41,99 €
Mehr erfahren.
- Herausgeber: John Wiley & Sons
- Kategorie: Fachliteratur
- Sprache: Englisch
Explores the transformative impact of artificial intelligence (AI) on the healthcare industry AI Doctor: The Rise of Artificial Intelligence in Healthcare provides a timely and authoritative overview of the current impact and future potential of AI technology in healthcare. With a reader-friendly narrative style, this comprehensive guide traces the evolution of AI in healthcare, describes methodological breakthroughs, drivers and barriers of its adoption, discusses use cases across clinical medicine, administration and operations, and life sciences, and examines the business models for the entrepreneurs, investors, and customers. Detailed yet accessible chapters help those in the business and practice of healthcare recognize the remarkable potential of AI in areas such as drug discovery and development, diagnostics, therapeutics, clinical workflows, personalized medicine, early disease prediction, population health management, and healthcare administration and operations. Throughout the text, author Ronald M. Razmi, MD offers valuable insights on harnessing AI to improve health of the world population, develop more efficient business models, accelerate long-term economic growth, and optimize healthcare budgets. Addressing the potential impact of AI on the clinical practice of medicine, the business of healthcare, and opportunities for investors, AI Doctor: The Rise of Artificial Intelligence in Healthcare: * Discusses what AI is currently doing in healthcare and its direction in the next decade * Examines the development and challenges for medical algorithms * Identifies the applications of AI in diagnostics, therapeutics, population health, clinical workflows, administration and operations, discovery and development of new clinical paradigms and more * Presents timely and relevant information on rapidly expanding generative AI technologies, such as Chat GPT * Describes the analysis that needs to be made by entrepreneurs and investors as they evaluate building or investing in health AI solutions * Features a wealth of relatable real-world examples that bring technical concepts to life * Explains the role of AI in the development of vaccines, diagnostics, and therapeutics during the COVID-19 pandemic AI Doctor: The Rise of Artificial Intelligence in Healthcare. A Guide for Users, Buyers, Builders, and Investors is a must-read for healthcare professionals, researchers, investors, entrepreneurs, medical and nursing students, and those building or designing systems for the commercial marketplace. The book's non-technical and reader-friendly narrative style also makes it an ideal read for everyone interested in learning about how AI will improve health and healthcare in the coming decades.
Sie lesen das E-Book in den Legimi-Apps auf:
Seitenzahl: 663
Veröffentlichungsjahr: 2024
Ähnliche
Table of Contents
Cover
Table of Contents
Title Page
Copyright Page
Dedication Page
About the Author
Foreword
Preface
Reference
Acknowledgments
Part I: Roadmap of AI in Healthcare
CHAPTER 1: History of AI and Its Promise in Healthcare
1.1 What is AI?
1.2 A Classification System for Underlying AI/ML Algorithms
1.3 AI and Deep Learning in Medicine
1.4 The Emergence of Multimodal and Multipurpose Models in Healthcare
References
CHAPTER 2: Building Robust Medical Algorithms
2.1 Obtaining Datasets That are Big Enough and Detailed Enough for Training
2.2 Data Access Laws and Regulatory Issues
2.3 Data Standardization and Its Integration into Clinical Workflows
2.4 Federated AI as a Possible Solution
2.5 Synthetic Data
2.6 Data Labeling and Transparency
2.7 Model Explainability
2.8 Model Performance in the Real World
2.9 Training on Local Data
2.10 Bias in Algorithms
2.11 Responsible AI
References
CHAPTER 3: Barriers to AI Adoption in Healthcare
3.1 Evidence Generation
3.2 Regulatory Issues
3.3 Reimbursement
3.4 Workflow Issues with Providers and Payers
3.5 Medical‐Legal Barriers
3.6 Governance
3.7 Cost and Scale of Implementation
3.8 Shortage of Talent
References
CHAPTER 4: Drivers of AI Adoption in Healthcare
4.1 Availability of Data
4.2 Powerful Computers, Cloud Computing, and Open Source Infrastructure
4.3 Increase in Investments
4.4 Improvements in Methodology
4.5 Policy and Regulatory
4.6 Reimbursement
4.7 Shortage of Healthcare Resources
4.8 Issues with Mistakes, Inefficient Care Pathways, and Non‐personalized Care
References
Part II: Applications of AI in Healthcare
CHAPTER 5: Diagnostics
5.1 Radiology
5.2 Pathology
5.3 Dermatology
5.4 Ophthalmology
5.5 Cardiology
5.6 Neurology
5.7 Musculoskeletal
5.8 Oncology
5.9 GI
5.10 COVID‐19
5.11 Genomics
5.12 Mental Health
5.13 Diagnostic Bots
5.14 At Home Diagnostics/Remote Monitoring
5.15 Sound AI
5.16 AI in Democratizing Care
References
CHAPTER 6: Therapeutics
6.1 Robotics
6.2 Mental Health
6.3 Precision Medicine
6.4 Chronic Disease Management
6.5 Medication Supply and Adherence
6.6 VR
References
CHAPTER 7: Clinical Decision Support
7.1 AI in Decision Support
7.2 Initial Use Cases
7.3 Primary Care
7.4 Specialty Care
7.5 Devices
7.6 End‐of‐Life AI
7.7 Patient Decision Support
References
CHAPTER 8: Population Health and Wellness
8.1 Nutrition
8.2 Fitness
8.3 Stress and Sleep
8.4 Population Health and Management
8.5 Risk Assessment
8.6 Use of Real World Data
8.7 Medication Adherence
8.8 Remote Engagement and Automation
8.9 SDOH
8.10 Aging in Place
References
CHAPTER 9: Clinical Workflows
9.1 Documentation Assistants
9.2 Quality Measurement
9.3 Nursing and Clinical Assistants
9.4 Virtual Assistants
References
CHAPTER 10: Administration and Operations
10.1 Providers
10.2 Payers
References
CHAPTER 11: AI Applications in Life Sciences
11.1 Drug Discovery
11.2 Clinical Trials
11.3 Medical Affairs and Commercial
References
Part III: The Business Case for AI in Healthcare
CHAPTER 12: Which Health AI Applications Are Ready for Their Moment?
