Main Machine Learning Applications Using Python: Cases Studies from Healthcare, Retail, and Finance
Machine Learning Applications Using Python: Cases Studies from Healthcare, Retail, and FinancePuneet Mathur
Gain practical skills in machine learning for finance, healthcare, and retail. This book uses a hands-on approach by providing case studies from each of these domains: you’ll see examples that demonstrate how to use machine learning as a tool for business enhancement. As a domain expert, you will not only discover how machine learning is used in finance, healthcare, and retail, but also work through practical case studies where machine learning has been implemented.
Machine Learning Applications Using Python is divided into three sections, one for each of the domains (healthcare, finance, and retail). Each section starts with an overview of machine learning and key technological advancements in that domain. You’ll then learn more by using case studies on how organizations are changing the game in their chosen markets. This book has practical case studies with Python code and domain-specific innovative ideas for monetizing machine learning.
What You Will Learn
• Discover applied machine learning processes and principles
• Implement machine learning in areas of healthcare, finance, and retail
• Avoid the pitfalls of implementing applied machine learning
• Build Python machine learning examples in the three subject areas
Who This Book Is For
Data scientists and machine learning professionals.
Machine Learning Applications Using Python is divided into three sections, one for each of the domains (healthcare, finance, and retail). Each section starts with an overview of machine learning and key technological advancements in that domain. You’ll then learn more by using case studies on how organizations are changing the game in their chosen markets. This book has practical case studies with Python code and domain-specific innovative ideas for monetizing machine learning.
What You Will Learn
• Discover applied machine learning processes and principles
• Implement machine learning in areas of healthcare, finance, and retail
• Avoid the pitfalls of implementing applied machine learning
• Build Python machine learning examples in the three subject areas
Who This Book Is For
Data scientists and machine learning professionals.
You may be interested in
Most frequently terms
Machine Learning Applications Using Python Cases Studies from Healthcare, Retail, and Finance — Puneet Mathur Machine Learning Applications Using Python Cases Studies from Healthcare, Retail, and Finance Puneet Mathur Machine Learning Applications Using Python: Cases Studies from Healthcare, Retail, and Finance Puneet Mathur Bangalore, Karnataka, India ISBN-13 (pbk): 978-1-4842-3786-1 https://doi.org/10.1007/978-1-4842-3787-8 ISBN-13 (electronic): 978-1-4842-3787-8 Library of Congress Control Number: 2018965933 Copyright © 2019 by Puneet Mathur This work is subject to copyright. All rights are reserved 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. Trademarked names, logos, and images may appear in this book. Rather than use a trademark symbol with every occurrence of a trademarked name, logo, or image we use the names, logos, and images only in an editorial fashion and to the benefit of the trademark owner, with no intention of infringement of the trademark. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Managing Director, Apress Media LLC: Welmoed Spahr Acquisitions Editor: Celestin Suresh John Development Editor: Matthew Moodie Coordinating Editor: Aditee Mirashi Cover designed by eStudioCalamar Cover image designed by Freepik (www.freepik.com) Distributed to the book trade worldwide by Springer Science+Business Media New York, 233 Spring Street, 6th Floor, New York, NY 10013. Phone 1-800-SPRINGER, fax (201) 348-4505, e-mail email@example.com, or visit www.springeronline.com. Apress Media, LLC is a California LLC and the sole member (owner) is Springer Science + Business Media Finance Inc (SSBM Finance Inc). SSBM Finance Inc is a Delaware corporation. For information on translations, please e-mail firstname.lastname@example.org, or visit http://www.apress.com/ rights-permissions. Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook versions and licenses are also available for most titles. For more information, reference our Print and eBook Bulk Sales web page at http://www.apress.com/bulk-sales. Any source code or other supplementary material referenced by the author in this book is available to readers on GitHub via the book’s product page, located at www.apress.com/978-1-4842-3786-1. For more detailed information, please visit http://www.apress.com/source-code. Printed on acid-free paper I dedicate this book in the lotus feet of Divine Mother. Table of Contents About the Author����������������������������������������������������������������������������������������������������� xi About the Technical Reviewer������������������������������������������������������������������������������� xiii Acknowledgments���������������������������������������������������������������������������������������������������xv Introduction�����������������������������������������������������������������������������������������������������������xvii Chapter 1: Overview of Machine Learning in Healthcare����������������������������������������� 1 Installing Python for the Exercises������������������������������������������������������������������������������������������������ 2 Process of Technology Adoption���������������������������������������������������������������������������������������������� 2 How Machine Learning Is Transforming Healthcare���������������������������������������������������������������� 8 End Notes������������������������������������������������������������������������������������������������������������������������������������ 11 Chapter 2: Key Technological advancements in Healthcare����������������������������������� 13 Scenario 2025����������������������������������������������������������������������������������������������������������������������������� 13 Narrow vs. Broad Machine Learning������������������������������������������������������������������������������������������� 14 Current State of Healthcare Institutions Around the World��������������������������������������������������������� 16 Importance of Machine Learning in Healthcare�������������������������������������������������������������������� 19 End Notes������������������������������������������������������������������������������������������������������������������������������������ 34 Chapter 3: How to Implement Machine Learning in Healthcare����������������������������� 37 Areas of Healthcare Research Where There is Huge Potential���������������������������������������������������� 37 Common Machine Learning Applications in Radiology��������������������������������������������������������������� 40 Working with a Healthcare Data Set������������������������������������������������������������������������������������������� 41 Life Cycle of Machine Learning Development����������������������������������������������������������������������� 41 Implementing a Patient Electronic Health Record Data Set�������������������������������������������������������� 44 Detecting Outliers������������������������������������������������������������������������������������������������������������������ 52 Data Preparation�������������������������������������������������������������������������������������������������������������������� 67 End Notes������������������������������������������������������������������������������������������������������������������������������������ 75 v Table of Contents Chapter 4: Case Studies in Healthcare AI��������������������������������������������������������������� 77 CASE STUDY 1: Lab Coordinator Problem����������������������������������������������������������������������������������� 78 CASE STUDY 2: Hospital Food Wastage Problem���������������������������������������������������������������������� 100 Chapter 5: Pitfalls to Avoid with Machine Learning in Healthcare����������������������� 121 Meeting the Business Objectives���������������������������������������������������������������������������������������������� 122 This is Not a Competition, It is Applied Business!