. While AI technologies do have the potential to transform
the practice of medicine, computer programs demonstrating expert-level perfor-
mance in diagnostic tasks have existed for decades, but significant challenges to
realizing the potential value of AI in health care—such as how such AI systems
might best be integrated into clinical practice—remain unresolved.
It is our view that the discipline of Cognitive Informatics (CI) [9], which brings
the perspective of the cognitive sciences to the study of medical decision making by
human beings and machines, is uniquely positioned to address many of these chal-
lenges. Through its roots in the study of medical reasoning, CI provides a sound
scientific basis from which to consider the relationship between current technolo-
gies and human intelligence. CI has extended its area of inquiry to include both
human-computer interaction and the study of the effects of technology on the flow
of work and information in clinical settings [10]. Accordingly, CI can inform the
integration of AI systems into clinical practice. In recent years, patient safety has
emerged as a research focus of the CI community, providing new insights into the
ways in which technology might mitigate or, despite best intentions, facilitate medi-
cal error. Consequently, a volume describing approaches to AIM from a CI perspec-
tive seemed like an excellent fit for the Springer Nature Cognitive Informatics book
series led by one of us (VLP). We fondly recall that this volume is a project that we
first discussed several years ago over a bottle of Merlot in proximity to a conference
in Washington, DC.
However, as our discussions developed over the course of subsequent meetings,
it became apparent that there was a need for a more comprehensive account of the
field. As educators, we considered the knowledge and skills that future researchers
and practitioners in the field might need in order to realize the transformative
Preface xiii

potential of AI methods for the practice of health care. We were aware of books on
the subject, such as cardiologist Eric Topol’s cogent account of the implications of
contemporary machine learning approaches for the practice of medicine [11], and
the development of pediatric cardiologist Anthony C. Chang’s excellent clinician-­
oriented introduction to AI methods with a focus on their practical application to
problems in health care and medicine [12].2 Books drawing together the perspec-
tives of multiple authors were also available, including a compendium of chapters
in which authors focus on their research interests within the field [13], and an
account of the organizational implications of “big data” and predictive models that
can be derived from such large collections of information [14]. However, none of
the books we encountered was developed with the focused intention to provide a
basis for curricular development in AIM, and we were hearing an increasing demand
for undergraduate, graduate, and postgraduate education from our own trainees—
students and aspiring physician-informaticians.
Consequently, we pivoted from our original goal of a volume primarily con-
cerned with highlighting the role of CI in AIM, to the goal of developing the first
comprehensive coauthored textbook in the AIM field, still with a CI emphasis. We
reached out to our friends and colleagues, prominent researchers with deep exper-
tise in the application of AI methods to clinical problems, several of whom have
been engaged with AIM since the inception of the field. This was a deliberate deci-
sion on our part, as we felt that in addition to lacking a cognitive perspective (despite
the emergence of the term “cognitive computing” as a catch-all for AI methods),
much work we encountered was presented without apparent consideration of the
history of the field. Our concern with this disconnect was not only a matter of the
academic impropriety of failure to acknowledge prior work. We were also con-
cerned that work conducted from this perspective would not be informed by the
many lessons learned from decades of work and careful consideration of the issues
involved in implementing AI at the point of care. Thus, in developing our ideas for
the structure of this volume, and in our selection of chapter authors, we endeavored
to make sure that presentations of current methods and applications were contextu-
alized in relation to the history of the field, and informed by a CI perspective.

I ntroducing Intelligent Systems in Medicine and Health:

The Role of AI

The result of these efforts is the current volume, a comprehensive textbook that
takes stock of the current state of the art of AIM, places current developments in
their historical context, and identifies cognitive and systemic factors that may help
or hinder their ability to improve patient care. It is our intention that a reader of this

2
A trained data scientist, Dr. Chang also contributed a chapter on the future of medical AI and data
science in this volume.
xiv Preface

volume will attain an accurate picture of the strengths and limitations of these
emerging technologies, emphasizing how they relate to the AI systems that pre-
ceded them, to the intelligence of human decision makers in medicine, and to the
needs and expectations of those who use the resulting tools and systems. This will
lay a foundation for an informed discussion of the potential of such technology to
enhance patient care, the obstacles that must be overcome for this to take place, and
the ways in which emerging and as-yet-undeveloped technologies may transform
the practice of health care.
With increasing interest and investment in AIM technologies will come the rec-
ognition of the need for professionals with the prerequisite expertise to see these
technologies through to the point of positive impact. Progress toward this goal will
require both advancement of the state of the science through scholarly research and
measurably successful deployment of AIM systems in clinical practice settings. As
the first comprehensive coauthored textbook in the field of AIM, this volume aims
to define and aggregate the knowledge that researchers and practitioners in the field
will require to advance it. As such, it draws together a range of expert perspectives
to provide a holistic picture of the current state of the field, to identify opportunities
for further research and development, and to provide guidance to inform the suc-
cessful integration of AIM into clinical and public health practice.
We intend to provide a sound basis for a seminar series or a university level
course on AIM. To this end, authors have been made aware of the context of their
chapters within the logical flow of the entire volume. We have sought to assure
coordination among authors to facilitate cross-references between chapters, and to
minimize either coverage gaps or redundancies. Furthermore, all chapters have fol-
lowed the same basic organizational structure, which includes explicit learning
objectives, questions for self-study, and annotated suggestions for further reading.
Chapters have been written for an intended audience of students in biomedical
informatics, AI, machine learning, cognitive engineering, and clinical decision sup-
port. We also offer the book to established researchers and practitioners of these
disciplines, as well as those in medicine, public health, and other health professions,
who would like to learn more about the potential for these emerging technologies to
transform their fields.

Structure and Content

The book is divided into four parts. They are designed to emphasize pertinent con-
cepts rather than technical detail. There are other excellent sources for exploring the
technical details of the topics we introduce. The Introduction provides readers with
an overview of the field. Chapter 1 provides an introduction to the fields of Artificial
Intelligence and Cognitive Informatics and describes how they relate to one another.
Chapter 2 provides a historical perspective, drawing attention to recurring themes,
issues, and approaches that emerged during the course of the development of early
AI systems, most of which remain highly relevant today. Chapter 3 provides an
Preface xv

overview of the landscape of biomedical data and information, to familiarize read-

ers with the resources AIM systems can draw upon for training and evaluation, and
as sources of structured knowledge. This chapter also serves as a bridge to subse-
quent parts in the book, introducing some of the methods, applications, and issues
that will be covered in greater detail later.
The Approaches part covers methods used by AIM systems. Chapter 4 com-
mences this part with a discussion of knowledge-based approaches, spanning from
knowledge modeling efforts, used in expert systems at the inception of AIM as a
field, to contemporary efforts to infuse machine learning models with structured
biomedical knowledge. Chapter 5 considers the idea of AIM from a cognitive per-
spective, beginning with an account of the parallel development of the two fields,
and advancing an argument for the development of complementary human/AI sys-
tems with superior clinical utility when compared with either of these components
alone. Chapter 6 introduces the machine learning approaches that have come to
predominate in the current wave of AIM systems, with the intention of providing
readers with a conceptual understanding of some key algorithms, and issues that
may arise during their application to modeling medical data. Chapter 7 considers
Natural Language Processing (NLP), which has been intertwined with AI at least
since Turing’s proposal of the ability to conduct a passably human conversation as
an observable surrogate for “thinking” [15]. The chapter focuses on biomedical
NLP, giving an account of the main problems to be addressed in this area, and the
methods that predominate at present, from rule-based methods through to deep
learning approaches. Chapter 8 considers approaches to the explanation of deci-
sions recommended by AIM models, with an emphasis on its importance in clinical
settings in which “black box” predictions may not—and arguably should not—be
taken at face value without some justification to earn trust and to permit detection of
predictions made on tenuous grounds. Chapter 9 revisits the issue of language, with
a focus on methods that support human-like conversational interactions with the
goal of supporting health care.
Having introduced the fundamental methods of the field, the book focuses in the
third part on Applications. Chapter 10 describes those in which AIM methods are
applied to support decision making in clinical care, through integration with exist-
ing platforms and workflows. Chapter 11 focuses on the prediction of clinical out-
comes, such as near-term readmission, with an emphasis on dynamic models of
variables and outcomes that change over time. Chapter 12 describes what is argu-
ably the most successful area of application of contemporary AIM methods to
date—the interpretation of medical images. Chapter 13 shifts from a focus on meth-
ods to support the care of individual patients to the population level, with a discus-
sion of the emerging role of AI in the field of public health. Chapter 14 then describes
AI applications at the molecular level: methods that yield clinically actionable
insights from -omics data. Chapter 15 discusses AIM applications that are of admin-
istrative importance to a healthcare system, such as approaches to optimize the use
of resources. This chapter also includes a discussion of the practical concerns that
may arise when attempting to implement AI solutions in the context of an opera-
tional healthcare system. Finally, Chap. 16 focuses on AI applications in medical
xvi Preface

education, including the roles of cognitive and learning sciences in informing how
clinicians-in-training should be educated about AIM, and how AI might support the
education of clinicians in their chosen fields.
The final part in the volume considers the road ahead for AIM. It addresses
issues that are likely to be of importance for the successful progression of the field,
including two potential stumbling blocks: inadequate evaluation, and failure to con-
sider the ethical issues that may accompany the deployment of AI systems in health-
care settings. Accordingly, the first chapter in this part, Chap. 17, is focused on
evaluation of AIM systems, including enabling capabilities such as usability and the
need to move beyond “in situ” evaluations of accuracy toward demonstrations of
acceptance and clinical utility in the natural world. Chapter 18 concerns the need for
a robust ethical framework to address issues proactively such as algorithmic bias,
and exacerbation of healthcare inequities due to limited portability of methods and
algorithms. Chapter 19 projects from the trajectory of current trends to anticipate
the future of AI in medicine with an emphasis on data science, and how broader
deployment of AI systems may affect the practice of medicine. Finally, Chap. 20
provides a summary and synthesis of the volume, including the editors’ perspectives
on the prospects and challenges for the field. The final part is then followed by a
detailed glossary that provides definitions of all terms displayed in bold throughout
the body of the book (with an indication of the chapter(s) in which each term was
used). The book closes with a subject index for the entire volume.

Guide to Use of This Book

This book is written as a textbook, such that it can be used in formal courses or
seminars. For this purpose, we would anticipate curricular design to follow the over-
all structure of the book, with a logical progression from introduction through
approaches to applications and projections for the future. For example, this structure
could support an undergraduate or graduate level course in a Computer Science,
Biomedical Informatics, or Cognitive Engineering program that aims to provide
students with a comprehensive survey of current applications and concerns in the
field. At the graduate level, this could be coupled to a student-led research project.
Alternatively, one might imagine an MS level course that aims to provide clinical
practitioners seeking additional training in clinical informatics with the knowledge
they will need to be informed users of AIM systems, in which case content could be
drawn from the book selectively with an emphasis on introductory content, and
clinical applications and issues that relate to them directly (Chaps. 1–3, 10–12, 15,
and 17–19). Of course, the book may be used for self-study and reference, and read-
ers may wish to explore particular topics in greater detail—starting with a particular
chapter (say, machine learning methods) and then exploring the cross-references in
this chapter to find out more about how this topic features in the context of particu-
lar applications, or issues that are anticipated to emerge as the field progresses.
Preface xvii

This is an exciting time to be working in the field of AIM, and an ideal time to
enter it. There is increasing support for AIM work, both through federal funding
initiatives such as the NIH-wide Bridge to Artificial Intelligence program3 and in
light of an acceleration in private investment in digital health technologies. Such
support has been stimulated in part by the field’s demonstrated utility and accep-
tance as a way to diagnose disease, to deliver care, and to support public health
efforts during the COVID-19 pandemic. On account of the pervasiveness of AI tech-
nologies across industries outside of health care, skepticism about the ability of
these technologies to deliver meaningful improvements is balanced by enthusiasm
for their potential to improve the practice of medicine. It is our goal that readers of
this volume will emerge equipped with the knowledge needed to realize this poten-
tial and to proceed to lead the advancement of health care through AIM.