12.1 Methodology
12.2 Clinical Care
12.3 Administrative and Operations
12.4 Life Sciences
References
CHAPTER 13: The Business Model for Buyers of Health AI Solutions
13.1 Clinical Care
13.2 Administrative and Operations
13.3 Life Sciences
13.4 Guide for Buyer Assessment of Health AI Solutions
References
CHAPTER 14: How to Build and Invest in the Best Health AI Companies
14.1 Barriers to Entry and Intellectual Property (IP)
14.2 Startups Versus Large Companies
14.3 Sales and Marketing
14.4 Initial Customers
14.5 Direct‐to‐Consumer (D2C)
14.6 Planning Your Entrepreneurial Health AI Journey
14.7 Assessment of Companies by Investors
References
Index
End User License Agreement
List of Illustrations
Chapter 1
FIGURE 1.1 1936–1969: Early progress, 1969–1986: AI winter, 1986: Hinton'...
FIGURE 1.2
FIGURE 1.3
FIGURE 1.4
FIGURE 1.5
FIGURE 1.6
FIGURE 1.7
FIGURE 1.8
FIGURE 1.9
FIGURE 1.10
FIGURE 1.11
FIGURE 1.12 Developing AI medical models using variety of data sources
FIGURE 1.13 Foundation Models for Generalist Medical Artificial Intelligence
Chapter 2
FIGURE 2.1 Big data in health care: Applications and challenges
FIGURE 2.2
FIGURE 2.3
FIGURE 2.4 Example of FHIR‐based EHR framework
FIGURE 2.5
FIGURE 2.6
FIGURE 2.7
FIGURE 2.8
FIGURE 2.9
FIGURE 2.10
FIGURE 2.11
FIGURE 2.12 Different types of biases in AI algorithms
FIGURE 2.13
FIGURE 2.14 Proposed Workflow for Monitoring for Bias in AI Models
FIGURE 2.15
FIGURE 2.16
FIGURE 2.17
Chapter 3
FIGURE 3.1
FIGURE 3.2 Challenges to AI Adoption in Healthcare
FIGURE 3.3
FIGURE 3.4
Chapter 4
FIGURE 4.1
FIGURE 4.2
FIGURE 4.3
FIGURE 4.4
FIGURE 4.5
FIGURE 4.6
FIGURE 4.7 Breast Mammograms in the United States
FIGURE 4.8 AI can help with the projected shortage of US clinicians
FIGURE 4.9 Projected Top 10 Applications by Value Delivered
FIGURE 4.10 Digital Twins
Chapter 5
FIGURE 5.1 AI in Radiology
FIGURE 5.2 AI allows for imaging using less radiation
FIGURE 5.3
FIGURE 5.4
FIGURE 5.5
FIGURE 5.6
FIGURE 5.7 AI assists in evaluating the characteristics of breast cancer ...
FIGURE 5.8
FIGURE 5.9 Passive monitoring allows monitoring of activities without using ...
Chapter 6
FIGURE 6.1
FIGURE 6.2
FIGURE 6.3
Chapter 7
FIGURE 7.1
FIGURE 7.2 The Different Types of Patient Date
FIGURE 7.3 Using AI for Decision Support Insights
FIGURE 7.4
FIGURE 7.5 Ten Ways Artificial Intelligence Will Transform Primary Care
Chapter 8
FIGURE 8.1 Different Types of Patient Data
FIGURE 8.2
Chapter 9
FIGURE 9.1 AI As Advisor to Patients and Providers
FIGURE 9.2
FIGURE 9.3 Microsoft Aims to Launch Clinical Ambient Intelligence
FIGURE 9.4 AI Results in Lessening of the Physician Documentation Work
FIGURE 9.5 An Example of the Workflow of a Virtual Assistant from Babylon He...
FIGURE 9.6
FIGURE 9.7 Consumer – Facing Chatbots in Healthcare Show Promise
Chapter 10
FIGURE 10.1
FIGURE 10.2
FIGURE 10.3
FIGURE 10.4
FIGURE 10.5
Chapter 11
FIGURE 11.1
FIGURE 11.2
FIGURE 11.3
FIGURE 11.4
FIGURE 11.5
FIGURE 11.6 Advances in AI will Shape the Future of Drug Discovery
FIGURE 11.7 AI will Uncover Key Insights Hidden in Current Body of Medical L...
FIGURE 11.8
FIGURE 11.9
FIGURE 11.10 Finding and Participating in Clinical Trials is a Challenging P...
FIGURE 11.11
FIGURE 11.12
FIGURE 11.13
FIGURE 11.14 AI can Help in Improving Medication Adherence
FIGURE 11.15 Aggregating and Normalizing Diverse Data Allows for Creation of...
Chapter 12
FIGURE 12.1
FIGURE 12.2
Chapter 13
FIGURE 13.1 Administrative Use Cases are High Priority for Buyers of Health ...
FIGURE 13.2
FIGURE 13.3
FIGURE 13.4 AI can Assist with Choosing the Best Next Step Based on the Pati...
FIGURE 13.5 Which AI Applications the Healthcare Stakeholders Trust Most ...
FIGURE 13.6
FIGURE 13.7
Guide
Cover Page
Table of Contents
Title Page
Copyright Page
Dedication Page
About the Author
Foreword
Preface
Acknowledgments
Begin Reading
Index
WILEY END USER LICENSE AGREEMENT
Pages
iii
iv
v
xi
xiii
xiv
xv
xvi
xvii
xix
xx
xxi
xxiii
1
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
333
334
335
336
337
338
339
340
341
342
343
344
345
AI Doctor: The Rise of Artificial Intelligence in Healthcare
A Guide for Users, Buyers, Builders, and Investors
RONALD M. RAZMI, MD
Copyright © 2024 by Ronald M. Razmi.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey.Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per‐copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750‐8400, fax (978) 750‐4470, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748‐6011, fax (201) 748‐6008, or online at http://www.wiley.com/go/permission.