��������������������������������������������������������������������� 123 Don’t Get Caught in the Planning and Design Flaws����������������������������������������������������������������� 126 Choosing the Best Algorithm for Your Prediction Model����������������������������������������������������������� 129 Are You Using Agile Machine Learning?������������������������������������������������������������������������������������ 130 Ascertaining Technical Risks in the Project������������������������������������������������������������������������������ 131 End Note����������������������������������������������������������������������������������������������������������������������������������� 134 Chapter 6: Monetizing Healthcare Machine Learning������������������������������������������� 135 Intro-Hospital Communication Apps����������������������������������������������������������������������������������������� 135 Connected Patient Data Networks�������������������������������������������������������������������������������������������� 140 IoT in Healthcare����������������������������������������������������������������������������������������������������������������������� 142 End Note����������������������������������������������������������������������������������������������������������������������������������� 145 Chapter 7: Overview of Machine Learning in Retail��������������������������������������������� 147 Retail Segments������������������������������������������������������������������������������������������������������������������������ 149 Retail Value Proposition������������������������������������������������������������������������������������������������������������ 151 The Process of Technology Adoption in the Retail Sector��������������������������������������������������������� 153 The Current State of Analytics in the Retail Sector������������������������������������������������������������������� 155 Chapter 8: Key Technological Advancements in Retail����������������������������������������� 159 Scenario 2025��������������������������������������������������������������������������������������������������������������������������� 159 Narrow vs Broad Machine Learning in Retail���������������������������������������������������������������������������� 161 The Current State of Retail Institutions Around the World��������������������������������������������������������� 162 Importance of Machine Learning in Retail�������������������������������������������������������������������������������� 164 Research Design Overview:������������������������������������������������������������������������������������������������������ 170 Data Collection Methods����������������������������������������������������������������������������������������������������������� 170 vi Table of Contents Data Analysis���������������������������������������������������������������������������������������������������������������������������� 171 Ethical Considerations�������������������������������������������������������������������������������������������������������������� 171 Limitations of the Study������������������������������������������������������������������������������������������������������������ 171 Examining the Study����������������������������������������������������������������������������������������������������������������� 172 Phases of Technology Adoption in Retail, 2018������������������������������������������������������������������� 179 End Notes���������������������������������������������������������������������������������������������������������������������������������� 181 Chapter 9: How to Implement Machine Learning in Retail����������������������������������� 183 Implementing Machine Learning Life Cycle in Retail���������������������������������������������������������������� 185 Unsupervised Learning�������������������������������������������������������������������������������������������������������� 186 Visualization and Plotting���������������������������������������������������������������������������������������������������� 190 Loading the Data Set����������������������������������������������������������������������������������������������������������� 193 Visualizing the Sample Data Set������������������������������������������������������������������������������������������ 198 Feature Engineering and Selection������������������������������������������������������������������������������������� 201 Visualizing the Feature Relationships���������������������������������������������������������������������������������� 204 Sample Transformation������������������������������������������������������������������������������������������������������� 206 Outlier Detection and Filtering��������������������������������������������������������������������������������������������� 207 Principal Component Analysis��������������������������������������������������������������������������������������������� 210 Clustering and Biplot Visualization Implementation������������������������������������������������������������ 212 End Notes���������������������������������������������������������������������������������������������������������������������������������� 216 Chapter 10: Case Studies in Retail AI������������������������������������������������������������������� 217 What Are Recommender Systems?������������������������������������������������������������������������������������������� 217 CASE STUDY 1: Recommendation Engine Creation for Online Retail Mart�������������������������������� 218 CASE STUDY 2: Talking Bots for AMDAP Retail Group��������������������������������������������������������������� 233 End Notes���������������������������������������������������������������������������������������������������������������������������������� 237 Chapter 11: Pitfalls to Avoid With Machine Learning in Retail����������������������������� 239 Supply Chain Management and Logistics��������������������������������������������������������������������������������� 239 Inventory Management������������������������������������������������������������������������������������������������������������� 241 Customer Management������������������������������������������������������������������������������������������������������������� 242 vii Table of Contents Internet of Things���������������������������������������������������������������������������������������������������������������������� 245 End Note����������������������������������������������������������������������������������������������������������������������������������� 247 Chapter 12: Monetizing Retail Machine Learning������������������������������������������������� 249 Connected Retail Stores������������������������������������������������������������������������������������������������������������ 249 Connected Warehouses������������������������������������������������������������������������������������������������������������� 252 Collaborative Community