References
1. Metz C. A.I. shows promise assisting physicians. The New York Times. 2019.
[cited 2021 Apr 15]. https://www.nytimes.com/2019/02/11/health/artificial-­
intelligence-­medical-­diagnosis.html.
2. O’Connor A. How artificial intelligence could transform medicine. The
New York Times. 2019. [cited 2021 Apr 15]. https://www.nytimes.com/
2019/03/11/well/live/how-­a rtificial-­i ntelligence-­c ould-­t ransform-­
medicine.html.
3. Cohn J. The robot will see you now. The Atlantic. 2013. [cited 2021 Jun 14].
https://www.theatlantic.com/magazine/archive/2013/03/the-­robot-­will-­see-­
you-­now/309216/.
4. Mukherjee S. A.I. versus M.D.. The New Yorker. [cited 2021 Apr 15]. https://
www.newyorker.com/magazine/2017/04/03/ai-­versus-­md.
5. Artificial intelligence will improve medical treatments. The Economist. 2018.
[cited 2021 Jun 16]. https://www.economist.com/science-­and-­technology/
2018/06/07/artificial-­intelligence-­will-­improve-­medical-­treatments.
6. Aaronovitch D. DeepMind, artificial intelligence and the future of the
NHS. [cited 2021 Jun 16]. https://www.thetimes.co.uk/article/deepmind-­
artificial-­intelligence-­and-­the-­future-­of-­the-­nhs-­r8c28v3j6.
7. Artificial intelligence has come to medicine. Are patients being put at risk?. Los
Angeles Times. 2020. [cited 2021 Jun 16]. https://www.latimes.com/business/
story/2020-­01-­03/artificial-­intelligence-­healthcare.
8. Experts argue whether computers could reason, and if they should. The
New York Times. [cited 2021 Apr 15]. https://www.nytimes.com/1977/05/08/
archives/experts-­a rgue-­w hether-­c omputers-­c ould-­r eason-­a nd-­i f-­t hey-­
should.html.
9. Patel VL, Kaufman DR, Cohen T, editors. Cognitive informatics in health and
biomedicine: case studies on critical care, complexity and errors. London:

3
https://commonfund.nih.gov/bridge2ai (accessed August 11, 2022).
xviii Preface

Springer; 2014. [cited 2021 Jun 16]. (Health Informatics). https://www.springer.

com/gp/book/9781447154891.
10. Patel VL, Kaufman DR. Cognitive science and biomedical informatics. In:
Shortliffe EH, Cimino JJ, editors. Biomedical informatics: computer applica-
tions in health care and biomedicine. 5th ed. New York: Springer; 2021.
p. 133–85.
11. Topol E. Deep medicine. How artificial intelligence can make healthcare human
again. Hachette UK; 2019.
12. Chang AC. Intelligence-based medicine: artificial intelligence and human cog-
nition in clinical medicine and healthcare. Academic Press; 2020.
13. Agah A, editor. Medical applications of artificial intelligence. Boca Raton:
CRC Press; 2013. 526 p.
14. Natarajan P, Frenzel JC, Smaltz DH. Demystifying big data and machine learn-
ing for healthcare. Boca Raton: CRC Press; 2017. 210 p.
15. Turing AM. Computing machinery and intelligence. Mind : a quarterly review
of psychology and philosophy. 1950;LIX:433.

Seattle, WA, USA Trevor A. Cohen

New York, NY, USA  Vimla L. Patel
New York, NY, USA  Edward H. Shortliffe
August 2022
Acknowledgments

While my coeditors are well known for their prescience in anticipating (and influ-
encing) the evolution of AIM, I think it is safe to say that none of us envisioned
developing a textbook together when we first met at the annual retreat of Columbia
University’s Department of Medical Informatics in the Catskills in 2002. At the
time, I had just joined the program as an incoming graduate student, attracted in part
by what I had learned of Ted’s work in medical AI and Vimla’s in medical cognition.
As these topics have remained core components of my subsequent research, it is
especially encouraging to see the upsurge in interest in the field. The expanding
pool of talented graduate students and physician-informaticians excited about the
potential of AIM was a key motivator for our development of this volume, as was
our recognition of the need for a textbook in the field that represented historical and
cognitive perspectives, in addition to recent methodological developments. However,
given the breadth of methods and biomedical applications that by now fall under the
AIM umbrella, it was apparent to us that we would need to draw upon the expertise
of leaders in relevant fields to develop a multiauthored textbook. Our efforts to
weave the perspectives of these authors into a coherent textbook benefited consider-
ably from Ted’s experience as lead editor of a multiauthored textbook in biomedical
informatics (fondly known as the “blue bible” of BMI, and now in its fifth edition).
We modeled both our approach to encouraging the integration of ideas across chap-
ters and the key structural elements of each chapter on the example of that book. We
were aided in this endeavor by our authors, who were highly responsive to our sug-
gestions for points of reference between chapters, as well as our recommendations
to reduce overlap. We especially appreciated our correspondence with authors dur-
ing the editing process, which broadened our own perspectives on AIM, and influ-
enced the themes we focused on when developing the final chapter (Reflections and
Projections). We owe additional gratitude to Grant Weston, Executive Editor of
Springer’s Medicine and Life Sciences division, for his steadfast support and guid-
ance throughout the development of this volume. We would also like to acknowl-
edge our Production Editor Rakesh Jotheeswaran, Project Manager Hashwini
Vytheswaran, and Editorial Assistant Leo Johnson for their assistance in keeping
the project on track. Special thanks also go to Bruce Buchanan (an AI luminary, key

xix
xx Acknowledgments

mentor for Ted’s work in the 1970s, and his long-term collaborator and colleague)
for his willingness to craft the foreword for this volume. The development of this
volume coincided with a global pandemic. Many involved in the project were
affected by this in a professional capacity through their practice of medicine, their
role in an informatics response at their institutions, or the position of their profes-
sional home within a healthcare system under siege. We trust the development of
their chapters provided a welcome respite for these authors, as ours did for us, and
hope that readers of this volume will be inspired to develop AIM solutions that
equip us to anticipate and manage global health crises more effectively in the future.

Seattle, WA, USA Trevor A. Cohen

August 2022
Contents

Part I Introduction
1 Introducing AI in Medicine��������������������������������������������������������������������    3
Trevor A. Cohen, Vimla L. Patel, and Edward H. Shortliffe
2  in Medicine: Some Pertinent History�����������������������������������������������   21
AI
Edward H. Shortliffe and Nigam H. Shah
3 
Data and Computation: A Contemporary Landscape ������������������������   51
Ida Sim and Marina Sirota
Part II Approaches
4 
Knowledge-Based Systems in Medicine ������������������������������������������������   75
Peter Szolovits and Emily Alsentzer
5 
Clinical Cognition and AI: From Emulation to Symbiosis������������������ 109
Vimla L. Patel and Trevor A. Cohen
6 Machine Learning Systems �������������������������������������������������������������������� 135
Devika Subramanian and Trevor A. Cohen
7 Natural Language Processing ���������������������������������������������������������������� 213
Hua Xu and Kirk Roberts
8 Explainability in Medical AI ������������������������������������������������������������������ 235
Ron C. Li, Naveen Muthu, Tina Hernandez-Boussard, Dev Dash,
and Nigam H. Shah
9 
Intelligent Agents and Dialog Systems �������������������������������������������������� 257
Timothy Bickmore and Byron Wallace
Part III Applications
10 
Integration of AI for Clinical Decision Support������������������������������������ 285
Shyam Visweswaran, Andrew J. King, and Gregory F. Cooper

xxi
xxii Contents

11 Predicting Medical Outcomes ���������������������������������������������������������������� 309

Riccardo Bellazzi, Arianna Dagliati, and Giovanna Nicora
12 Interpreting Medical Images������������������������������������������������������������������ 343
Zongwei Zhou, Michael B. Gotway, and Jianming Liang
13 Public Health Applications���������������������������������������������������������������������� 373
David L. Buckeridge
14  in Translational Bioinformatics and Precision Medicine���������������� 391
AI
Thanh M. Nguyen and Jake Y. Chen
15 Health Systems Management������������������������������������������������������������������ 431
Adam B. Wilcox and Bethene D. Britt
16 
Intelligent Systems in Learning and Education������������������������������������ 449
Vimla L. Patel and Parvati Dev
Part IV The Future of AI in Medicine: Prospects and Challenges
17 
Framework for the Evaluation of Clinical AI Systems������������������������ 479
Edward H. Shortliffe, Martìn-Josè Sepùlveda, and Vimla L. Patel
18 
Ethical and Policy Issues ������������������������������������������������������������������������ 505
Diane M. Korngiebel, Anthony Solomonides,
and Kenneth W. Goodman
19 Anticipating the Future of Artificial Intelligence
in Medicine and Health Care: A Clinical Data Science
Perspective������������������������������������������������������������������������������������������������� 527
Anthony C. Chang
20 Reflections and Projections �������������������������������������������������������������������� 539
Trevor A. Cohen, Vimla L. Patel, and Edward H. Shortliffe
Terms and Definitions�������������������������������������������������������������������������������������� 553
Index������������������������������������������������������������������������������������������������������������������ 585
Contributors