Trademarks: Wiley and the Wiley logo are trademarks or registered trademarks of John Wiley & Sons, Inc. and/or its affiliates in the United States and other countries and may not be used without written permission. All other trademarks are the property of their respective owners. John Wiley & Sons, Inc. is not associated with any product or vendor mentioned in this book.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this work was written and when it is read. Neither the publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages.
For general information on our other products and services or for technical support, please contact our Customer Care Department within the United States at (800) 762‐2974, outside the United States at (317) 572‐3993 or fax (317) 572‐4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic formats. For more information about Wiley products, visit our web site at www.wiley.com.
Library of Congress Cataloging‐in‐Publication Data applied for9781394240166 (paperback); 9781394240180 (ePDF); 9781394240173 (ePUB); 9781394240197 (oBook)
Cover Design: WileyCover Images: © Kittipong Jirasukhanont/Alamy Stock Photo; Sandipkumar Patel/Getty Images
This work is dedicated to my amazing parents and brothers, whose love and support have always been the foundation upon which I've built. To friends and mentors whom I’ve learned from and who have always helped me pursue my goals.
About the Author
DR. RONALD M. RAZMI IS A CARDIOLOGIST who’s occupied many roles in healthcare and has a 360‐degree view of the practice and business of healthcare. He completed his medical training at the Mayo Clinic and was one of the pioneers in studying the applications of emerging digital technologies in managing cardiac patients. During his training, he was involved in research and published several peer‐reviewed articles in scientific journals. He is a co‐author of the Handbook of Cardiovascular Magnetic Resonance Imaging (CRC Press 2006) and launched one of the first training centers in the world for cardiologists to learn emerging technologies for managing their patients.
Ron earned his MBA from Northwestern University’s Kellogg School of Management and joined McKinsey & Company in their healthcare group, where he advised large and small companies in corporate strategy, M&A, buyout deals, and investments in emerging health technologies. In 2011, he founded Acupera, a software company that uses healthcare data and analytics to improve the management of patients with chronic diseases. Ron was the CEO of Acupera for six years and saw the challenges of bringing innovation into healthcare organizations or achieving patient adoption. Much of what he learned is covered in this book to educate innovators, buyers, and investors about the roadmap for technology adoption in healthcare. Ron has been one of the prominent voices in the digital revolution in healthcare and has written many articles and bylines, participated in interviews, and spoken at conferences.
Ron is the co‐founder and managing director of Zoi Capital, a venture capital firm that invests in digital health technologies with a focus on the applications of AI in healthcare. In 2021, he launched AI Doctor, one of the only blogs focused solely on the applications of AI in healthcare. He’s a frequent speaker at conferences, a guest on podcasts, and the author of medical and business articles.
Foreword
FROM THE EARLIEST DAYS of Alan Turing’s 1950s speculations about the possibility of computers developing some general form of intelligence, scholars have thought that healthcare would be an ideal application area for such capabilities. After all, everyone desires more effective health care, medical practice has never approached perfection, and computers and their applications have increased by nearly a billion‐fold in computational abilities over these years. So, it is not beyond the realm of imagination to think that the computer and its sophisticated machine learning abilities can form an ideal technology to help revolutionize the practice of medicine and public health. After all, if we can capture all the healthcare situations in which patients have found themselves, record the decisions their clinicians have made in their care, and evaluate the outcomes in those myriad trials, we (and our tools) should be able to learn what works best for which kinds of patients in what conditions, and perhaps even why, based on a growing understanding of the biology that underlies medicine. That knowledge should then form the basis of clinical decision support, which should help doctors, nurses, technicians, and patients themselves make optimal choices and thus lead to an era of precision medicine, achieving the best possible outcomes for each individual patient.
Despite the optimism embedded in such projections, this vision has largely not been fulfilled. Of course, the practice of medicine has improved greatly over the past seventy years, but mostly because we have developed better tools to examine the operations of the human body, a much‐improved understanding of the biological and environmental processes and genetic influences that cause disease, and highly targeted medications and procedures that can interfere with disease processes. Yet most of the traditional tasks of medicine—diagnosis, prognosis, and treatment—are still rife with uncertainty.
The vast increases in computer power have enabled widespread application of imaging tools such as CT and MRI, have contributed to sequencing and analyzing the human genome, and have made it possible to create electronic health records for billions of patients around the world, whose data had previously largely remained in inaccessible paper collections of handwritten notes. Nevertheless, despite tremendous advances in what we now call artificial intelligence (AI), we see surprisingly few successful applications of those technologies in healthcare. The ones that are successful tend to focus on very specific clinical problems, such as interpreting retinal fundus images to identify patients developing diabetic retinopathy or examining mammograms to estimate the risk of a patient developing breast cancer. And even in such narrow domains, success is often elusive.
Ron Razmi has spent the past three years wrestling with this conundrum, and in this book reviews what he has learned from that struggle. Ron is a Renaissance man of healthcare, who trained and practiced as a cardiologist, then was a consultant for one of the world’s largest consulting companies, became a digital health entrepreneur, and now serves as a venture capitalist, helping others to realize the above dreams.
My own interest in medical AI began in 1974 when I joined the computer science faculty at MIT and quickly found myself attracted to the problems of how one could represent medical knowledge in a formal way in computer programs and how those programs could use that knowledge to apply diagnostic reasoning, therapy planning, and prognoses to individual patient cases. I began to work with a cohort of brilliant doctors who were excited to formalize these ideas for a very different reason: they taught the next generation of medical students and wanted to know how to teach the best ways to reason about clinical decision‐making. They thought that capturing what they knew and how they reasoned would crystalize what they wanted to pass on to their students. My motivation was to build decision support systems that could improve the practice of most doctors by providing a built‐in “second opinion” on their challenging cases. Replacing doctors’ decision‐making was never the goal, as we realized that our systems were unlikely to be perfect, so a man‐machine team was likely to make better decisions than either alone.
At that time, most medical records were still kept as handwritten paper notes and the machine learning methods we today take for granted had not yet been developed, so our approach was to ask experienced doctors how they thought through difficult problems and then to write computer programs to simulate that thinking. Those programs, based on symbolic pattern matching or on chaining of symbolic inference rules, did a good job of handling the “textbook” cases but broke down on more complex cases where, for example, a patient suffered from an unusual combination of disorders. We tried to fix this problem by developing methods that reasoned about the pathophysiology of multiple diseases, but unfortunately, even today, with our much better understanding of biology and genetics, variations among patients and their responses to treatments are often unexplainable.