Mobile Stores����������������������������������������������������������������������������������� 254 End Notes���������������������������������������������������������������������������������������������������������������������������������� 257 Chapter 13: Overview of Machine Learning in Finance���������������������������������������� 259 Financial Segments������������������������������������������������������������������������������������������������������������������ 261 Finance Value Proposition��������������������������������������������������������������������������������������������������������� 262 The Process of Technology Adoption in the Finance Sector������������������������������������������������������ 265 End Notes���������������������������������������������������������������������������������������������������������������������������������� 270 Chapter 14: Key Technological Advancements in Finance����������������������������������� 271 Scenario 2027��������������������������������������������������������������������������������������������������������������������������� 271 Narrow vs Broad Machine Learning in Finance������������������������������������������������������������������������ 272 The Current State of Finance Institutions Around the World����������������������������������������������������� 274 Importance of Machine Learning in Finance����������������������������������������������������������������������������� 274 Research Design Overview������������������������������������������������������������������������������������������������������� 280 Data Collection Methods����������������������������������������������������������������������������������������������������������� 281 Data Analysis���������������������������������������������������������������������������������������������������������������������������� 281 Ethical Considerations�������������������������������������������������������������������������������������������������������������� 282 Limitations of the Study������������������������������������������������������������������������������������������������������������ 282 Examining the Study����������������������������������������������������������������������������������������������������������������� 282 Phases of Technology Adoption in Finance, 2018�������������������������������������������������������������������� 290 End Notes���������������������������������������������������������������������������������������������������������������������������������� 292 viii Table of Contents Chapter 15: How to Implement Machine Learning in Finance������������������������������ 295 Implementing Machine Learning Life Cycle in Finance������������������������������������������������������������ 297 Starting the Code����������������������������������������������������������������������������������������������������������������� 299 Feature Importance������������������������������������������������������������������������������������������������������������� 304 Looking at the Outliers�������������������������������������������������������������������������������������������������������� 306 Preparing the Data Set�������������������������������������������������������������������������������������������������������� 309 Encoding Columns��������������������������������������������������������������������������������������������������������������� 312 Splitting the Data into Features������������������������������������������������������������������������������������������� 313 Evaluating Model Performance������������������������������������������������������������������������������������������� 313 Determining Features���������������������������������������������������������������������������������������������������������� 321 The Final Parameters���������������������������������������������������������������������������������������������������������� 324 End Note����������������������������������������������������������������������������������������������������������������������������������� 324 Chapter 16: Case Studies in Finance AI���������������������������������������������������������������� 325 CASE STUDY 1: Stock Market Movement Prediction����������������������������������������������������������������� 325 Questions for the Case Study���������������������������������������������������������������������������������������������� 327 Proposed Solution for the Case Study��������������������������������������������������������������������������������� 328 CASE STUDY 2: Detecting Financial Statements Fraud������������������������������������������������������������� 347 Questions for the Case Study���������������������������������������������������������������������������������������������� 349 Discussion on Solution to the Case Study:�������������������������������������������������������������������������� 349 End Notes���������������������������������������������������������������������������������������������������������������������������������� 354 Chapter 17: Pitfalls to Avoid with Machine Learning in Finance�������������������������� 355 The Regulatory Pitfall���������������������������������������������������������������������������������������������������������������� 355 Government Laws and an Administrative Controller, the Securities and Trade Commission (SEC)������������������������������������������������������������������������������������������������������ 358 States Laws and Controllers������������������������������������������������������������������������������������������������ 358 Self-Regulatory Organization���������������������������������������������������������������������������������������������� 359 The Data Privacy Pitfall������������������������������������������������������������������������������������������������������������� 360 End Note����������������������������������������������������������������������������������������������������������������������������������� 362 ix Table of Contents Chapter 18: Monetizing Finance Machine Learning��������������������������������������������� 363 Connected Bank������������������������������������������������������������������������������������������������������������������������ 363 Fly-In Financial Markets����������������������������������������������������������������������������������������������������������� 367 Financial Asset Exchange��������������������������������������������������������������������������������������������������������� 369 End Note����������������������������������������������������������������������������������������������������������������������������������� 372 Index��������������������������������������������������������������������������������������������������������������������� 373 x About the Author Puneet Mathur Advisory Board Member & Senior Machine Learning Consultant. Puneet is an experienced hands-on machine learning consultant working for clients from large corporations to startups and on multiple projects involving machine learning in healthcare, retail, finance, publishing, airlines, and other domains. He is an IIM Bangalore alumni of BAI and Machine Learning Engineer Nanodegree Graduate from Udacity. He is also an open source Python library volunteer and contributor for machine learning scikit-learn. For the past 6 years, he has been working as a Machine Learning Consultant for clients around the globe, by guiding and mentoring client teams stuck with machine learning problems. He also conducts leadership and motivational workshops and machine learning hands-on workshops. He is an author of nine self-published books and his new two-volume book