Emily Alsentzer Massachusetts Institute of Technology, Cambridge, MA, USA

Riccardo Bellazzi Department of Electrical, Computer and Biomedical
Engineering, University of Pavia, Pavia, Italy
Laboratory of Informatics and Systems Engineering for Clinical Research, Istituti
Clinici Scientifici Maugeri, Pavia, Italy
Timothy Bickmore Khoury College of Computer Sciences, Northeastern
University, Boston, MA, USA
Bethene D. Britt UW Medicine, University of Washington, Seattle, WA, USA
David L. Buckeridge McGill University, Montreal, QC, Canada
Anthony C. Chang Medical Intelligence and Innovation Institute (MI3), Children’s
Health of Orange County, Orange, CA, USA
Jake Y. Chen Informatics Institute, University of Alabama at Birmingham,
Birmingham, AL, USA
Trevor A. Cohen University of Washington, Seattle, WA, USA
Gregory F. Cooper Department of Biomedical Informatics, University of
Pittsburgh, Pittsburgh, PA, USA
Arianna Dagliati Department of Electrical, Computer and Biomedical
Engineering, University of Pavia, Pavia, Italy
Dev Dash Stanford University School of Medicine, Stanford, CA, USA
Parvati Dev SimTabs LLC, Los Altos Hills, CA, USA
Kenneth W. Goodman Institute for Bioethics and Health Policy, University of
Miami, Miami, FL, USA
Michael B. Gotway Mayo Clinic, Scottsdale, AZ, USA

xxiii
xxiv Contributors

Tina Hernandez-Boussard Stanford University School of Medicine,

Stanford, CA, USA
Andrew J. King Department of Critical Care Medicine, University of Pittsburgh,
Pittsburgh, PA, USA
Diane M. Korngiebel Department of Biomedical Informatics and Medical
Education, University of Washington, Seattle, WA, USA
Ron C. Li Stanford University School of Medicine, Stanford, CA, USA
Jianming Liang Arizona State University, Phoenix, AZ, USA
Naveen Muthu University of Pennsylvania School of Medicine,
Philadelphia, PA, USA
Children’s Hospital of Philadelphia, Philadelphia, PA, USA
Thanh M. Nguyen Informatics Institute, University of Alabama at Birmingham,
Birmingham, AL, USA
Giovanna Nicora Department of Electrical, Computer and Biomedical
Engineering, University of Pavia, Pavia, Italy
Vimla L. Patel The New York Academy of Medicine, New York, NY, USA
Kirk Roberts The University of Texas Health Science Center at Houston,
Houston, TX, USA
Martìn-Josè Sepùlveda Florida International University, Miami, FL, USA
Nigam H. Shah Stanford University School of Medicine, Stanford, CA, USA
Edward H. Shortliffe Vagelos College of Physicians and Surgeons, Columbia
University, New York, NY, USA
Ida Sim University of California San Francisco, San Francisco, CA, USA
Marina Sirota University of California San Francisco, San Francisco, CA, USA
Anthony Solomonides Research Institute, NorthShore University HealthSystem,
Evanston, IL, USA
Devika Subramanian Rice University, Houston, TX, USA
Peter Szolovits Massachusetts Institute of Technology, Cambridge, MA, USA
Shyam Visweswaran Department of Biomedical Informatics, University of
Pittsburgh, Pittsburgh, PA, USA
Byron Wallace Khoury College of Computer Sciences, Northeastern University,
Boston, MA, USA
Adam B. Wilcox Washington University in St. Louis School of Medicine, St.
Louis, MO, USA
Contributors xxv

Hua Xu The University of Texas Health Science Center at Houston,

Houston, TX, USA
Zongwei Zhou Johns Hopkins University, Baltimore, MD, USA
 Part I
Introduction
Chapter 1
Introducing AI in Medicine

Trevor A. Cohen, Vimla L. Patel, and Edward H. Shortliffe

After reading this chapter, you should know the answers to these questions:
• How does one define artificial intelligence (AI)? What are some ways in
which AI has been applied to the practice of medicine and to health care more
broadly?
• How does one define cognitive informatics (CI)? How can the CI perspective
inform the development, evaluation and implementation of AI-based tools to
support clinical decision making?
• What are some factors that have driven the current wave of interest in AI
methods?
• How can one compare and contrast knowledge-based systems with machine
learning models? What are some of the relative advantages and disadvantages of
these approaches?
• Considering the current state of progress, where is research and development
most urgently needed in the field and why?

T. A. Cohen (*)
University of Washington, Seattle, WA, USA
e-mail: cohenta@uw.edu
V. L. Patel
New York Academy of Medicine, New York, NY, USA
E. H. Shortliffe
Columbia University, New York, NY, USA

© The Author(s), under exclusive license to Springer Nature 3

Switzerland AG 2022
T. A. Cohen et al. (eds.), Intelligent Systems in Medicine and Health, Cognitive
Informatics in Biomedicine and Healthcare,
https://doi.org/10.1007/978-3-031-09108-7_1
4 T. A. Cohen et al.

The Rise of AIM

Knowledge-Based Systems

The term “artificial intelligence” (AI) can first be found in a proposal for a confer-
ence that took place at Dartmouth College in 1956, which was written by John
McCarthy and his colleagues [1]. The research to be conducted in this two-month
conference was built upon the “conjecture that every aspect of learning or any
other feature of intelligence can in principle be so precisely described that a
machine can be made to simulate it.” This conference is considered a seminal event
in AI, and was followed by a steady growth of interest in the field that is reflected
by the frequency with which the term ‘artificial intelligence’ appeared in books of
this era (Fig. 1.1). There was a first peak of activity in the mid-1980s that followed
a period of rapid progress in the development of knowledge-based expert systems,
systems that were developed by eliciting knowledge from human experts and ren-
dering this content in computer-interpretable form. Diagnostic reasoning in medi-
cine was one of the first focus areas for the development of such systems, providing
proof that AI methods could approach human performance in tasks demanding a
command of a rich base of knowledge [3]. This shows that medical decision mak-
ing has long been considered a paradigmatic example of intelligent human behav-
ior, and has been a focus of—and has had an influence on—AI research for decades.
The historical trend in term usage in Fig. 1.1 also reveals a dip in enthusiasm and
in support for AI endeavors following the peak in the 1980s (one of the so-called ‘AI
Winters’), for reasons that are discussed in Chap. 2. For the purpose of this introduc-
tion, we focus on the events of recent years, which have seen rapid growth in inter-
est in AIM applications driven by media attention to AI in general (evident to the
right of Fig. 1.1), coupled with high profile medical demonstrations of diagnostic

Fig. 1.1 Frequency with Frequency of Term Artificial Intelligence

1e–6
which the term ‘artificial
intelligence’ appears in 3.0
books published between
1950 and 2019 and 2.5
digitized by Google (data
obtained from the Google 2.0
Books n-gram viewer
website [2]). 1e-6 indicates 1.5
the order of frequency of
occurrence of the term 1.0
(e.g. approximately 2.5
occurrences per million 0.5
bigrams at the peak in the
0.0
late eighties)
1960 1970 1980 1990 2000 2010 2020
1 Introducing AI in Medicine 5

accuracy, particularly in image recognition. This growth is part of a larger picture in

which the capabilities of artificial neural networks—originally conceived as mod-
els of human information processing and learning [4, 5]—have been enhanced
through a convergence of the availability of large data sets for training, refinements
in training approaches, and increases in computational power.

Neural Networks and Deep Learning

Loosely inspired by the interconnections between neurons in the human brain, arti-
ficial neural networks consist of interconnected functional units named neurons,
each producing an output signal determined by their input data, weights assigned to
incoming connections, and an activation function that transforms cumulative
incoming signals into an output that is passed on to a next layer of the network. The
weights of a neural network serve as parameters that can be altered during training
of a model, so that the output of the neural network better approximates a desired
result, such as assigning a high probability to the correct diagnostic label for a radio-
logical image. When used in this way, neural networks exemplify the paradigm of
supervised machine learning, in which models learn from labels (such as diagno-
ses) assigned to training data. This approach is very different in nature from the
deliberate engineering of human knowledge that supported the expert systems in the
first wave of AIM (see Chap. 2 and, for detailed accounts of knowledge modeling
and machine learning methods, see Chaps. 4 and 6 respectively).
While machine learning models can learn to make impressively accurate predic-
tions, especially when large data sets are available for training, systems leveraging
explicitly modeled human knowledge—systems intended to reason as humans do—
are much better positioned to explain themselves (for an example, see Box 1.1) than
systems that have been developed to optimize accuracy without considering human
cognition. Explanation has long been recognized as a desirable property of AI sys-
tems for automated diagnosis, and as a prerequisite for their acceptance by clini-
cians [6] (and see Chap. 8). However, the general trend in machine learning has
been that accuracy comes at the cost of interpretability, to the point at which restor-
ing some semblance of interpretability to the predictions made by contemporary
machine learning models has emerged as a field of research in its own right—
explainable AI—with support from the Defense Advanced Research Projects
Agency (DARPA),1 the same agency that initiated the research program on network
protocols that ultimately led to a consumer-accessible internet.

1
See https://www.darpa.mil/program/explainable-artificial-intelligence (accessed August 18,
2022) for details.
6 T. A. Cohen et al.

Box 1.1 An explanation provided by the MYCIN system in response to a

user entering “WHY”: From Shortliffe et al. 1974 [7]
–– **WHY
–– [1.0] It is important to find out whether there is therapeutically significant
disease associated with this occurrence of ORGANISM-1.
–– It has already been established that:
–– [1.1] the site of the culture is not one of those which are normally sterile, and
–– [1.2] the method of collection is sterile
–– Therefore, if:
–– [1.3] the organism has been observed in significant numbers
–– Then: there is strongly suggestive evidence (.9) that there is therapeutically
significant disease associated with this occurrence of the organism
–– [Also: there is strongly suggestive evidence (.8) that the organism is not a
contaminant]

This trend toward accurate but opaque predictions has accelerated with the advent
of deep learning models—neural networks that have multiple intervening layers of
neurons between input data and output predictions. While deep neural network
architectures are not new phenomena (see for example the important paper by
Hinton et al. [8]), their performance when trained on large data sets has produced
dramatic improvements in results attained across fundamental tasks such as speech
recognition, question answering and image recognition.
Figure 1.2 shows the extent of recent improvements for three key benchmarks:
the Stanford Question Answering Dataset (SQUAD [9])—over 100,000 compre-
hension questions related to short articles; ImageNet—over 14 million images each
assigned one of two hundred possible class labels [10]; and LibriSpeech—over
1000 hours of speech with matching text from audiobooks [11]. Of note, with both
SQUAD and ImageNet, human performance on the tasks concerned has been esti-
mated, and superseded by deep learning models.
Conceptually, the advantages of deep learning models over previous machine
learning approaches have been attributed to their capacity for representation learn-
ing [12]. With prior machine learning approaches, performance generally advanced
through engineering ways to represent incoming data (such as pixels of an image
representing a handwritten digit) that led to better downstream machine learning
performance (representations such as a count of the number of loops in a handwrit-
ten digit [13]). With deep learning models, the lower layers of a network can learn
to represent incoming data in ways that facilitate task performance automatically.2
Of particular importance for domains such as medicine, where large labeled data

2
While deep learning models excel at learning representations that lead to better predictive model-
ing performance, representation learning is broader than deep learning and includes a number
of previously established methods. For a review of developments up to 2013, see [14].
1 Introducing AI in Medicine 7

SQUAD1.1 (F1)
ImageNet (top 5 acc)
97.5 LibriSpeech (100-wer)

ImageNet human: 94.9

95.0

92.5
Squad1.1 human: 91.2
best performance

90.0

87.5

85.0

82.5

80.0

2012 2014 2016 2018 2020

Year

Fig. 1.2 Best documented performance, by year, on three key benchmarks (data from the 2021 AI
Index Report [15, 16]). (1) SQUAD1.1 = Stanford Question Answering Dataset (version 1.1).
Performance metric is “F1” (the balanced f-measure; see Chap. 6); (2) ImageNet - performance
metric is “top 5 acc” (the percent of images in which the correct label, among 200 possibilities,
appeared in the top 5 predictions); (3) LibriSpeech - performance metrics is “100-wer” (a transfor-
mation of the word error rate, with 100 indicating every word in a recording was recognized cor-
rectly). Dashed lines indicate documented human performance on the task concerned, which has
been superseded by AI in both cases

sets are relatively difficult to obtain, the ability to extract useful representations for
one task can often be learned from training on another related one. This ability to
apply information learned from one task or data set to another is known as transfer
learning, and is perhaps best exemplified by what has become a standard approach
to classifying medical images (see Chap. 12): adding a classification layer to a deep
neural network that has been pretrained on the task of recognizing non-medical
images in ImageNet [17]. Similarly, fine-tuning of models such as Google’s BERT
and Open-AI’s GPT series, which were originally trained to predict held-out words
in large amounts of text from a range of digital data sources, has advanced perfor-
mance across a broad range of natural language processing (NLP) tasks [18, 19].
8 T. A. Cohen et al.