Fortunately, by the mid‐1990s, many academic medical centers had implemented electronic medical/health/patient records, so it became possible to learn relationships among symptoms, lab data, drug prescriptions, procedures, and their outcomes from such empirical data. By the end of the 2000s, government subsidy for EHR adoption had made such systems ubiquitous, today implemented in about 98% of all US hospitals. At the same time, research on machine learning, starting with the tried‐and‐true methods of statistics, had been extended to model much more complex relationships among data.
In the 2010s, a new method was developed to represent concepts as vectors in a high‐dimensional space—i.e. as long lists of numbers—derived from the frequency and nearness of co‐occurrence of the concepts in papers, medical notes, etc. Concepts whose meanings are similar to each other are often found in similar contexts, so their representations wind up near each other in that space. Furthermore, relations also take on a numerical relationship. For example, the vector of the distance and direction of the difference between the embeddings of “bone marrow” and “red blood cells” is similar to that between “adrenal gland” and “cortisol”. So, that vector is approximately a representation of “produces”. At the same time, an older idea, to build learning systems inspired by neuroscience, became practical because of the enormous improvements in computer power. To learn a model that can, say, predict the outcome of a medical intervention on a patient in a certain clinical state, we can start with a random neural network and train it on a large body of known cases of the state of a patient, the treatment, and the outcome. Initially, the network will predict randomly, but each time it makes an error, we can assign blame for that error proportionally to the parts of the network that computed the wrong outcome and adjust their influence to reduce the error. As we keep doing so, often repeatedly for thousands or even millions of cases, the error is reduced and the model becomes more accurate. The numerical representation of concepts makes this possible, so those two insightful methods now account for most machine learning approaches. Indeed, the Large Language Models that are so much in the news today are trained very much as just described, where their task is simply to predict the next word from previous ones in a vast number of human‐written (or spoken) texts.
In the past dozen or so years, therefore, many projects have succeeded in using repositories of clinical data to learn predictive models that could estimate the likelihood that a patient would develop some disease within a certain period of time, whether particular interventions were likely to cure them, how long they might live with an incurable disease, etc.
Nevertheless, much technical work remains to be done. We learned that systems built from data at one hospital might not work nearly as well at another because the patient populations differed in genetics, socioeconomic status, attitudes toward complying with medical advice, the prevalence of environmental influences in their neighborhood, etc. Medical records were often incomplete: a patient treated at multiple institutions might have records at only some of them, clerical errors dropped some data, heterogeneous vendors could have incompatible data formats that prevented their matching, etc. Clinical practice, often based on local traditions, might differ, so the tests and treatments used in different hospitals may not be the same. Most significantly, medical records do not typically record why some action was taken, yet that reasoning may be the most useful predictor of how well a patient eventually thrives. Finally, faced with a clinical situation in which two alternative therapies seem reasonable, in each case only one of them is chosen, so we have no way to know what would have happened had the other—the counterfactual therapy—been chosen.
Clinical trials address this problem by randomizing the choice of intervention, so we can know that nothing about the case has influenced the treatment choice, so its success or failure must depend only on the patient’s condition and treatment, and not on confounders that in a non‐trial context probably influenced the choice of treatment. However, most clinical decisions have not been studied by clinical trials. Also, the trials that have been done are typically done with a limited number of patients and for a limited period of time, so they may easily miss phenomena that arise rarely or only in the long term. They also tend to recruit subjects without comorbidities unrelated to the trial; thus, most of the patients whose care is ultimately determined by the outcome of the trial may not have been eligible to participate in it. Thus, clinical trials are also an imperfect method to determine the right interventions, so, despite methodological difficulties, analyzing the vast amount of non‐trial clinical data must have great value to improve healthcare.
Focusing on the technical problems of how to build the best possible models makes researchers such as me myopic. Only with decades of experience have I come to realize that a working predictive model will not necessarily be adopted and used. Institutions’ choices are driven by their own perceptions of what they need most, what they can afford to do, what they have the energy to take on, and what is consistent with their culture. These are not computational questions but fall more into areas such as management and sociology. Thus, improvement of healthcare through machine learning is a socio‐technical problem that requires insights and collaboration among many specialists.
This book is organized into three major sections. The first section (“Roadmap”) introduces the development of AI technologies over many decades and indicates the current state of the art, especially how it might be applied in healthcare. It then asks the question of how we can build robust applications and enters a very detailed discussion of what data are needed to train AI models, how these may be obtained (with difficulty), how they are processed, and how their availability and nature influence the quality of the models we can build from them. The need for robustness also raises critical questions about whether a model can be used on different populations and related issues of bias and fairness. This section then presents a bad news/good news pair of chapters, first summarizing the impediments to the vision the book outlines, followed by a more optimistic chapter describing the forces that are likely to make the vision come to fruition.
Section II (“Applications of AI in Healthcare”) examines in great detail the opportunities for AI health care applications in diagnostics, therapeutics, decision support, population health, clinical workflows, administration, and the related basic life sciences. Each subsection delves into Ron’s insights about specific clinical areas of opportunity, e.g. radiology, ophthalmology, cardiology, neurology, etc. It also details examples of particular projects and companies that are currently pursuing these opportunities, thus painting a vivid picture of where the field stands. Although the tone is overall optimistic, these discussions also cover potential problems and places where we don’t yet know how things will turn out.
The third section (“The Business Case for AI in Healthcare”) is Ron’s assessment of the business case for the possibilities described in the earlier chapters. Doctors, patients, hospitals, clinics, insurers, etc., each have a limited attention span and many competing potential projects vying for their attention. Thus, even applications that technically work are often not adopted if they are not seen as solving critical problems, doing so rapidly, while offering significant financial returns on investment. Ron uses his experience as an entrepreneur to discuss the myriad problems encountered in taking an application to successful adoption and then his experience as an investor to show the constraints on developing new applications placed by financial considerations. This section serves as a cautionary tale on what one might expect from new ventures but also includes suggestions on the kinds of applications that are ripe for contemporary exploitation, those expected to become practical in the medium term, and those whose potential is in the more distant future.