series, The Predictive Program Manager based on Data Science and Machine Learning, is his latest work. He is currently writing books on Artificial Intelligence, Robotics, and Machine Learning. You can learn more about him on http://www.PuneetMathur.me/. xi About the Technical Reviewer Manohar Swamynathan is a data science practitioner and an avid programmer, with over 14+ years of experience in various data science-related areas including data warehousing, Business Intelligence (BI), analytical tool development, ad-hoc analysis, predictive modeling, data science product development, consulting, formulating strategy, and executing analytics program. He’s had a career covering life cycles of data across different domains such as US mortgage banking, retail/e-commerce, insurance, and industrial IoT. He has a bachelor’s degree with a specialization in physics, mathematics, and computers and a master’s degree in project management. He’s currently living in Bengaluru, the Silicon Valley of India. He has authored the book Mastering Machine Learning With Python - In Six Steps and been involved in technical review of books around Python & R. You can learn more about his various other activities on his site http://www.mswamynathan.com. xiii Acknowledgments First of all, I would like to thank my publisher Apress and its team of dedicated professionals who have made this book writing journey very painless and simple, including Acquisition Editor Celestin John, Coordinator Editor Aditee Mirashi, Development Editor Matthew Moodle, and many who have worked in the background to make this book a success. This book has been possible because of many people with whom I have been professionally connected in different ways. Many of my clients prefer to remain nameless due to non-disclosure treaties; however, I have learned the most from them. The business problems they presented to me and the solutions that worked well and did not work well in those situations is the essence of a professional career as a machine learning consultant. I also thank the contributions of hundreds of healthcare, retail, and finance professionals who interacted with me and were willing to spend time and explain their problems in the industry sectors in which they were working. Your patience, time, and effort has borne fruit in this book, and I sincerely acknowledge your contributions toward this book. The experts from healthcare, retail, and finance domains came together and agreed to give their selfless feedback in the form of Delphi Method surveys, which are part of this book. It is not possible to individually thank all of them, but I know without your contributions this book would not have been in the excellent form that it is being presented to the reader today. I wish to acknowledge my immediate family, my wife, my son, and my dog ,who gave me the emotional support that was needed to complete the book. I must tell you that I am also a BOT Father; yes, I have many bots that have helped me in the creation of this book, and they do deserve to be named as part of their contributions to this book. I have a Bot named KEYWY, which I made for the purpose of getting the choicest keywords by looking at the subject matter of the book and spidering the web to get the most relevant ones is one that has made my SEO life simple. Then there is another Bot, PLAGI, whom I created to check each paragraph of this book, and it then spidered the web to see if a duplicate content existed and warned me if there was xv Acknowledgments one. The uniqueness of this bot is that it can check programming language source codes such as Python and Java. He is a life-saver as far as plagiarism is concerned. The last Bot that needs mention is GRAMMERY. She relentlessly checked the grammar and helped me correct it as soon I was finished writing. She is the only grammar BOT that I know of and that I created that has the ability to correct not just English text but also source code text like Python and Java. xvi Introduction The idea of writing this book came up when I was planning a machine learning workshop in Bangalore in 2016. When I interacted with people, I found out that although many said they knew machine learning and had mostly learned it through self-study mode, they were not able to answer interviewers’ questions on applying machine learning to practical business problems. Even some of the experienced machine learning professionals said they had implementation experience of computer vision in a particular area like manufacturing; however, they did not have the experiential knowledge on how it can be applied in other domains. As of the writing of this book, the three most progressive and promising areas for implementation are healthcare, retail, and finance. I call them promising because there are some applications that have been built in areas like healthcare (e.g., with expert robotic processes like surgical operations); however, there are more applications that are being discovered every day. Retail affects everyday lives of everybody on this planet, as you need to shop for your personal needs. Whether you buy from a grocery store or a retail chain, online machine learning and artificial intelligence is going to change the customer experience by predicting their needs and making sure the right solutions are available at the right time. Finance is another area that holds a lot of promise and has seen less application of machine learning and artificial intelligence in comparison to the other sectors. The primary reason for that is because it is the sector with the maximum regulations and law enforcement taking place heavily here. It is also the sector which forms the backbone of the economy. Without finance, there is no other sector that can operate. Readers, be they those who are just starting off with machine learning or with experience in Python and machine learning implementation in projects other than these sectors, will definitely gain an experiential knowledge that I share with you the through the case studies presented in this book. The reader will get motivation from my famous quote on artificial intelligence and machine learning it is not the Artificial Intelligence but the Human Intelligence behind the Artificial Intelligence that is going to change the way we live our lives in the future. There are three sections in this book, and I think each of these could have been printed as separate books in themselves. The good thing that the reader will find is xvii Introduction that the structure of these three sections is identical. Each section starts off with an overview section where you will understand the current scenarios for that segment, such as healthcare, retail, or finance. Then there is the technological advancement chapter common to all the three segments, where the state of machine learning has been discussed in detail. It is also the section where I present to you the results of the Delphi Method expert survey for each of those domains. Then there is a chapter on how to implement machine learning in that particular domain. This is where you will learn how to use an industry-emulated or modeled data set and how to implement it using Python code, step-by-step. Some of this code and approach you will be able to directly apply in your project implementations. In each section, you will find two case studies taken from practical business problems, again modeled on some of the practical business problems that are commonly faced by businesses in that industry segment. Each case study is unique and has its own questions that you must carefully study and try to answer independently. I have given the solution for only one of the case studies using