a b

Fig. 1.3 Recognition of a subtle diagnostic cue by a deep neural network trained to detect thyroid
cancer in different ultrasound images of the same nodule. Each image (top row) is annotated with
the probability of malignancy according to the model, and is paired with a visualization of the
pixels attended to by the deep learning model when making a prediction for whether an image is
in the “malignant class”, developed using the GradCam method [20]. Only the second image from
the left exhibits the diagnostic feature of interrupted eggshell calcification, in which the rim of the
opaque “shell” of calcification (blue arrows in the top row) is disrupted (red arrow). The GradCam
visualization reveals the model has learned to attend to this subtle diagnostic feature. Image cour-
tesy of Dr. Nikita Pozdeyev

Machine Learning and Medical Practice

Of course, outperforming humans on the repetitive and mundane task of selecting

among hundreds of possible labels for a given image, or surpassing their accuracy
in answering multiple choice questions about particular passages, does not neces-
sarily provide an indication that deep neural networks could meet the requirements
for flexibility, prioritization and adaptive decision making under uncertainty needed
to replace medical practitioners in a busy clinical environment (audiobooks are also
far less challenging to transcribe than recordings captured in a naturalistic environ-
ment—see Chap. 9 for a related discussion of automated medical transcription).
Nonetheless, the ability to recognize diagnostically important features is a funda-
mental task in interpreting medical images (as illustrated in Fig. 1.3—see also Chap.
12). A system capable of answering clinical questions accurately on the basis of
written notes would make the information that these notes contain amenable to
downstream computational processing for decision support or observational research
(methods to achieve such ends are discussed in detail in Chap. 7). Furthermore, simi-
lar advances in performance have been achieved by predictive models in medicine,
due in part to the large volume of digitized medical data that has accompanied the
adoption of electronic health record (EHR) systems,3 and the widespread use of
digital platforms for image storage and retrieval (see Chap. 3) [22].

3
In the United States this increase in adoption is attributable to the incentivization structures pro-
vided by the Health Information Technology for Economic and Clinical Health (HITECH) act of
2009 [21].
1 Introducing AI in Medicine 9

For example, a 2016 paper in the Journal of the American Medical Association
describes an impressively accurate deep learning system for the diagnosis of diabetes-­
related eye disease in images of the retina [23]. Similarly, a widely-cited 2017 paper
in Nature describes the application of deep learning to detect skin cancer [24], with
the resulting system performing as well as 21 board-certified dermatologists in iden-
tifying two types of neoplastic skin lesions. These systems leveraged recent advances
in AI, including deep neural network architectures and approaches to train them effi-
ciently, as well as large sets of labeled data that were used to train the networks—over
125,000 images in each study. The dermatology system benefitted further from pre-
training on over 1.25 million non-medical images labeled with 1000 object categories.
Beyond imaging, deep learning models trained on EHR data have learned to predict
in-hospital mortality, unplanned readmission, prolonged length of stay, and final dis-
charge diagnosis—in many cases outperforming traditional predictive models that are
still widely used in clinical practice [25]. In this case, models were trained on data
from over 200,000 hospitalized adult patients from two academic medical centers,
considering over 40 billion sequential data points collectively.
These advances have attracted a great deal of press attention, with frequent arti-
cles in prominent media outlets considering the potential of AI to enhance—or dis-
rupt—the practice of medicine [26–28]. As we have discussed in the preface to this
volume, neither AI systems with physician-level performance nor media attention to
such systems are without precedent, even in the days before advances in computa-
tional power and methodology mediated the current explosive interest in machine
learning. However, the convergence of an unprecedented availability of clinical data
with the maturation of machine learning models (and the computational resources
to train them at scale) has allowed the rapid development of AI-based predictive
models in medicine. Many demonstrate impressive results beyond those we have
briefly described here. Furthermore, the proven commercial viability and public
acceptance of such models in other areas have offset some of the skepticism with
which AI models were greeted initially. Having seen the effectiveness with which
machine learning models leverage data to deliver our entertainment and shopping
recommendations on a daily basis, why would we not wish such systems to assist
our clinicians in their medical practice? A strong indicator of the commercial poten-
tial of AI-based systems in medicine is the emergence of regulatory frameworks for
their application in practice (see also Chap. 18) [29], with a number of AI systems
already approved for medical use in the United States (Fig. 1.4) and Europe [30].

The Scope of AIM

A fundamental question in the study (and regulation) of AIM systems concerns the
definition of the term “Artificial Intelligence”. Given the breadth of approaches that
have been categorized as related to AI, it is perhaps not surprising that there is no
universally-accepted definition of this term, and that the extent to which contempo-
rary deep learning approaches constitute AI is still vigorously debated [32, 33]. A
representative sample of AI definitions is provided in Box 1.2. While there are
10 T. A. Cohen et al.

SPECIALITY
Cardiology
Radiology
70 Neurology
Psychiatry
General medicine
Hospital monitoring
Ophthalmology
Endocrinology
Orthopedics
60 Internal Medicine
Urology / General practise

50

40
COUNT

30

20

10

2013 2014 2015 2016 2017 2018 2019 2020 2021

DATE

Fig. 1.4 FDA approvals for AI-related products by specialty (data drawn from medicalfuturist.
com [30, 31]) with radiology systems (X) the most common category

Box 1.2 Sample Definitions of Artificial Intelligence

• “The study of complex information processing problems that often have
their roots in some aspect of biological information processing” (Marr,
1977) [34]
• “…the study of ideas that enable computers to do the things that make
human beings seem intelligent: the ability to reason symbolically, the abil-
ity to acquire and apply knowledge, and the ability to manipulate and com-
municate ideas” (Winston, 1977) [35]
• “….the part of computer science concerned with designing intelligent
computer systems, that is, systems that exhibit the characteristics we asso-
ciate with intelligence in human behavior – understanding, language,
learning, reasoning, solving problems and so on” (Barr et al., vol 1, 1981,
p. 3) [36]
1 Introducing AI in Medicine 11

• “The branch of computer science that is concerned with the automation of

intelligent behavior” (Luger and Stubblefield, 1993) [37]
• “It is the science and engineering of making intelligent machines, espe-
cially intelligent computer programs. It is related to the similar task of
using computers to understand human intelligence, but AI does not have to
confine itself to methods that are biologically observable”. (McCarthy,
2007) [38]

clearly common threads that run among them, notably the emphasis on intelligence
(loosely defined by Barr as exhibiting the characteristics we associate with intelli-
gence in human behavior, or by Winston as emphasizing the use of knowledge and
an ability to communicate ideas), the definitions also reflect a departure from the
cognitive motivations of AI at its inception—performance of tasks as humans do—
to the more pragmatic motivations of the performance-oriented systems that are
commonly termed AI today. Note that McCarthy in particular asserts explicitly that
biological constraints need not apply. Of course, motivations for understanding how
machines might solve a problem presumed to require human intelligence are not
exclusively pragmatic, as this topic is also of considerable academic interest.
As one might anticipate given the fluidity of definitions of AI in general, the
notion of what qualifies as AI in medicine is also in flux. At the inception of the
field, the focus was on systems that could reason, leveraging encoded knowledge
(including probabilistic estimates or uncertainty) derived from clinical experts.
Such formalization of knowledge to render it computable also underlies the clinical
decision support rules embedded in contemporary EHR systems. However, few
would argue that the individual rules firing alerts in such systems constitute AI, even
when considered collectively (see the discussion of warnings and alerts in Chap.
17). It seems, therefore, that the perceived difficulty of the tasks accomplished by a
system determine whether it is thought to have exhibited intelligent behavior. Today,
machine learning approaches (including deep neural networks) are strongly associ-
ated with the term AI. These systems are not designed to reason, but instead learn to
recognize patterns, such as diagnostic features of radiology images, leading to per-
formance on constrained tasks that is comparable to that of highly trained physi-
cians. As such it is easy to argue that they exhibit intelligent human behavior, at
least in the context of a task for which large amounts of labeled training data are
readily available. Furthermore, such models can make predictions that are beyond
the capabilities of human experts at times, such as prediction of cardiovascular risk
factor status from retinal fundus photographs [39], or prediction of 3-D protein
structure from an amino acid sequence [40]. Perhaps as a consequence of the lack
of funding for research associated with the term AI during periods in which it was
out of favor (see Chap. 2), a great deal of machine learning work in the field was not
framed as AI research, but would be perceived this way in retrospect. Analogous to
the case with rule-based models, this raises the question of how sophisticated a
machine learning model is required to qualify as AI. For example, would a system
12 T. A. Cohen et al.

based on a logistic regression model trained on a handful of features, with less than
ten trainable parameters constitute AI? Perhaps, as with rules, the main question
concerns the nature of the task that the model is able to accomplish, with a bench-
mark for AIM being the automated accomplishment of tasks that would be chal-
lenging for a highly trained human.

From Accurate Predictions to Clinically Useful AIM

However, irrespective of whether the engineers of AIM systems attempt to emulate

human-like problem-solving processes, the ultimate goal of such efforts is often to
support decision making by human clinicians at the point of care. The role of AIM
in improving the quality, efficiency and safety of clinical practice exists within a
larger system that includes human decision makers [41]. As such, both the remark-
able capabilities and recognized constraints of human information processing must
also be considered when designing and deploying AI-based systems, even if the
systems concerned do not explicitly attempt to emulate human information process-
ing methods. The consideration of the broader context in which AI-based systems
must operate to influence patient care reveals a number of challenges that must be
overcome in order to bridge the gulf between systems that perform well in the con-
text of a constrained reference set, and systems that provide clinical utility at the
point of care. Many of these challenges have been recognized since the inception of
the field. In a 1975 paper, Shortliffe and Davis identified a series of seven consider-
ations for expert system evaluation that suggest a path from conception of a system
to clinical utility (Table 1.1; see also Chap. 17).
Of note, most of the work on accurate automated medical image interpretation
we have discussed addresses only the second consideration in Table 1.1, and improv-
ing the ability of machine learning models to approach (or even surpass) the accu-
racy of expert clinicians has remained the focus of much recent work [43]. However,
such models must be embedded in systems that are both usable and acceptable to
clinicians if they are to exert an effect on management to improve outcomes for
patients or to advance institutional or societal priorities such as cost-effectiveness.
Furthermore, the design of AI systems should be motivated by the needs of clini-
cians, which are best understood in the context of the processes and environmental
constraints in which they work [41].

Table 1.1 Overview: seven Demonstration Impact

considerations for system
1 Need 5 Management
evaluation [42]
2 Expert-level 6 Patient outcome
performance
3 Usability 7 Cost-effectiveness
4 Acceptance by
clinicians
1 Introducing AI in Medicine 13

The Cognitive Informatics Perspective

Why CI?