The value of this book is in explaining the technical background that makes sophisticated healthcare applications possible, surveying the field of such applications currently being fielded or developed, and describing the landscape into which any such efforts must fit. The specific projects will, of course, change rapidly as the trends outlined here make more data available to learn from in a more standardized way, as the field figures out how to conduct trials of these technologies that can convincingly pave the way for their adoption, and as the priorities of providers and payers adjust to the growing capabilities of the systems.
The underlying technologies will also change rapidly. The convolutional neural networks that underlie most machine vision and image interpretation algorithms only became practical a little over a decade ago. The large language models that drive generative AI applications have only become capable in the past two or three years and continue to develop rapidly. Pioneers of these technologies anticipate that newer methods will arise that can much more accurately identify subtle image features and generate answers to queries that are free of the “hallucinations” that sometimes plague current methods. Such technical advances, if realized, will open up new avenues of application in healthcare and will change the ROI calculations that influence adoption. Despite the consequent changes that I would expect in an edition of this book five years from now, the principles identified by Ron and the way of thinking exemplified by the analyses in this book are going to continue to be of great value.
Peter Szolovits, PhD
Professor of Computer Science and EngineeringMassachusetts Institute of TechnologyCambridge, MAJanuary 15, 2024
Preface
OVER THE COURSE OF my career as a cardiologist, consultant, CEO, and investor, I’ve always been astonished at the pace of medical innovation and how we're making rapid progress in maintaining health and treating diseases. Only 100 years ago, we didn't have anesthesia for surgeries, antibiotics for infections, cancer treatments, or the knowledge of what caused heart disease (or much else, for that matter). How did we get this far in 100 years or so when for 300,000 years, progress was negligible? Well, the industrial revolution that started about 250 years ago set in motion the kind of technological progress and wealth creation that paved the way for the medical breakthroughs we've seen over the last few decades. You need resources and know‐how to do good medical research.
Much of the progress in healthcare stems from the creation of institutions that established standards for research, private sector incentives such as patents for discoveries and inventions, and robust public sector funding. Life expectancy is way up and the pace of progress is accelerating every day. This has a ripple effect throughout the rest of the economy. When people are healthier and live longer, they're happier and more productive. This results in economic growth, which means more employment and better standards of living. It also means higher tax revenues for the government, the entity that funds most of the basic research that leads to new industries, creating future employment opportunities and improving our lives. Examples of this include the internet, the pharmaceutical industry, and now artificial intelligence (AI). According to the McKinsey Global Institute, about a third of all economic growth in advanced economies in the past 100 years can be attributed to health improvements among the global population.1
Although AI has been discussed for decades, it's only in the last decade that breakthroughs in its core methodologies have pushed it to the forefront. In healthcare, we've only recently accepted its potential to become a breakthrough technology. It was only after it started to be used in other sectors and people noticed its power to solve complex problems that innovators in healthcare realized how well‐suited it is for medical research and care delivery. After all, AI is at its most powerful when it has access to a large amount of data with many dimensions. Perhaps there's no industry with more of that than healthcare, from genetics and proteomics to epigenetics, clinical data, social data, and more!
But that's not always been the case, and until a few decades ago, we had limited data to work with. Gradually, as the number of labs we could order increased, we carried out more sophisticated radiological imaging and tissue analysis with pathology, while genetics was introduced into certain specialties. However, reviewing all of this information and putting it all together has remained a clinician's job. It will most likely remain their job, but the amount of data has exploded and keeps growing. Molecular data (and the insights it provides) introduces a whole new dimension to the practice of medicine. When you add this to data from smartphones, smartwatches and wearables, and environmental and socioeconomic data, you have a lot to analyze. Figuring out patterns in all of this data to predict, diagnose, or manage disease is beyond the capabilities of the human mind and the traditional analytical methods we've relied on.
Since this explosion of health data only started in the last couple of decades, AI is arriving at just the right moment. It's the technology of the moment and can help us to take advantage of all of this data. It can help us to make the data usable, to figure out key relationships, to make predictions, and to do things that aren't even on our radar today. There's a wealth of insights inside existing medical research that we're yet to uncover. AI will eventually help us to arrive at those insights. It will also allow us to sift through genomics and other 'omics data and uncover new insights to come up with the treatments of the future. It can transform how medicine is practiced in terms of risk analysis, decision support, better workflows, automated administration, and many other areas.
You'll notice that throughout this book, I try to separate fact from fiction. Many of these promising areas are works‐in‐progress and will take years to materialize. That shouldn't lead to disillusionment or disappointment, though. We need to accept a less‐than‐perfect start and use AI in all of these areas to figure out where it's falling short and to focus on the next round of AI research. A prime example of that is natural language processing (NLP), which will be critical for AI in healthcare because so much of the data is unstructured. However, it's currently not as advanced as machine learning algorithms for structured data. This means more manual work in the near future to abstract key concepts and data from unstructured healthcare data. It's not ideal, but that's what we have to work with for now, until NLP methodologies improve. Recent advances in large language models that underpin self‐supervised learning and generative AI (e.g. ChatGPT) are a pleasant surprise that also holds great promise in healthcare, but those applications are not yet commonplace in everyday practice of medicine.
This book is separated into three parts. In the first part, I focus on defining AI and examining its history, its promise in healthcare, the drivers and barriers of its applications in healthcare, and the data issues that will ultimately be the key factor in fulfilling its promise. The next part is about its applications to date and where it's poised to make the biggest impact. The last part is about the business of AI in healthcare. It will examine which applications are ready today from a methodological point of view to solve mission‐critical issues for the buyers, as well as which use cases can see short‐ and medium‐term demand from those same buyers: health systems, life science companies, long‐term care facilities, telemedicine companies, etc. Lastly, we examine the key issues that every entrepreneur and investor needs to analyze before embarking on building or investing in a health AI company. Ultimately, the business case for the innovators and buyers needs to be established before adoption accelerates. Having been both a physician and a digital technology entrepreneur and investor in healthcare, I've experienced how issues like workflow implications, data ownership, and reimbursement impact the bottom line and can make or break new innovations.