Python code, and I have let the second case study in each section be a discussion-only solution. The reason for doing this is because I want you to apply your own mind to solve them after looking at how I have solved the first one. Please remember each business is different, and each solution has to also be different. However, the machine learning approach does not differ much. I know for sure that many of you who read this book are highly experienced machine learning professionals in your field and that is why you are looking for expert advice on how to avoid common gotchas or pitfalls during machine learning in that domain, such as healthcare or retail or finance. Each sector has its own set of pitfalls, as the nature of the business is very different. There could be many readers who could belong to the startup eco-system and would like to get new ideas on implementation of machine learning and artificial intelligence in these areas. In each of the three sections, you will find three innovative ideas that I present to you that you could immediately take and start implementing. If you are looking for a book that gives you experiential and practical knowledge of how to use Python and solve some of the problems in the real world, then you will be highly satisfied. All the Python code and the data sets in the book are available on my website URL: http://www.PuneetMathur.me/Book009/. You will need to register there using your e-mail ID and the link to download the code, and data sets will be sent to you as part of the registration process. xviii CHAPTER 1 Overview of Machine Learning in Healthcare In late January 2018, I sat in a plush downtown hotel in Bangalore with one of the fine minds in healthcare, as well as a renowned international doctor and a Python programmer. The discussion was around how to apply machine learning in healthcare and finance. Being one of my clients, he wanted to not just get ideas but to see in practical Python code terms how data could be put to use in some of the work that was done in his large hospital chains as well as his investments in the areas of the stock market, commodities investments, etc. My discussions during the 4 days of meetings were not just intense but deep into the business domain of the healthcare industry. After having studied such similar patterns in many of my healthcare projects with my clients, in this book I present to you fine practical examples of implementation that are not just workable but also make a lot of business sense. While most of the work I do falls under non-disclosure agreements, thus not allowing me to reveal the confidential stuff, in this book you will find many examples of implementation of ideas that are from the real world. The real data has not been used. Most of the data I present in this book is from the public domain. I shall be using Python version 3.x compatible code throughout this book. Note Python version 3.x has been used throughout the book. If you have an older version of Python, the code examples may not work. You need a version of Python 3.x or later to be able to run them successfully. © Puneet Mathur 2019 P. Mathur, Machine Learning Applications Using Python, https://doi.org/10.1007/978-1-4842-3787-8_1 1 Chapter 1 Overview of Machine Learning in Healthcare Installing Python for the Exercises For running the exercises in this book, you will need Python 3.x. I recommend you use WinPython for this purpose. WinPython is a simple Python distribution, and it does not require any installation whatsoever like Anaconda. You can just copy it in a folder in Windows, change your $PYTHONPATH to the folder where you copied WinPython, and you are done. WinPython has pre-installed all the major packages that we need in this book. So you’ll save time if you use WinPython. You can download WinPython from https://winpython.github.io/ on github. Choose from 64-bit or 32-bit versions of the distribution, as per your computer requirement. As of the writing of this book, the WinPython website has a release of WinPython 3.5.4 1Qt-64bit. All the code exercises in this book work on this version of WinPython. If, however, you want to work on Windows, I would recommend you go with Anaconda for Python on Linux installers, given here: https://anaconda.org/anaconda/python. Process of Technology Adoption Before we begin to look at how machine learning is transforming healthcare, let us look at machine learning technology and the process of its adoption. This process is common to all sectors and industries. I will also explain this with examples as to how the adoption process has worked in some of the areas of healthcare, retail, and finance. As per the definition of machine learning, it is a field of computer science that gives computer systems the ability to “learn” (i.e., progressively improve performance on a specific task) with data, without being explicitly programmed . The later part of the definition, “without being explicitly programmed,” is controversial, as there are hardly any computers that do not require programming to learn. But what this could mean for applying machine learning in business is the use of supervised and unsupervised machine learning techniques. Supervised learning techniques are the ones where the computer needs references of past data and explicit categorization and explanation of patterns, trends, and facts from it. However, for unsupervised learning this is not a requirement; we let the computer learn on its own to find the patterns, trends, and facts. This is also known as auto-discovery or auto-data-mining. So when we use unsupervised learning, you can say that the computer program is not being explicitly programmed to learn. It is learning on its own by discovering the facts, patterns, and trends. But we do program it by selecting the algorithms it will use to discover them. It does not select the algorithms by itself. To give you an example of how 2 Chapter 1 Overview of Machine Learning in Healthcare this happens, let us say we want to develop a machine learning algorithm for developing and finding out if hospital consumer data has any given patterns for predicting whether a particular outpatient would be admitted to the hospital or not. Simply put, are there any hidden patterns in the data to find out the profile of a patient? This can be done in two ways: the first uses a human machine learning engineer who can start to look at the hospital outpatient and in-patient data sets and then see if there are any patterns; the second uses unsupervised learning and lets the computer select clustering algorithms to find out if there are any clusters in both the outpatient and in-patient data sets. We will look at this example with code and how to implement this in Chapter 3. Now we look at Figure 1-1 Machine learning technology adoption process below. Figure 1-1. Machine learning technology adoption process Now let us look at how this machine learning technology adoption takes place in the industry. I am outlining here the process that is common to all sectors, regardless of their technological complexity. In the technology adoption diagram in Figure 1-1, you will find four phases of the technology adoption that takes place in any sector. The first phase is quick applications. This phase is marked with certain characteristics. This is the stage where the business tries to apply machine learning technology on the low-hanging fruits. As an example, a company may want to automate its social media analysis or sentiment analysis. It would also look to automate some of the less-than-1-minute tasks performed by its employees. This task would be low on technological complexity. It would also like its employees to list the repetitive tasks and to do things like root cause analysis for any failures or issues in the business systems. The focus here would be hindsight. This means that the business is trying to focus on such issues or problems and trying to address those that have caused failures in the past. As an early adopter of the technology, the business is still trying to understand how machine learning is going to help them advance their applications for business growth. 