It is our view that the discipline of cognitive informatics (CI) [44–46], which brings
the perspective of the cognitive sciences to the study of medical decision making by
human beings and machines, is uniquely positioned to address these challenges.
Through its roots in the study of medical reasoning [47–49], CI provides a sound
scientific basis from which to consider the relationship between current technolo-
gies and human intelligence. CI has extended its area of inquiry to include both
human-computer interaction and the study of the effects of technology on the flow
of work and information in clinical settings [50–53]. Accordingly CI is well-­
positioned to inform the integration of AIM systems into clinical practice, and more
broadly to inform the design of AI systems that complement the cognitive capabili-
ties of human decision makers, in alignment with seminal ideas concerning the
potential of cooperative human-machine systems [54].

The Complementarity of Human and Machine Intelligence

As is discussed in Chap. 5, evaluations in the context of image processing tasks have

demonstrated that the performance of human beings and machines working in con-
cert can result in better diagnostic accuracy than either machines or human
beings alone [55–57]. In some ways this is not surprising, given the different strate-
gies human experts and machines employ to achieve diagnostic accuracy. Cognitive
studies of radiologists have shown that experts in this domain integrate their knowl-
edge of anatomical structures and their projections onto two-dimensional images,
with their knowledge of general physiology and specific disease processes. This
allows radiologists to generate initial hypotheses that narrow the focus of their
search for a definitive diagnosis [47]. In contrast, contemporary neural network
models learn to identify radiological abnormalities by training two-dimensional
“feature detectors” to recognize regions that are useful in distinguishing between
diagnostic categories in the training data (as illustrated previously, in Fig. 1.3), irre-
spective of where within an image these regions may occur [58]. Differences in the
processes through which neural networks and human experts interpret images can
also be detected empirically. Recent work has shown that human beings and
machines focus on different features when interpreting histology slides [59].
Acknowledgment of these differences leads naturally to the conclusion that a
human/AI collaborative team has the potential to make better decisions than those
that would emerge from a fully automated or exclusively manual process (see, for
example, the discussion of QMR in Chap. 2). However, many open questions remain
regarding how best to realize this potential. A promising proposal concerns deliber-
ately designing AI systems to compensate for known “blind spots” in clinical
14 T. A. Cohen et al.

decision making [60], such as biases in diagnostic reasoning that have been identi-
fied through cognitive research [61], or distracted attention in busy clinical settings
[62]. Alternatively, one might envision developing ways to distribute labor across a
human/AI collaborative system to maximize the expected utility of this system, tak-
ing into account both the accuracy of model predictions and the time required for a
human actor to reassess them. Recent work has developed an approach to optimiz-
ing collaborative systems in this way, resulting in some experiments in systems that
increase high-confidence predictions (i.e. predictions to which the model assigns
extremely high or low probability) at the expense of its accuracy in edge cases (i.e.
predictions close to the model’s decision boundary), where human input could
resolve model uncertainty [63].

Mediating Safe and Effective Human Use of AI-Based Tools

CI methods are already well established as means to evaluate the usability of deci-
sion support tools [45, 46]. Findings from this line of research have led to recom-
mendations that the usability of clinical systems should be prioritized as a means
to enhance their acceptability and safety [64]. In contrast to system-centric meth-
ods of usability evaluation, such as heuristic evaluations by usability experts [65],
CI approaches attempt to understand the thought process of a user, which is par-
ticularly important in knowledge-rich domains, such as medicine, where both
knowledge of the system being used and of the domain are required to perform
tasks correctly [66]. This can be accomplished through analysis of a think-aloud
protocol, collected by prompting users to verbalize their thoughts during the pro-
cess of completing representative tasks [67]. This approach is similarly well-suited
to the study of clinician interactions with AI-based systems, where users must
make clinical decisions on the basis of their estimation of the veracity of sys-
tem output.
Critical questions concerning the nature of these interactions remain to be
answered. One such question concerns how best to represent model predictions. For
example, recent work in dermatology diagnosis found that advantages in perfor-
mance for a human-computer collective were contingent upon the granularity (prob-
abilities of all of the diseases in the differential diagnosis vs. a single global risk of
malignancy) and cognitive demand of the representation used to convey predictions
to physicians [57]. Analysis of verbal protocols collected during interactions with
interfaces, using alternative representations of the same predictions, could inform our
understanding of why this is the case by revealing the reasoning dermatologists use
when deciding whether to accept a particular recommendation. Another important
question concerns the role of explanations provided by a system in influencing human
decision making. Intriguingly, recent research has shown that revealing the influence
of input features (here, words in a passage of text) on model predictions increases the
likelihood that users will accept the validity of these predictions, irrespective of
whether they are accurate [68]. This suggests that displaying feature salience may not
1 Introducing AI in Medicine 15

be adequate to support the fault detection procedures that are a prerequisite to safe
and resilient human/AI collaborative systems. CI methods are well-suited to identify
the thought processes through which faulty AI decisions are (or are not) identified
when considering explanations, to inform the development of effective systems in
which process are both highly automated and subject to human control. This should
arguably be the case for systems making critical medical decisions, where mistakes
have irreversible consequences [69].

Concluding Remarks

In this chapter, we have provided an introduction to AIM, with a focus on recent

developments. In doing so, we have highlighted some key challenges that AI models
must meet if they are to achieve the goal of improving the efficiency, safety and
quality of health care. We have argued that the field of CI is well-suited to address
these challenges, by providing greater insight into the role of the human component
of human/AI collaborative systems, to inform their design and evaluation.
Consideration of the cognitive processes through which human beings evaluate,
interpret and act upon the recommendations made by AI systems is fundamental to
the development of solutions that enhance the capabilities of clinicians and research-
ers in the biomedical domain. Accordingly, one of our goals in developing this vol-
ume has been to provide a resource to support the multidisciplinary training required
to design and implement AI methods with the potential to enhance the practice of
medicine as well as life science research in human biology.
Questions for Discussion
• What is an example of a recent technological advancement in AIM, and what are
its implications for clinical practice?
• Provide your own definition of AIM that reflects the discussion in this chapter
(i.e., do not simply pick one from Box 1.2). Do any aspects of the field of which
you are aware fall outside the scope of this definition?
• What are the main application areas and techniques for AIM?
• AI in medicine has a long history, and AIM technologies have been proposed as
a potential disruptor of the healthcare industry before. What current contextual
factors might increase or limit the potential for broad adoption?

Further Reading
Chang, AC. Intelligence-Based Medicine: Artificial Intelligence and Human
Cognition in Clinical Medicine and Healthcare. Academic Press (Elsevier); July
8th 2020.
• This book provides a survey of AI methods from clinical and data science per-
spectives, with an emphasis on their implementation in, and impact upon, medi-
cine and its subspecialties.
16 T. A. Cohen et al.

Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare:
review, opportunities and challenges. Briefings in Bioinformatics. 2018 Nov 27;19
(6):1236–1246.
• This paper provides an overview and of deep learning applications in healthcare
up to 2018, and introduces a number of issues that are addressed in the cur-
rent volume.
Patel VL, Kannampallil TG. Cognitive informatics in biomedicine and healthcare.
Journal of biomedical informatics. 2015 Feb 1;53:3–14.
• This paper provides a definition and overview of the field of cognitive informat-
ics, with a focus on biomedical applications.
Topol EJ. High-performance medicine: the convergence of human and artificial
intelligence. Nature Medicine. Nature Publishing Group; 2019 Jan;25 (1):44–56.
• This paper provides an overview of AI applications in healthcare, including a
thoughtful account of challenges that distinguish this domain from others in
which AI applications have established their value.
Zhang D, Mishra S, Brynjolfsson E, Etchemendy J, Ganguli D, Grosz B, Lyons T,
Manyika J, Niebles JC, Sellitto M, Shoham Y, Clark J, Perrault R. The AI Index
2021 Annual Report. arXiv:210306312 [cs] [Internet]. 2021 Mar 8 [cited 2021 Apr
24]; Available from: http://arxiv.org/abs/2103.06312
• Stanford’s AI Index Report provides an overview of national and global AI trends
in research and industry.

References

1. McCarthy J, Minsky ML, Rochester N, Shannon CE. A proposal for the Dartmouth summer
research project on artificial intelligence, August 31, 1955. AIMag. 2006;27(4):12.
2. Google Books Ngram Viewer [Internet]. [cited 2021 June 25]. Available from: https://books.
google.com/ngrams.
3. Yu VL, Buchanan BG, Shortliffe EH, Wraith SM, Davis R, Scott AC, Cohen SN. Evaluating the
performance of a computer-based consultant. Comput Programs Biomed. 1979;9(1):95–102.
4. Rosenblatt F. The perceptron: a probabilistic model for information storage and organization
in the brain. Psychol Rev. 1958;65(6):386.
5. McClelland JL, Rumelhart DE, Group PR. Parallel distributed processing. Boston, MA: MIT
Press; 1986. p. 1.
6. Swartout WR. Explaining and justifying expert consulting programs. Computer-assisted medi-
cal decision making. Springer; 1985. p. 254–71.
7. Shortliffe EH, Davis R, Axline SG, Buchanan BG, Green CC, Cohen SN. Computer-based
consultations in clinical therapeutics: explanation and rule acquisition capabilities of the
MYCIN system. Comput Biomed Res. 1975;8(4):303–20.
8. Hinton GE, Osindero S, Teh Y-W. A fast learning algorithm for deep belief nets. Neural
Comput. 2006;18(7):1527–54.
1 Introducing AI in Medicine 17

9. Rajpurkar P, Zhang J, Lopyrev K, Liang P. Squad: 100,000+ questions for machine compre-
hension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural
Language Processing, pages 2383–2392, Austin, Texas. Association for Computational
Linguistics.
10. Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L. Imagenet: a large-scale hierarchical image
database. 2009 IEEE conference on computer vision and pattern recognition. IEEE; 2009.
p. 248–55.
11. Panayotov V, Chen G, Povey D, Khudanpur S. Librispeech: an ASR corpus based on public
domain audio books. 2015 IEEE international conference on acoustics, speech and signal pro-
cessing (ICASSP). 2015. p. 5206–5210.
12. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–44.
13. Kumar G, Bhatia PK. A detailed review of feature extraction in image processing systems.
2014 fourth international conference on advanced computing communication technologies.
2014. p. 5–12.
14. Bengio Y, Courville A, Vincent P. Representation learning: a review and new perspectives.
IEEE Trans Pattern Anal Mach Intell. 2013;35(8):1798–828.
15. Zhang D, Mishra S, Brynjolfsson E, Etchemendy J, Ganguli D, Grosz B, Lyons T, Manyika
J, Niebles JC, Sellitto M, Shoham Y, Clark J, Perrault R. The AI Index 2021 annual report.
arXiv:210306312 [cs] [Internet]. 2021 Mar 8 [cited 2021 Apr 24]. Available from: http://arxiv.
org/abs/2103.06312.
16. AI Index 2021 [Internet]. Stanford HAI. [cited 2021 June 25]. Available from: https://hai.
stanford.edu/research/ai-­index-­2021.
17. Shin H-C, Roth HR, Gao M, Lu L, Xu Z, Nogues I, Yao J, Mollura D, Summers RM. Deep
convolutional neural networks for computer-aided detection: CNN architectures, dataset char-
acteristics and transfer learning. IEEE Trans Med Imaging. 2016;35(5):1285–98.
18. Devlin J, Chang M-W, Lee K, Toutanova K. BERT: pre-training of deep bidirectional trans-
formers for language understanding. Proceedings of the 2019 conference of the North
American Chapter of the Association for computational linguistics: human language technolo-
gies, Vol. 1 (Long and Short Papers). 2019. p. 4171–4186.
19. Radford A, Wu J, Child R, Luan D, Amodei D, Sutskever I. Language models are unsupervised
multitask learners. OpenAI Blog. 2019;1(8):9.
20. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual expla-
nations from deep networks via gradient-based localization. Proceedings of the IEEE interna-
tional conference on computer vision. 2017. p. 618–626.
21. Adler-Milstein J, Jha AK. HITECH act drove large gains in hospital electronic health record
adoption. Health Aff. 2017;36(8):1416–22.
22. Bauman RA, Gell G, Dwyer SJ. Large picture archiving and communication systems of the
world--part 1. J Digit Imaging. 1996;9(3):99–103.
23. Gulshan V, Peng L, Coram M, Stumpe MC, Wu D, Narayanaswamy A, Venugopalan S, Widner
K, Madams T, Cuadros J. Development and validation of a deep learning algorithm for detec-
tion of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316(22):2402–10.
24. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level
classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115–8.
25. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M,
Sundberg P, Yee H, Zhang K, Zhang Y, Flores G, Duggan GE, Irvine J, Le Q, Litsch K, Mossin
A, Tansuwan J, Wang D, Wexler J, Wilson J, Ludwig D, Volchenboum SL, Chou K, Pearson
M, Madabushi S, Shah NH, Butte AJ, Howell MD, Cui C, Corrado GS, Dean J. Scalable and
accurate deep learning with electronic health records. NPJ Digit Med. 2018;1(1):1–10.
26. Mukherjee S. A.I. versus M.D. [Internet]. The New Yorker. [cited 2021 Apr 15]. https://www.
newyorker.com/magazine/2017/04/03/ai-­versus-­md.
27. Metz C. A.I. shows promise assisting physicians. The New York Times [Internet]. 2019 Feb
11 [cited 2021 Apr 15]. https://www.nytimes.com/2019/02/11/health/artificial-­intelligence-­
medical-­diagnosis.html.
18 T. A. Cohen et al.