I'm optimistic about what lies ahead in healthcare. I see great progress ahead but make no unrealistic or wildly optimistic predictions. The state of data in healthcare is chaotic and there are powerful technical and business barriers to the adoption of these technologies. As such, my focus is to lay out the promise, discuss the drivers and barriers, and examine how innovators can successfully navigate them to fulfill that promise.
I deliberately wrote this book to cover technical, clinical, and business issues that will be critical for this technology to fulfill its immense promise in healthcare. It is meant to be technical and clinical enough for those who are involved in the field and want to better understand the current state of affairs, what's holding back faster adoption, which technologies or use cases are ready now, and how they can create or invest in a successful business model on either side of the commercial transaction. On the other hand, I've tried to use plain language and to avoid getting too technical so that those who are trying to familiarize themselves with this field can follow along.
I hope you find it useful during your journey in this exciting field.
Ronald M. Razmi, MD
New YorkJanuary 15, 2024
Reference
1. Remes, J., Linzer, K., Singhal, S. et al. (2022, August 4).
Prioritizing Health: A Prescription for Prosperity
. McKinsey & Company
https://www.mckinsey.com/industries/healthcare‐systems‐and‐services/our‐insights/prioritizing‐health‐a‐prescription‐for‐prosperity
.
Acknowledgments
MANY PEOPLE HELPED IMPROVE this book. They reviewed and edited each chapter. I'm extremely grateful to each one of them for their contributions. Special thanks to Pete Szolovits, PhD, Seymour Dunker, Matthew Kittay, JD, Eli Khedouri, Kang Zhang, MD, Justin Graham, MD, Greg Wolff, Gerasimos Petratos, MD, and Alex Fair.
Part IRoadmap of AI in Healthcare
CHAPTER 1History of AI and Its Promise in Healthcare
ARTIFICIAL INTELLIGENCE TECHNOLOGY has been “around” for some 80 years. Although it's gained significant traction recently and its applications are transforming almost every industry, its foundations date back to the Second World War. Alan Turing is considered one of the pioneers in computing and artificial intelligence (AI), and his papers from the time reveal some of the earliest mentions of machines that could mimic the human mind and its ability to reason. Back then, scientists were beginning to build computer systems that were meant to process information in the same way as the human brain. A paper by Warren McCulloch and Walter Pitts, who proposed an artificial neuron as early as 1943, refers to a computational model of the “nerve net” in the brain and is one of the first mentions of the topic in scientific literature.1 This laid the foundation for the first wave of research into the topic.
In the 1950s, in his paper Computing Machinery and Intelligence, Turing offered a framework for building intelligent machines and methods for testing their intelligence. Bernard Widrow and Ted Hoff at Stanford University developed a neural network application for reducing noise in phone lines in the late 1950s.2 Around the same time, Frank Rosenblatt built the Perceptron while working as an academic and with the US government.3 The Perceptron was based on the concept of a neural network and was meant to be able to perform tasks that humans usually performed, such as recognizing images and even walking and talking. However, the Perceptron only had a single layer of neurons and was limited in how much it could do. Marvin Minsky, a colleague of Rosenblatt's and an old high school classmate of his, and Seymour Papert wrote a book called Perceptrons: An Introduction to Computational Geometry in which he detailed the limitations of the Perceptron and neural networks. This led to an AI winter that lasted until the mid‐1980s.4
By 1986, there was renewed interest in neural networks by physicists who were coming up with novel mathematical techniques. There was also a landmark paper by Geoffrey Hinton about the applications of back propagation in neural networks as a way to overcome some of their limitations, although some practitioners point to a Finnish mathematician, Seppo Linnainmaa, as having invented back propagation in the 1960s.5, 6 This led to a revival of the field and the creation of some of the first practical applications, such as detecting fraud in credit card transactions.
The 1986 paper by Geoffrey Hinton and few others in Canada highlighted the potential of multi‐layered neural networks.5 This, along with the potential it subsequently showed in speech recognition, re‐ignited interest in neural networks and was enough to start attracting interest and research dollars. At Carnegie Melon in the late 1980s, Dean Pomerleau built a self‐driving car using neural networks. NYU's Yann LeCun started using neural networks for image recognition and released a paper in 1998 that introduced convolutional neural networks (CNNs) as a way to mimic the human visual cortex.7 In parallel, John Hopfield popularized the Hopfield network, which was the first recurrent neural network (RNN).8 This was expanded upon by Jurgen Schmidhuber and Sepp Hochreiter in 1997 with the introduction of long short‐term memory, which greatly improved the efficiency and practicality of RNNs.9 Although these applications in the 1980s and 1990s created momentum for the field, they soon reached their limits due to a lack of data and insufficient computing power. There was another AI winter—fortunately, a shorter one—for about a decade or so.
As computing power and the amount of data increased, over the next few years, companies like Microsoft and Google ramped up their research in the field dramatically. In 2012, Hinton and two of his students highlighted the power of deep learning (DL) when they obtained significant results in the well‐known ImageNet competition, based on a dataset collated by Fei‐Fei Li and others.10
Seymour Duncker, CEO of Mindscale.ai, who's developed many AI‐based products in healthcare and other sectors, told me, “Fei‐Fei’s work that led to ImageNet was a watershed moment that changed how the field of AI came to view data and its key role for making progress in the field of AI. Before ImageNet, the predominant thinking was that a better algorithm would make better decisions, regardless of the data. Fei‐Fei had the key insight that even the best algorithms wouldn’t be able to perform if the data they operated on wasn’t representative of the world. That’s why she went out and curated a large dataset representing visual objects found in the world and then made it open source. The ImageNet competition spawned a range of vision models that took the accuracy of classifying objects from 71.8% to 97.3%, beating human abilities. Essentially, Fei‐Fei showed that big data leads to better decisions.”