3 Chapter 1 Overview of Machine Learning in Healthcare The next stage is that of early applications of machine learning, where the business will try to create learning operations. This means that they are trying to look at the past data and find out what can be learned from it. The business is also trying to address the low-efficiency test so it may carry out an efficiency audit in its operations to help find out identify those areas where it can learn and be more efficient in its business operations. In early applications of machine learning, the business could also think of reducing the cost of its existing operations. And in this it could also carry out cost audit for its various business operations carried out by its employees. It could, as an early adopter, target those operations that are high cost and high growth in nature. It is also to diagnose clearly the business, which would look at the business problems and the reasons for the issues it is facing and focus on how to avoid them in the future. The business would also look at building problem detection systems, such as building a credit card fraud detection system. In this case, as well as in the earlier applications, the business is trying to focus and gain hindsight. Now I move to the third phase of technology adoption, where there are assisted applications of machine learning. Here there is application of low-level intelligence to assist the experts in highly skilled tasks. The focus of automation here is to augment the human capability for business growth and advancement. The effort here is to predict the business requirements from data and to make use of such predictions for enhancing the business. Here the focus of the business is to gain an insight and not to just automate its operations but also to gain from the hidden patterns, facts, or trends that may have been lying hidden in its data. In this stage, the organization is discovering about its customers, its employees, and also its operations and, as a result, trying to understand the things that have been troubling it in the form of business issues or problems. This is actually where the business organization will start to look to apply machine learning-supervised techniques with the unsupervised techniques. Now we move on to the fourth and the last phase of technology adoption, which is independent applications of operations using machine learning. This is a stage where the automation of a company has reached its fullest capability. Most of its operations are robotic in nature. This is also the stage where there is an expert human replacement happening. In this stage, there is also foresight and prescription on a future course of action for a particular business strategy or vision or mission. As I said before, this is the stage where the human specialist is being looked at being replaced or assisted at a high level. So here the machine learning is being used at a level where the learning by the machine is at its fullest extent. The machine is capable of learning from the huge 4 Chapter 1 Overview of Machine Learning in Healthcare data generation happening inside the business operations. It has also developed skills for finding out hidden patterns, facts, and trends to prescribe to its business leaders the future course correction or actions that need to take place in order for the business to grow. This machine learning capability can also be used for averting any kind of debacle, such as financial crisis or scams that may happen in the future or may be reflected in the current data of the business organization. In this stage, the business is using foresight, and it is this foresight that actually gives its operations the course correction capability. This is the maximum extent that a business operation can use machine learning to gain advantage in the market against its competitors. I am not suggesting that the entire company operations be run in an auto-mode. That is not what this phase represents. This state is that of an organization that has intelligent automation in place. By intelligent automation, I mean that the key business functions, such as finance marketing purchase, are sufficiently automated to provide foresight about the business operations. The company also has the ability to gather data from its business environment and to avoid any tragic incidents that may occur not due to the company’s fault but due to the nature of the business environment, such as recession, market crashes, etc. I now present in tabular format the characteristic feature of each phase so that you gain a clear understanding of the entire process. Table 1-1. Phases of Technological Adoption and Advancement Phase Characteristics Focus Analytics used Quick applications 1) Low technological complexity 2) Replacement of repetitive and mundane tasks 3) Solutions for common issues and problems ProblemDescriptive solving of analytics day-to-day issues faced in its operations Level of prediction Technological complexity Hindsight Low (continued) 5 Chapter 1 Overview of Machine Learning in Healthcare Table 1-1. (continued) Phase Characteristics Focus Analytics used Level of prediction Technological complexity Early applications 1) Improve efficiency and productivity 2) Reduce cost of operations 3) Diagnosing business problems and issues faced in the past 4) Building problem detection systems Learning from the problems faced in its operations Diagnostic analytics Hindsight Medium Assisted applications 1) Assist highly skilled Automation professionals in business operations 2) Augment human specialist capabilities 3) Predictions of business requirements Predictive analytics Insight Medium to Complex Independent operations 1) Robotic operations Cognition 2) Robots gain specialist capabilities after learning from the human specialists. 3) Prediction on future events and capability to course correct in advance 4) Cognitive capability building Prescriptive Foresight analytics 6 Highly complex Chapter 1 Overview of Machine Learning in Healthcare From Table 1-1 we can clearly see what I have described in Figure 1-1 and also understand a few more aspects of the process of technology adoption. I will also explain this table in detail by taking examples in healthcare where some organizations have used these capabilities. I have added the aspect of analytics in Table 1-1, which I have not discussed in this book so far, so let’s look at what these forms of analytics are and how they can be used by healthcare organizations. I have explained these analytics types in my book The Predictive Program Manager Volume 1 (Chapter 2, page 17) and I take the definitions of analytics from there . Descriptive Analytics: This field of analytics is invoked to know about the answers to questions for projects that have already happened, such as “What is the status of X Project?” Diagnostic Analytics: This field of analytics is used to know the root cause of a phenomenon, such as a project’s success or failure. Why did the X Project fail? What are the positive lessons we can learn from this project’s success? All such questions can be answered using diagnostic analytics. Predictive Analytics: This type of analytics is used for determining the outcome of an event in