28. O’Connor A. How artificial intelligence could transform medicine. The New York Times
[Internet]. 2019 Mar 11 [cited 2021 Apr 15]. https://www.nytimes.com/2019/03/11/well/live/
how-­artificial-­intelligence-­could-­transform-­medicine.html.
29. Health C for D and R. Artificial intelligence and machine learning in soft-
ware as a medical device. FDA [Internet]. FDA; 2021 Jan 11 [cited 2021 Apr 19].
https://www.fda.gov/medical-­d evices/software-­m edical-­d evice-­s amd/artificial-
intelligence-­and-­machine-­learning-­software-­medical-­device.
30. Benjamens S, Dhunnoo P, Meskó B. The state of artificial intelligence-based FDA-approved
medical devices and algorithms: an online database. NPJ Digit Med. 2020;3(1):1–8.
31. The Medical Futurist [Internet]. The Medical Futurist. [cited 2021 Apr 19]. Available from:
https://medicalfuturist.com/fda-­approved-­ai-­based-­algorithms.
32. Marcus G. Deep learning: a critical appraisal. arXiv preprint arXiv:180100631. 2018.
33. Zador AM. A critique of pure learning and what artificial neural networks can learn from ani-
mal brains. Nat Commun. 2019;10(1):1–7.
34. Marr D. Artificial intelligence—a personal view. Artif Intell. 1977;9(1):37–48.
35. Winston PH. Artificial Intelligence. Reading, MA: Addison-Wesley; 1977.
36. Barr A, Feigenbaum EA. The handbook of artificial intelligence (Vol. 1). Los Altos, CA:
William Kaufman; 1981.
37. Luger GF, Stubblefield WA. Artificial intelligence (2nd ed.): structures and strategies for com-
plex problem-solving. USA: Benjamin-Cummings Publishing Co., Inc.; 1993.
38. McCarthy J. What is artificial intelligence? [Internet]. What is artificial intelligence? 2007
[cited 2021 Apr 20]. http://www-­formal.stanford.edu/jmc/whatisai/whatisai.html.
39. Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, Peng L, Webster
DR. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learn-
ing. Nat Biomed Eng. 2018;2(3):158–64.
40. Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K,
Bates R, Žídek A, Potapenko A, Bridgland A, Meyer C, Kohl SAA, Ballard AJ, Cowie A,
Romera-Paredes B, Nikolov S, Jain R, Adler J, Back T, Petersen S, Reiman D, Clancy E,
Zielinski M, Steinegger M, Pacholska M, Berghammer T, Bodenstein S, Silver D, Vinyals O,
Senior AW, Kavukcuoglu K, Kohli P, Hassabis D. Highly accurate protein structure prediction
with AlphaFold. Nature. 2021;15:1–11.
41. Berg M. Patient care information systems and health care work: a sociotechnical approach. Int
J Med Inform. 1999;55:87–101.
42. Shortliffe T, Davis R. Some considerations for the implementation of knowledge-based expert
systems. SIGART Bull. 1975;55:9–12.
43. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence.
Nat Med. 2019;25(1):44–56.
44. Wang Y. The theoretical framework of cognitive informatics. Int J Cogn Inform Nat Intell.
2007;1(1):1–27.
45. Patel VL, Kaufman DR. Cognitive science and biomedical informatics. In: Shortliffe EH,
Cimino JJ, editors. Biomedical informatics: computer applications in health care and biomedi-
cine. 5th ed. New York: Springer; 2021. p. 133–85.
46. Patel VL, Kannampallil TG. Cognitive informatics in biomedicine and healthcare. J Biomed
Inform. 2015;53:3–14.
47. Lesgold A, Rubinson H, Feltovich P, Glaser R, Klopfer D, Wang Y. Expertise in a complex
skill: diagnosing x-ray pictures. In: Chi MTH, Glaser R, Farr MJ, editors. The nature of exper-
tise. Hillsdale, NJ: Lawrence Erlbaum; 1988. p. 311–42.
48. Elstein AS, Shulman LS, Sprafka SA. Medical problem solving: an analysis of clinical reason-
ing. Cambridge, MA: Harvard University Press; 1978.
49. Patel VL, Arocha JF, Kaufman DR. Diagnostic reasoning and medical expertise. Psychol
Learn Motiv. 1994;31:187–252.
50. Kushniruk AW, Patel VL, Cimino JJ. Usability testing in medical informatics: cognitive
approaches to evaluation of information systems and user interfaces. Proceedings/AMIA
annual fall symposium. 1997. p. 218–222.
1 Introducing AI in Medicine 19

51. Kushniruk AW, Patel VL. Cognitive and usability engineering methods for the evaluation of
clinical information systems. J Biomed Inform. 2004;37:56–76.
52. Malhotra S, Jordan D, Shortliffe E, Patel VL. Workflow modeling in critical care: piecing
together your own puzzle. J Biomed Inform. 2007;40:81–92.
53. Cohen T, Blatter B, Almeida C, Shortliffe E, Patel V. A cognitive blueprint of collaboration
in context: distributed cognition in the psychiatric emergency department. Artif Intell Med.
2006;37:73–83.
54. Licklider JC. Man-computer symbiosis. IRE transactions on human factors in electronics.
IEEE. 1960;1:4–11.
55. Patel BN, Rosenberg L, Willcox G, Baltaxe D, Lyons M, Irvin J, Rajpurkar P, Amrhein T,
Gupta R, Halabi S, Langlotz C, Lo E, Mammarappallil J, Mariano AJ, Riley G, Seekins J, Shen
L, Zucker E, Lungren MP. Human–machine partnership with artificial intelligence for chest
radiograph diagnosis. NPJ Digit Med. 2019;2(1):1–10.
56. Hekler A, Utikal JS, Enk AH, Hauschild A, Weichenthal M, Maron RC, Berking C, Haferkamp
S, Klode J, Schadendorf D, Schilling B, Holland-Letz T, Izar B, von Kalle C, Fröhling S,
Brinker TJ, Schmitt L, Peitsch WK, Hoffmann F, Becker JC, Drusio C, Jansen P, Klode J,
Lodde G, Sammet S, Schadendorf D, Sondermann W, Ugurel S, Zader J, Enk A, Salzmann
M, Schäfer S, Schäkel K, Winkler J, Wölbing P, Asper H, Bohne A-S, Brown V, Burba B,
Deffaa S, Dietrich C, Dietrich M, Drerup KA, Egberts F, Erkens A-S, Greven S, Harde V,
Jost M, Kaeding M, Kosova K, Lischner S, Maagk M, Messinger AL, Metzner M, Motamedi
R, Rosenthal A-C, Seidl U, Stemmermann J, Torz K, Velez JG, Haiduk J, Alter M, Bär C,
Bergenthal P, Gerlach A, Holtorf C, Karoglan A, Kindermann S, Kraas L, Felcht M, Gaiser
MR, Klemke C-D, Kurzen H, Leibing T, Müller V, Reinhard RR, Utikal J, Winter F, Berking C,
Eicher L, Hartmann D, Heppt M, Kilian K, Krammer S, Lill D, Niesert A-C, Oppel E, Sattler
E, Senner S, Wallmichrath J, Wolff H, Gesierich A, Giner T, Glutsch V, Kerstan A, Presser D,
Schrüfer P, Schummer P, Stolze I, Weber J, Drexler K, Haferkamp S, Mickler M, Stauner CT,
Thiem A. Superior skin cancer classification by the combination of human and artificial intel-
ligence. Eur J Cancer. 2019;120:114–21.
57. Tschandl P, Rinner C, Apalla Z, Argenziano G, Codella N, Halpern A, Janda M, Lallas A,
Longo C, Malvehy J, Paoli J, Puig S, Rosendahl C, Soyer HP, Zalaudek I, Kittler H. Human–
computer collaboration for skin cancer recognition. Nat Med. 2020;26(8):1229–34.
58. Soffer S, Ben-Cohen A, Shimon O, Amitai MM, Greenspan H, Klang E. Convolutional
neural networks for radiologic images: a Radiologist’s guide. Radiology. 2019;290(3):
590–606.
59. Kimeswenger S, Tschandl P, Noack P, Hofmarcher M, Rumetshofer E, Kindermann H, Silye
R, Hochreiter S, Kaltenbrunner M, Guenova E, Klambauer G, Hoetzenecker W. Artificial neu-
ral networks and pathologists recognize basal cell carcinomas based on different histological
patterns. Mod Pathol. 2020;13:1–9.
60. Horvitz E. One hundred year study on artificial intelligence: reflections and framing. Microsoft
com. 2014
61. Chapman GB, Elstein AS. Cognitive processes and biases in medical decision-making. In:
Chapman GB, Sonnenberg FS, editors. Decision-making in health care: theory, psychology,
and applications. Cambridge: Cambridge University Press; 2000. p. 183–210.
62. Franklin A, Liu Y, Li Z, Nguyen V, Johnson TR, Robinson D, Okafor N, King B, Patel VL,
Zhang J. Opportunistic decision making and complexity in emergency care. J Biomed Inform.
2011;44(3):469–76.
63. Bansal G, Nushi B, Kamar E, Horvitz E, Weld DS. Is the Most accurate AI the best teammate?
Optimizing AI for teamwork. Proc AAAI Conf Artif Intell. 2021;35(13):11405–14.
64. Middleton B, Bloomrosen M, Dente MA, Hashmat B, Koppel R, Overhage JM, Payne TH,
Rosenbloom ST, Weaver C, Zhang J. Enhancing patient safety and quality of care by improv-
ing the usability of electronic health record systems: recommendations from AMIA. J Am Med
Inform Assoc. 2013;20(e1):e2–8.
65. Nielsen J, Molich R. Heuristic evaluation of user interfaces. Proceedings of the SIGCHI con-
ference on human factors in computing systems. 1990. p. 249–256.
20 T. A. Cohen et al.