At the same time as Hinton's work in 2012, Jeffrey Dean and Andrew Ng were doing breakthrough work on large‐scale image recognition at Google Brain.11 DL also enhanced the existing field of reinforcement learning, thanks largely to researchers like Richard Sutton, leading to the game‐playing successes of systems developed by DeepMind.12 Given the impressive results that this algorithm showed the world, everyone woke up to the potential of DL and neural networks. In 2014, Ian Goodfellow published his paper on generative adversarial networks (GANs), which along with reinforcement learning has become the focus of much of the recent research in the field.13
This led to an initiative at Stanford University called the One Hundred Year Study on Artificial Intelligence. The study was founded by Eric Horvitz and aimed to build on the existing research that Horvitz and his team had carried out at Microsoft. The rest, as they say, is history! Progress came fast and in various sectors thanks to large amounts of digitized data and significantly stronger computers. Figure 1.1 summarizes some of the key milestones in the evolution of AI.
Today, we're at the dawn of AI's true potential. Most of the applications to date use supervised learning, which teaches algorithms by giving them annotated data (thousands or millions of records) and allowing the algorithm to learn from that data to identify patterns and make predictions. The long‐term power of neural nets will reside in unsupervised learning, where the algorithms can learn without being trained on annotated data but by just being given the data. Generative AI is surprising everyone with its rapid progress and the level of sophistication it's showing in generating text, images, voice, art, etc. Reinforcement learning, which mimics human learning mechanisms, will also lead the charge toward AI's full potential.15
1.1 What is AI?
AI isn't magic, and nor is it going to spark a robot uprising or replace your doctor entirely. Mathematical terms like machine learning (ML) and DL are used as easy ways to explain statistical computer algorithms that use data to identify patterns and make accurate predictions. The term AI is used to refer to a range of technologies that can be combined in different ways to sense, comprehend, and act, as well as to learn from experience and to adapt over time (Figure 1.2).16
As an umbrella term, AI can refer to both natural language processing (NLP) and ML. NLP powers translations, understanding meaning in written or spoken language, pattern recognition, as well as smart assistants like Google Assistant, Siri, and Alexa. ML is one of the most exciting areas of AI and uses computational approaches to make sense of data and to provide insights into what that data shows. It's a dynamic, iterative process that uses examples and experiences as opposed to predefined rules (Figure 1.3). With ML, instead of providing a set of rules that define what a cat looks like, the operator provides the algorithm with a bunch of cat photos and leaves the software to arrive at its own conclusions. ML allows computers to retain information and to get smarter over time, just like human beings. But unlike human beings, these algorithms aren't susceptible to sleep deprivation, distractions, information overload, and short‐term memory loss. That's why this powerful technology is so exciting.18
FIGURE 1.1
(source: original research14):
1936–1969: Early progress, 1969–1986: AI winter, 1986: Hinton's paper on back propagation in neural networks, 1997–2012: Progress in AI methodologies: 1997 IBM beat Kasperov, 2007 ImageNet, 2011 IBM beat Jeopardy, 2012–Present: Rapid progress in deep‐learning applications
FIGURE 1.2
(source: Accenture)16
FIGURE 1.3
(source: HIMSS 2019)17
Then there's cognitive computing, an approach that uses computers to simulate human understanding, reasoning, and thought‐processing.
The thing to remember is that AI isn't like other software and you can't just install it and get going. Instead, it's a sophisticated combination of interrelated technologies that needs to be customized for each application. Done properly, it can process data and come to its own conclusions about what that data means, and it can then carry out a set of actions based on what it's learned. It's true that AI can mimic the human brain, but it can also outperform us mere humans by discovering complex patterns that no human being could ever process and identify.
And not all AI applications are created equally. Some of the simpler use cases for AI include chatbots and automated phone screeners, which are able to provide basic responses to voice or text inputs. More complex AI algorithms can process unfathomably large sets of data to discover underlying trends and to answer questions ranging from which film a Netflix subscriber is most likely to enjoy to which treatment option is best for any given patient. They work by making predictions on how to interpret what's happening and how best to respond (Figure 1.4).
ML is a subcategory of AI in which algorithms and statistical models are used to allow computers to perform tasks. The interesting thing about ML is that it goes above and beyond classic rules‐based approaches and instead taps into inference and patterns to address sophisticated challenges. It's able to tap into DL and other complex mathematical techniques to parse large datasets and make predictions. As time goes on, the algorithm figures out which patterns and approaches will deliver the best results and uses that to adapt itself for the future.
FIGURE 1.4
(source: IQVIA)19
In general, the more data that they have access to, the more ML algorithms can improve themselves. Recent technological advances in big data have made it more practical to apply ML to massive datasets to derive insights that were previously impossible to detect.
We typically use ML to build inference tools where we find patterns in existing data that allow us—when presented with new data—to infer something interesting about that data, such as recognizing abnormalities in an MRI image.20 The great thing about ML is that it doesn't require interference or human intuition because it's driven entirely by data.
Neural networks are a subcategory of ML that simulates connected artificial neurons. This allows them to model themselves on the natural neurons and the way that they interact in the human brain. Computational models that are inspired by neural connections have been studied since the 1950s and are more relevant than ever before because processing power has continued to increase and we have access to even larger training datasets that can allow algorithms to analyze input data such as images, video, and speech.
AI practitioners refer to these techniques as DL, since neural networks have many (“deep”) layers of simulated, interconnected neurons.19 DL is a branch of ML that's emerged in the last decade as a breakthrough technology and its applications hold significant promise in the coming decades. DL is possible due to the architecture and design of neural networks, which mimic the neural connections in the human brain.
Artificial neural networks (ANNs) are clusters of interconnected nodes, like brain neurons. ANNs with multiple layers of connected nodes are able to take advantage of DL. They typically also use convolutional layers, which take the input data and group it together into blocks. These blocks are then fed into multiple deep processing layers, which filter them and then feed that filtered data into further layers with more filters. They identify features that are inherent to the original data and combine those features to produce a hierarchic estimation of patterns, called concepts.
Seymour Duncker explains, “These hierarchic estimations are represented by numerical weights, each of which represents the relative importance of two nodes in the network. A trained AI model is represented by a matrix of these numerical weights which gets combined with live data to produce a prediction.” That might sound complicated, but that's what we humans do every day. When machines do this, it's called AI.