the future, such as project success or failure, project budget overrun, or a schedule slippage for an ongoing project. Prescriptive Analytics: In this field of analytics the maximum value of analytics is achieved as it builds upon the prediction made based on predictive analytics, and it prescribes actions that should be taken for the future. I have used descriptive analytics for a client in the US for detecting whether a healthcare institution was using racial discrimination practices in its operations. I was given data on the patient records and their backgrounds. Patient data was given with their racial orientation, such as Asian, Native American, etc., along with data on admissions to the ICU, operations, and admissions to hospital wards and private rooms. I had to analyze and give conclusive evidence using statistical techniques as to whether there was any racial bias. By using descriptive analytics and looking at the patient records, I was able to say with confidence that there was not much evidence of such acts in the data. My findings were later used for evidence in legal proceedings as well. So I had to be careful to analyze data from all angles to confirm that there was no such pattern present in the data set. Diagnostic analytics is used in the life of every healthcare professional. The industry is very diagnostic-driven, as it tries to diagnose the disease based on symptoms. So building systems that diagnose issues and problems is not very difficult. Genomics is a field where much diagnostic research is taking place at IBM Watson project for 7 Chapter 1 Overview of Machine Learning in Healthcare Genomics is at the forefront in such research . IBM Watson is an analytics engine built by IBM for use in machine learning and artificial intelligence. The machine learning engine IBM Watson is helping find solutions for individual treatment of cancer patients using its huge data sets comprised of medical literature, clinical study results, pharmacopeia, etc., to find cures for cancer patients. This is public research available to oncologists worldwide and is helping unearth possible new cures for various forms of cancer . Predictive analytics is the next level of implementation of machine learning in the healthcare industry. In such an implementation, for example, the focus would be on predicting the likely group of people who could develop cancer. A system so developed would be able to predict accurately the age and type of people who are likely to develop a particular type of cancer. It would have the ability to create a profile of cancer patients, and as such a person comes in contact with this type of analytical system, it would throw up an alarm on the likely case of developing cancer. Prescriptive analytics is being used by an IBM Watson for Genomics project, where it not just diagnoses the disease but also gives a prediction and then a likely prescription for the type of cancer by looking at clinical drug trials and their results. Although this system is undergoing rigorous testing, it will yield significant results when it is able to increase its predictive and prescriptive accuracy. How Machine Learning Is Transforming Healthcare Let us now look at some of the ways that machine learning is transforming the healthcare segment of business. The healthcare industry is one of the most labor-intensive industries around the world. It requires the presence of humans to take care of people at various stages of their illnesses. I was at the AI Conclave held by Amazon in 2017 in Bangalore and was amazed to see how an acute problem of staff scarcity, which has been plaguing the healthcare industry in the United Kingdom, has been aptly solved by creating artificial tabletop bots that would take care of elderly patients needs (1). The artificial tabletop bots remind elderly patients to take their pills, track their prescriptions, and track and suggest wakeup routines. At the heart of Echo Alexa (as it is known) is the machine learning developed by the Amazon team using its cloud infrastructure Amazon Web Services (AWS). At the heart of Alexa is the Python machine learning code that helps it to perform tasks and learn from them through a feedback mechanism. The wonderful part of this service is that Echo Alexa is available to a common Python developer to use and develop their own programs and products based on Amazon’s infrastructure. 8 Chapter 1 Overview of Machine Learning in Healthcare In another DataHack Summit in 2017, I had an opportunity to see the demo of IBM Watson for healthcare services. Developers built their own applications on top of this base analytics engine. IBM has proven to use its analytics engine in applications such as testing genetic results, drug discovery, oncology, and care management, to name just a few. One more area where not just IBM but other analytics engines are making headway is in diagnosing disease using imaging. In healthcare imaging, such as X-ray images or CAT scan images, all have traditionally been interpreted by humans. However, there are some reasons why we need machines to do this work more efficiently: –– High volume of imaging data with increased patients. –– Stress on doctors due to high volumes makes them more error-prone. Machines can handle large sets of imaging data with a lower error rate. –– Inability of healthcare professionals to link and see the big picture from imaging data. Machines can help them by assessing large numbers of image datasets and determine whether there are any patterns or any connections among groups of patients or groups of localities, for example. –– Replace doctors or specialist at times of their absence. This is a key operation that a machine can do—when a specialist is not available, it can replace the human specialist and provide diagnosis in even critical cases. In my opinion this function of a machine will be used more and more, and the day is not far when the entire area of image diagnosis will be done by machines with no human intervention. –– Drug discovery is a very key area for the healthcare industry. Research in the pharmaceutical companies for diseases like cancer or HIV is continuously happening. Machine learning is helping speed up drug discovery by analyzing medicinal data and providing prediction models on drug reactions even before they are injected into subjects in a controlled environment. This saves both time and money, as the simulation of drug reactions gives an estimate on likely cure patterns and reactions to the drug. –– Patient Research in difficult fields like Cancer, etc. There is a lot of data available in this field for both patient and clinical trials of medicines. Clinical trials are time-consuming and require collection of subject data on reactions in the body. This is either collected invasively, such as via a blood test, or non-invasively, such as through urine tests or putting probes on certain body parts of the subject. 