66. Horsky J, Kaufman DR, Oppenheim MI, Patel VL. A framework for analyzing the cognitive
complexity of computer-assisted clinical ordering. J Biomed Inform. 2003;36:4–22.
67. Ericsson KA, Simon HA. Protocol analysis: verbal reports as data. Cambridge, MA: MIT
Press; 1993.
68. Bansal G, Wu T, Zhou J, Fok R, Nushi B, Kamar E, Ribeiro MT, Weld D. Does the whole exceed
its parts? The effect of AI explanations on complementary team performance. Proceedings of
the 2021 CHI conference on human factors in computing systems. New York, NY: Association
for Computing Machinery; 2021. p. 1–16. https://doi.org/10.1145/3411764.3445717.
69. Shneiderman B. Human-centered artificial intelligence: reliable, safe & trustworthy. Int J Hum
Comput Interact. 2020;36(6):495–504.
Chapter 2
AI in Medicine: Some Pertinent History

Edward H. Shortliffe and Nigam H. Shah

After reading this chapter, you should know the answers to these questions:
• What are the roots of artificial intelligence in human history, even before the
general introduction of digital computers?
• How did computer science emerge as an academic and research discipline and
how was AI identified as a component of that revolution?
• How did a medical focus on AI applications emerge from the early general prin-
ciples of the field?
• How did the field of cognitive science influence early work on AI in Medicine
(AIM) and how have those synergies evolved to the present?
• What were the early medical applications of AI and how were they received in
the clinical and medical research communities?
• How has the focus of medical AI research and application evolved in parallel
with AI itself, and with the progress in computing power, communications tech-
nology, and interactive devices?
• To what extent are the early problems and methods developed by early AIM
researchers still relevant today? What has been lost and what has been gained?
• How have the advances in hardware and the availability of labeled data made
certain forms of AI popular? How can we combine these recent advances with
what we learned from the previous 40 years?
• How might we anticipate the further evolution of AI in medicine in light of the
way the field has evolved to date and its likely trajectory?

E. H. Shortliffe (*)
Vagelos College of Physicians and Surgeons, Columbia University, New York, NY, USA
e-mail: ted@shortliffe.net
N. H. Shah
Stanford University School of Medicine, Stanford, CA, USA
e-mail: nigam@stanford.edu

© The Author(s), under exclusive license to Springer Nature 21

Switzerland AG 2022
T. A. Cohen et al. (eds.), Intelligent Systems in Medicine and Health, Cognitive
Informatics in Biomedicine and Healthcare,
https://doi.org/10.1007/978-3-031-09108-7_2
22 E. H. Shortliffe and N. H. Shah

Introduction

The history of artificial intelligence in medicine (AIM) is intimately tied to the his-
tory of AI itself, since some of the earliest work in applied AI dealt with biomedi-
cine. In this chapter we provide a brief overview of the early history of AI, but then
focus on AI in medicine (and in human biology), providing a summary of how the
field has evolved since the earliest recognition of the potential role of computers in
the modeling of medical reasoning and in the support of clinical decision making.
The growth of medical AI has been influenced not only by the evolution of AI itself,
but also by the remarkable ongoing changes in computing and communication tech-
nologies. Accordingly, this chapter anticipates many of the topics that are covered
in subsequent chapters, providing a concise overview that lays out the concepts and
progression that are reflected in the rest of this volume.

Artificial Intelligence: The Early Years

As was discussed in Chap. 1, AI is a diverse field that addresses a wide variety of

topics regarding human intelligence and expertise, with an emphasis on how to
model and simulate these topics in computer systems. Thus studies of how human
beings reason are part of AI, but so are the creation of devices (such as robots) that
incorporate human-like features. Viewed in this framework, notions relevant to AI
emerged early in human history as people studied the workings of the human mind
or imagined creations that might duplicate those capabilities.
For example, fantastical non-human intelligent entities were imagined as far
back as Greek mythology. Hephaestus was a mythical blacksmith who manufac-
tured mechanical servants, and there were even early tales that involved the concept
of intelligent robots. But perhaps the most important early harbinger of AI was
Aristotle’s invention of syllogistic logic (a formal deductive reasoning system) in
the fourth century BC.
Mechanical inventions that attempted the creation of human-like machines are
known to have existed as early as the thirteenth century, when talking heads were
created as novelty items and Al-Jazari, an Arab inventor, designed what is believed
to be the first programmable humanoid robot (a boat carrying four mechanical
musicians, powered by water flow). There are many other examples that could be
mentioned from periods prior to the twentieth century.1

1
For more discussion, see “A Brief History of AI” at https://aitopics.org/misc/brief-history
(accessed August 13, 2022) and “History of Artificial Intelligence” at https://en.wikipedia.org/
wiki/History_of_artificial_intelligence. (accessed August 13, 2022).
2 AI in Medicine: Some Pertinent History 23

In the early twentieth century Bertrand Russell and Alfred North Whitehead
published Principia Mathematica, which revolutionized formal logic [1].
Subsequent philosophers pursued the logical analysis of knowledge. The first use of
the word “robot” in English occurred in a play by Karel Capek that was produced
in 1921.2 Thereafter a mechanical man, Electro, was introduced by Westinghouse
Electricat at the New York World’s Fair in 1939 (along with a mechanical dog
named Sparko). It was a few years earlier (1936–37) that Alan Turing proposed the
universal Turing Machine concept and proved notions of computability.3 Turing’s
analysis imagined an abstract machine that can manipulate symbols on a strip of
tape, guided by a set of rules. He showed that such a simple machine was capable
of simulating the logic of any computer algorithm that could be constructed. Also
relevant (in 1943) were the introduction of the term cybernetics, the publication by
McCulloch and Pitts of A Logical Calculus of the Ideas Immanent in Nervous
Activity (an early stimulus to the notion of artificial neural networks) [2], and
Emil Post’s proof that production systems are a general computational mecha-
nism [3].
Especially important for AI was George Polya’s 1945 book How to Solve It,
which introduced the notion of heuristic problem solving [4]—a key influential
concept in the AI community to this day. That same year Vannevar Bush published
As We May Think, which offered a remarkable vision of how, in the future, comput-
ers could assist human beings in a wide range of activities [5]. In 1950, Turing
published Computing Machinery and Intelligence, which introduced the Turing
Test as a way of defining and testing for intelligent behavior [6]. In that same year,
Claude Shannon (of information theory fame) published a detailed analysis show-
ing that chess playing could be viewed as search (Programming A Computer to
Play Chess) [7]. The dawn of computational artificial intelligence was upon us as
computers became viable and increasingly accessible devices.

Modern History of AI

The history of AI, as we think of it today, began with the development of stored-­
program digital computers and the ground-breaking work of John von Neumann
and his team at Princeton University in the 1950s. As the potential of computers

2
Čapek K. Rossumovi Univerzální Roboti (Rossum’s Universal Robots). It premiered on 25
January 1921 and introduced the word “robot” to the English language and to science fiction as a
whole. https://en.wikipedia.org/wiki/R.U.R. (accessed August 13, 2022).
3
Turing submitted his paper on 31 May 1936 to the London Mathematical Society for its
Proceedings, but it was published in early 1937. https://en.wikipedia.org/wiki/Turing_machine.
(accessed August 13, 2022).
24 E. H. Shortliffe and N. H. Shah

began to be appreciated, academic engineering scientists began to pursue concepts

that would evolve to be known as computer science. The history and capabilities of
AI have subsequently been tied to the evolution of computers and their associated
technologies.
As is mentioned in Chap. 1, it was at a conference at Dartmouth University in
1956 that a group of early computer scientists gathered to discuss the notion of
simulating human reasoning by computer. One attendee, John McCarthy from
Massachusetts Institute of Technology (MIT) (who later spent most of his pro-
fessional life at Stanford University), coined a name for the developing field:
artificial intelligence. At Carnegie Mellon University (then known as Carnegie
Tech), psychologist Allen Newell, economist/psychologist Herbert Simon, and
systems programmer (from the Rand Corporation) John Clifford Shaw intro-
duced the Logic Theorist system4—arguably the first AI program—which was
followed by their General Problem Solver in 1957.5 At about the same time
(1958), Frank Rosenblatt invented the perceptron algorithm at the Cornell
Aeronautical Laboratory [8]. This introduced the notion of connectionism in
AI, where networks of circuits or connected units were used to simulate intelli-
gent behavior.
The notion of machine learning was first explored by Arthur Samuel (IBM)
between 1958 and 1962 [9]. He developed a checker-playing program that learned
strategy and novel methods by having it mounted on two machines and then having
it play against itself thousands of times—resulting in a program that was able to
beat the world champion. Another key development during that era (1958) was John
McCarthy’s creation of the LISP programming language6—which dominated as the
basis for AI research and development for several decades (including in the medical
AI community).
During the 1960s there was an explosion in creative AI work, initially at MIT
and Carnegie Mellon, but later in the decade at other universities in the
US. International explorations of AI were also underway, especially in the United
Kingdom (where the first Machine Intelligence workshop was held in Edinburgh
in 1966). By the end of the decade, as early computer science departments began
to be formed, AI groups began to appear more broadly (with notable efforts under-
way at the University of California Berkeley and Stanford University). The first
industrial robot company was formed (1962) and a series of influential AI PhD

4
http://shelf1.library.cmu.edu/IMLS/MindModels/logictheorymachine.pdf. (accessed August
13, 2022).
5
http://bitsavers.informatik.uni-stuttgart.de/pdf/rand/ipl/P-­1 584_Report_On_A_General_
Problem-­Solving_Program_Feb59.pdf. (accessed August 13, 2022).
6
McCarthy’s original paper is available at http://www-formal.stanford.edu/jmc/recursive.html.
(accessed August 13, 2022).
2 AI in Medicine: Some Pertinent History 25

dissertations emerged—particularly at MIT where the students of Marvin Minsky

had a huge impact on the evolving field [10]. Also noteworthy was the invention
of the mouse pointing device by Doug Engelbart at Stanford Research Institute
(SRI7), which was to revolutionize the way in which human beings would interact
with computers. In 1969, also at SRI, scientists developed “Shakey”, a mobile
robot that had a problem-solver embedded in addition to locomotion (wheels) and
perception (cameras with image processing).8 The first International Joint
Conference on Artificial Intelligence (IJCAI) was held in Washington, DC in
1969. Meanwhile, that same year at MIT, Minsky and Seymour Papert published
Perceptrons [11], an influential book that discussed the computational approach
that Rosenblatt had introduced a decade earlier, outlining the limits of what per-
ceptrons could do. This led to a decrease in interest in connectionist concepts and
arguably held back the pace of development for what eventually became known as
neural networks in the 1980s and in turn led to today’s deep learning approaches
(see Chap. 6).
AI research topics in the 1960s seem remarkably similar to those that domi-
nate today. Machine learning, natural language processing, speech understand-
ing, image analysis, robotics, and simulation of human problem solving were all
major areas of research focus. Much of the funding for such research in the US
came from the Department of Defense (DOD), which envisioned eventual mili-
tary applications of AI but provided extensive support for basic methodology
development that had no immediate military application. The DOD also sup-
ported communications research, which in turn became a great facilitator of AI
development work. Perhaps most notable was the introduction of a nationwide
network for interconnecting major research computers that were located at aca-
demic institutions and in research centers for military contractors. The DOD’s
Advanced Research Projects Agency (ARPA)9 supported much of the AI and
communications research in the country. This network for research computers,
was built on the notion of packet switching and became known as the ARPA
Network or, simply, the ARPAnet. Collaborative AI research among universities
became heavily dependent on this network, and the notion of electronic messag-
ing among researchers across the various sites evolved into the email that we
take for granted today. Similarly, the ARPAnet, and its packet switching technol-
ogy, were eventually taken over by the National Science Foundation (NSF) and,

7
See https://en.wikipedia.org/wiki/Douglas_Engelbart. (accessed August 13, 2022). SRI became
an independent entity outside of Stanford University and is known today simply as SRI
International.
8
https://www.sri.com/hoi/shakey-the-robot/. (accessed August 13, 2022).
9
Also often called DARPA, for Defense Advanced Research Projects Agency.
26 E. H. Shortliffe and N. H. Shah

in turn, became a coordinated independent entity that is today known as the

Internet.