Neural networks run datasets through what the ML community calls a black box. This black box represents a series of mathematical calculations and statistical computations that are far beyond our understanding. Learning typically happens by feeding back an error signal from incorrect predictions, to alter the myriad weights the network uses to compute an answer from the inputs. The actual decision‐making process can't usually be traced from beginning to end in a way that can be easily understood.
Neural networks and DL can and should be used as tools to aid the medical community. But these algorithms and this technology won't replace the expertise of medical professionals. Human experts will always be needed to train these algorithms and to identify the unique situations that exceed the scope of what can easily be defined.
Here are the four different types of neural networks and their benefits (Figure 1.5):
Feed‐forward neural networks:
This is the simplest type of ANN in which information moves in only one direction (forward) with no loops in the network. Information typically starts out from the input layer and moves through the hidden layers to the output layer. The first single‐neuron network was proposed by AI pioneer Frank Rosenblatt back in 1958.
Recurrent neural networks (RNNs):
These are ANNs with connections between the neurons, which are perfectly suited for processing sequential data like speech and language. RNNs are composed of an additional hidden state vector that contains “memory” about the data that it's observed over time. As Seymour Duncker told me, “Recurrent neural networks have largely been replaced by the transformer architecture, which provides a more effective way of representing patterns in sequences and enables parallel processing, allowing algorithms to process extremely large volumes of data more efficiently than the original RNNs.”
FIGURE 1.5
(source: McKinsey Global Institute analysis/https://www.mckinsey.com/featured‐insights/artificial‐intelligence/notes‐from‐the‐ai‐frontier‐applications‐and‐value‐of‐deep‐learning)
Convolutional neural networks (CNNs):
CNNs are ANNs where the connections between neural layers are inspired by biological visual cortexes, the portion of the brain that processes images. This makes them ideal for perceptual tasks.
Generative adversarial networks (GANs)
: This approach uses two neural networks that compete against one other in a zero‐sum game framework (which is why we call them “adversarial”). GANs can learn to mimic various distributions of data (such as text, speech, and images), which makes them useful for generating test datasets when they're not readily available. This is where we're seeing the exciting new developments with ChatGPT and other generative AI and possibly the foundation for the next breakthroughs in AI capabilities.
I mentioned earlier about the astounding results that generative AI is showing. You could say that generative AI is a branch of AI that involves creating new content. This new content can be in the form of audio, code, images (art!), text, and videos. Recent advancements in the field mean that AI is on its way to being one step ahead of interpreting data and to start creating the types of content that have previously been only the domain of human beings.
Generative AI is a form of ML algorithm that uses self‐supervised learning to train on large amounts of historic data. For example, with ChatGPT, which generates text, the training involved feeding the model a massive amount of text from the internet so that it was able to generate predictions. It's a breakthrough in that rather than simply perceiving and classifying data, ML is now able to create content on demand. Given their ability to do self‐supervised learning and not require labeled training data (as in supervised learning), the models can be trained on large datasets much faster. For example, a large language model can be trained on historic medical literature and be ready to answer medical questions in short order. Up until recently, our thinking has been that the medical literature needs to be annotated and key concepts abstracted in order to use ML models to extract insights.
Generative AI outputs are carefully calibrated combinations of the data used to train the algorithms. Because the amount of data used to train these algorithms is so large, the models can appear to be “creative” when producing outputs. What's more, the models usually have random elements, which means they can produce a variety of outputs from one input request—making them seem even more lifelike.21
In healthcare, the initial applications of generative AI could include generating images for training radiology models, creating clinical content for educating patients about their conditions, answering general questions about various conditions, carrying out administrative tasks such as medical coding and documentation, supporting mental health services, and helping with research.
Now, you may be wondering how AI, ML, DL, and other related technologies are different from statistics or analytics. The short answer is that the new methods are better suited for large volumes of data and high‐dimensional data like pictures, video, and audio files. Also, AI is better suited for data where we're not aware of all of the complex relationships within the dataset and so we can't provide specific instructions to the model. AI is particularly effective at uncovering hidden relationships in the datasets, uncovering patterns, and making predictions.
Classical statistics methods require more input from humans about the variables, the relationship between the variables, and the desired outcome. Traditional statistical analyses aim to find inferences about sample or population parameters, while ML aims to algorithmically represent the data structure and to make predictions or classifications. ML typically starts out with fewer assumptions about the underlying data than traditional statistical analysis and results in algorithms that are much more accurate when it comes to prediction and classification (Figure 1.5).
Healthcare analytics have traditionally been rooted in understanding a given set of data by using business intelligence‐focused tools.22 The people using those tools are typically analysts, statisticians, and business users as opposed to engineers. The problem with traditional enterprise data analytics is that you don't learn from the data, you only understand what's in it. To learn from it, you need to use ML and effective feedback loops from all of the relevant stakeholders. This will help you to uncover hidden patterns in the data, especially when there are non‐linear relationships that aren't easily identifiable to humans.
For DL models to get good at classification and to perform at the same level as humans, they require thousands—or in some cases, millions—of data records. One estimate found that supervised DL algorithms generally achieve acceptable performance with around 5000 labeled examples per category and match or exceed human performance when trained with at least 10 million.23 In some cases, so much data is available (often millions or billions of rows per dataset) that AI is the most appropriate technique. However, if a threshold of data volume isn't reached, AI might not add value to traditional analytics techniques.
It can be difficult to source these massive datasets, and labeling remains a challenge. Most current AI models are trained with supervised learning, which requires humans to label and categorize the underlying data. However, promising new techniques are emerging to overcome these data bottlenecks, such as large language models, reinforcement learning, transfer learning, and “one‐shot learning,” which allows a trained AI model to learn about a subject based on a small number of real‐world demonstrations or examples—and sometimes just one.
Neural AI techniques are particularly good at analyzing images, videos, and audio data because they're complex and multidimensional. AI practitioners often call this “high dimensionality.” Neural networks are well suited to high dimensionality because multiple layers in a network can learn to represent the many different features that are present in the data. For example, with facial recognition, the first layer in the network could focus on raw pixels, the next on edges and lines, another on generic facial features, and the final layer on identifying the face (Figure 1.6