AI Meets Medicine and Biology: The 1960s and 1970s

As AI was developing as a research discipline, it is not surprising that some of the

challenging problems that attracted investigators were drawn from biomedical sci-
ence. An early example from 1965 was MIT work by Joseph Weizenbaum who was
exploring chatbot technology (conversational natural language processing and
response generation; see Chap. 9). He developed a program known as “The Doctor”,
but more affectionately referred to as “Eliza”, which attempted to provide psychiat-
ric assessments of patients. The focus was on maintaining the conversation intelli-
gently rather than actually reaching a psychiatric diagnosis. The program became a
popular, easy-to-use “toy” at AI centers since it was available for conversations over
the ARPAnet, and it did respond in ways that suggested, at some level, that it under-
stood what the user was saying. A few years later, at Stanford, a psychiatrist on the
medical school faculty, Ken Colby, worked with AI researchers to develop a conver-
sational program, known as “Parry”, that would simulate the behavior of a patient
with paranoid schizophrenia. He undertook the work largely for educational pur-
poses, and his students and residents enjoyed “interviewing” the program to learn
about its thought disorder and to try to keep the “patient” from shutting down and
refusing to communicate further. Of course, as Parry became known in the AI com-
munity, it was inevitable that people would begin to wonder how Eliza would han-
dle a therapeutic session with Parry. Accordingly, in 1972, an ARPAnet link was
created between Eliza at MIT (Cambridge, MA) and Parry at Stanford (Palo Alto,
CA). Without human intervention, the two programs had a conversation,10 and this
somewhat hysterical match-up has become part of AI lore [12].

Emergence of AIM Research at Stanford University

A more serious and ground-breaking AI research effort in biomedicine was the

Dendral Project at Stanford University. It began as an effort developed by a remark-
able scientist, Joshua Lederberg, who had been attracted to Stanford as founding
chair of their Department of Genetics in the late 1950s. He arrived shortly after
receiving the Nobel Prize in Physiology or Medicine (at age 33!) for his ground-­
breaking work, at the University of Wisconsin, on genetic transfer between bacteria.
Then, in the mid-1960s, a young researcher, Edward Feigenbaum, joined the faculty
in Stanford’s nascent computer science department, arriving from UC Berkeley

10
See https://tools.ietf.org/html/rfc439 for a transcript of the interchange between the two pro-
grams. (accessed August 13, 2022).
Random documents with unrelated
content Scribd suggests to you:
 A SUBJECT FOR SYMPATHY.
DISTRESSING POSITION OF CHARLES, WHO DOES NOT FEEL
WELL, AND WHO IS KEENLY ALIVE TO THE FACT THAT AMY IS
LOOKING AT HIM THROUGH HER OPERA-GLASS.
 LITTLE DUCKS.
Georgy. "THERE NOW, CLARA—I CALL IT VERY PEEVISH OF YOU.
YOU PROMISED ME, IF I LET YOU GO IN FIRST, THAT YOU
WOULDN'T BE LONG, AND I DECLARE YOU HAVE BEEN EXACTLY
AN HOUR AND TWENTY MINUTES."
[Pouts.

SOLICITUDE.
Wife. "NOW PROMISE ME ONE THING, ADOLPHUS. YOU WON'T
GO FLYING OVER ANY HEDGES OR FIVE-BARRED GATES?"
 AN UNCONSCIOUS VICTIM.
FEARFUL PRACTICAL JOKE PLAYED WITH A CHILD'S BALLOON
UPON A SWELL.
 A HORRID BOY.
Frank. "OH, I SAY, EMILY! AIN'T THE SEA-SIDE JOLLY?"
Emily (who is reading The Corsair to Kate). "I DO NOT KNOW,
FRANK, WHAT YOU MEAN BY JOLLY.—IT IS VERY BEAUTIFUL!—IT
IS VERY LOVELY!"
Frank. "HAH! AND DON'T IT MAKE YOU ALWAYS READY FOR YOUR
GRUB, NEITHER?"
[Exit Young Ladies, very properly disgusted.
 VERY ARTFUL CONTRIVANCE.
Clara. "WHY, DEAR ME! WHAT DO YOU WEAR YOUR HAT IN THE
WATER FOR?"
Mrs. Walrus. "OH, I ALWAYS WEAR IT WHEN I BATHE; FOR THEN,
YOU SEE, DEAR, NO ONE CAN RECOGNISE ME FROM THE BEACH!"
AT THE PLAY.
 IN A HURRY.
Boy. "NOW THEN, SIR!—THE MORE YOU LOOK THE LESS YOU'LL
LOIKE IT!—GET OVER, OR ELSE LET US COME!"
 ANGLING DELIGHTS.
ON ARRIVING AT THE BEST PART OF YOUR FISHING, YOU ARE OF
COURSE CHARMED TO FIND THAT OLD MUFFINS AND HIS LITTLE
BOY HAVE BEEN WHIPPING THE STREAM ALL THE AFTERNOON.
 THE BEARD MOVEMENT.
Young Snobley (a regular Lady-killer). "HOW THE GIRLS DO STARE
AT ONE'S BEARD! I SUPPOSE THEY THINK I'M A HORFICER JUST
COME FROM THE CRIMEAR!"
 CONSOLATION.
Young Snobley. "AH, JIM! NOBLE BIRTH MUST BE A GREAT
ADVANTAGE TO A COVE."
Jim (one of Nature's nobility). "H'M! P'RAPS!—BUT EGAD!
PERSONAL BEAUTY AIN'T A BAD SUBSTITUTE!"

AN UNDESIGNED COINCIDENCE.
TOMKINS RETIRES TO A SECLUDED VILLAGE, THAT HE MAY
GROW HIS MOUSTACHES, AND SO CUT OUT HIS ODIOUS RIVAL,
JONES. JONES, IT SO HAPPENS, HAS COME TO THE SAME PLACE
WITH THE SAME OBJECT.—FRIGHTFUL MEETING!
GOING TO THE PARK.
 A LITTLE SHOOTING IN IRELAND.
"NO HIT AGAIN, I'M AFRAID, TIM!"
"O, NIVER MIND, YER 'ONOR! SURE, YE DO IT VERY NIST.
THERE'S SOME JINTLEMEN NOW COMES, AND THEY BLAZE AWAY,
AND THEY WOWNDES THE POOR BIRRDS IN THE LIGS AND THE
WINGS, AND SUCH LIKE; BUT YER 'ONOR! O, YE FIRES, AND
FIRES, AND ALWAYS MISSES 'EM CLANE AND CLEVER!"
 LITTLE DINNER AT GREENWICH.
Fish Swell. "HERE, WAITAW!—ARE THE WHITEMEN PRETTY
GOOD?"
AFTER THE BATH.
 POETRY AND PROSE.
Blanche. "OH, IS THERE NOT, DEAR EMILY, SOMETHING,
DELICIOUS ABOUT SPRING?—WE SHALL SOON HAVE ALL THE
DEAR LITTLE BIRDS SINGING, AND THE BANKS AND THE GREEN
FIELDS COVERED WITH BEAUTIFUL FLOWERS!"
Emily. "OH, YES!—AND WITH IT WILL COME ALL THE NEW
BONNET SHAPES FROM PARIS, AND THE LOVELY NEW PATTERNS
FOR MORNING DRESSES!"
 A HAPPY NOTION.
Johnny. "OH, I SAY, GRANMA! S'POSE YOU PRETEND BEING A
LITTLE PONY, AND I RIDE ABOUT ON YOUR BACK ROUND THE
SQUARE!"—(N.B. Granma feels the heat a good deal.)

CRICKET.—CAPITAL PRACTICE FOR THE ROUND BOWLING OF THE

PERIOD.
A NICE BRACING DAY AT THE SEA-SIDE.
THE MERMAIDS' HAUNT.
 QUITE A NEW SENSATION.
Swell (on top of Omnibus). "LOOK HERE, GUS, MY BOY! SUCH A
CAPITAL I-DEAW! I RIDE UP AND DOWN FROM BAYSWATER TO
THE WHITE CHAPEL AND EAT PERIWINKLES WITH A PIN!"
 SPORTING INTELLIGENCE.
MASTER OF FOX HOUNDS PUNCH TAKES COMMAND OF HIS FIELD
IN DEFENCE OF HIS COUNTRY.
 A CAREFUL RIDER.
"A STILE, EH? AHEM! THAT'S A SORT OF THING THAT REQUIRES A
GOOD DEAL OF JUDGMENT."
 AT ALDERSHOT—A RATHER DIFFICULT MANOEUVRE.
CLARA AND HARRIET, AND THE TWO NICE GIRLS WHO ARE
STAYING WITH THEM, LED BY MAMMA, GET INTO COUSIN
HERBERT'S HUT AND HAVE LUNCH.
 FANCY SKETCH.
PORTRAIT OF THE OLD PARTY WHO RATHER LIKES ORGAN-
GRINDING.
 AN EASY MATTER.
Bathing Woman. "TEACH YER TO SWIM? LOR BLESS YER, MY
LOVE, WHY, OF COURSE I CAN!"
 A COMPLACENT BELIEF.
Old Aunt. "WELL, MY LOVE—SO YOU'VE GOT A HAT LIKE MINE, I
SEE."
A SYREN.
 AT THE SEA-SIDE.
Augustus. "ISN'T IT JOLLY, FRANK, BEING DOWN HERE FOR THE
HOLIDAYS?"
Incipient Swell. "H'M! PRETTY WELL FOR THAT. I CONFESS I MISS
THE GAIETY OF TOWN."
MR. BRIGGS TRIES HIS SHOOTING PONY.
CAMP LIFE—A NIGHT SURPRISE.
CAMP LIFE—A DAY SURPRISE.
THE CURRENCY QUESTION; OR, THE STOCK EXCHANGE OUT FOR
THE DAY.
Jones. "I SAY, BROWN, THINGS ARE DEUCED BAD IN THE CITY."
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com