9 Chapter 1 Overview of Machine Learning in Healthcare One of the common fears that I hear with healthcare professionals is their fear that AI will replace them. The machines may make their jobs redundant. That fear is valid and is not without evidence. This report comes from The Sun  China where a Robot named “Xiao Yi” has passed China’s National Medical Licensing Examination successfully and has achieved all the skills to practice medicine. Some people say this is a scary trend. Some say it is a clear sign that robots are going to rule the humans. However, I say this is just the tip of the iceberg. The following are some of the trends we are likely to see in the healthcare world as far as machines are concerned: –– Robots replace workers in low-paying jobs first, where humans do not want to do the mundane work, such as the case of Amazon’s Echo Alexa replacing elderly healthcare due to staff shortage. –– Robots become assistants to senior specialists, like neurosurgeons, and learn the skills for diagnosis and surgery. –– Robots will replace senior specialists in diagnosis, as it requires more analysis and processing. Humans can't process large information and spot patterns in big data sets. This is where robots will score significantly higher in accuracy of diagnosis than a human specialist. –– Surgery will be done by humans with assistance from robots. This has already been demonstrated by the University of Oxford Surgeons . So in my view, it is possible as more and more robots are built to do precision operations on humans and are successful, they will work jointly with human specialists to carry out complex, precision-based surgeries. This trend will start to emerge in the next 2 to 3 years. They may be termed as Auto-doctors and Guided-doctors. Auto-doctors would use unsupervised learning techniques to treat a patient for new discovery diseases. Guided-doctors would use supervised learning techniques. They would work for known diseases on known lines of treatments. We will be looking at an in-depth example of a Python program for supervised learning in Chapter 3, “How to Implement Machine Learning in Healthcare.” 10 Chapter 1 Overview of Machine Learning in Healthcare E nd Notes  IBM Watson Genomics, https://www.mskcc.org/ibm-watsonand-quest-diagnostics-launch-genomic-sequencing-serviceusing-data-msk  https://www.portsmouth.co.uk/news/hampshire-councilto-use-amazon-echo-technology-for-elderly-social-carepatients-1-8122146  For the First Time, a Robot Passed a Medical Licensing Exam, Dom Galeon, https://futurism.com/first-time-robot-passedmedical-licensing-exam/  Robot Passed a Medical Licensing Exam: https://www.thesun. co.uk/tech/4943624/robot-doctor-medical-exam-chinabeijing/  Definition of Machine Learning, https://en.wikipedia.org/ wiki/Machine_learning  Page 17, Chapter 2, The Predictive Program Manager Volume 1, Puneet Mathur  World first for robot eye operation, http://www.ox.ac.uk/ news/2016-09-12-world-first-robot-eye-operation# 11 CHAPTER 2 Key Technological advancements in Healthcare S cenario 2025 In the not so distant future in the year 2025, one fine morning an old lady receives an alert on her personal home management device that she is going to develop cancer in the near future. This report has been sent by her robot doctor, after her visit last week for a checkup. She is mildly shocked to hear such news. She then decides to get a second opinion from a human doctor. The human doctors are very few in numbers now in her city and are more expensive than the robot doctors. So she decides to visit the human doctor nearest to her home. She visits the doctor and shows him her report, which was sent to her by the robot doctor this morning. The human doctor carefully looks at the report and finds that the robot had mentioned a clinical study that was done in the year 2019 where it was proven that people with a sleeping disorder lasting more than 3 weeks in a row had a 90 percent chance of getting a certain type of cancer. Using its probe sensors installed in the patient’s house, the robot doctor had detected that she had experienced a disturbed sleeping pattern for more than 6 weeks in continuation. Based on this fact, the robot doctor had looked at her vital statistics data, such as her heart rate, blood pressure, breathing patterns, etc., and had reached the conclusion that she was on the path to get cancer. The human doctor, on the other hand, checks her vital statistics again and asks her to conduct some blood tests and other required tests for determining her current medical condition. After a few days, when her medical reports arrive, the human doctor declares that she does not have any signs of cancer. Does this sound far-fetched and something too distant? © Puneet Mathur 2019 P. Mathur, Machine Learning Applications Using Python, https://doi.org/10.1007/978-1-4842-3787-8_2 13 Chapter 2 Key Technological advancements in Healthcare This is not an unlikely scenario but something that we may witness once the robot doctors become a reality. We have seen in Chapter 1 that there is a robot in China that has already successfully passed the medical examination and has attained the medical degree of a doctor. What questions arise in your mind once you read the situation? What would you do if something like this happened to you? Would you trust the robot doctor? Would you trust the human doctor more? Would you dismiss the report by the robot doctor as false and ignore it after the human doctor gave you a clean chit on your current medical condition? These are some of the questions that the future society is going to have to deal with once we accept robots as specialists in the healthcare industry. If you noticed, this is a scenario where the human expert does not have the ability to prescribe any medicine based on the patterns that it is observing in a human being. In this case, the robot doctor is better prepared to predict and prescribe course-corrective medication to a human being based on the data that it gets from its connected probes or sensors. The healthcare industry in particular deals with human beings and their lives. This is one of those industries where a simple judgmental error could cause death to a patient. However, when we talk about building prediction models based on machine learning (ML), which is the brain behind any robot, we know that no matter what algorithm is selected for predicting the outcome from any data set, there is going to be a percentage of errors in the final prediction by the model. In the case of human beings, a human being or a human doctor or a healthcare professional is also prone to errors. This is something that we know as human error. A recent research by Hopkins Medical Organization or the Johns Hopkins Medical Organization shows that 10 percent of all the U.S. states happened due to medical errors by the doctor and it is the third highest cause of death in the US . So if we were to build and create a replacement or a competitor for a human doctor, we know that it would have to do better than this error rate. It can only survive if it gives predictive diagnosis at a lower error rate than that of the human doctor. Since we are dealing with human life in the healthcare industry, we need a gradual and careful way of adopting technology, as a lot is at stake. The requirement is to build robust algorithms with prediction models with higher accuracy levels. Narrow vs. Broad Machine Learning Now let us understand the brain behind robotics, which is ML. There are two types of ML applications: one is narrow in nature, and the second is broad in nature. Narrow ML deals with creating programs algorithms and robotics software that caters to a narrow 14 Chapter 2 Key Technological advancements in Healthcare focused set of activities or skill set. Here, the narrow means that the area of application is a specialized skill. It relates to an expert and its purpose is to emulate and exceed the human expert in their specialized skill. Narrow ML works best when it has an expert to learn from and to copy. An example of narrow ML robots would be the robotic arms belt for doing heart operations, such as removing blood clots from arteries. This example is of a robot that requires assistance from a human being in order to carry out its operation. We can see this in Figure 2-1. Figure 2-1. Narrow versus broad ML 15 Chapter 2 Key Technological advancements in Healthcare In Figure 2-1, we can clearly see that narrow ML concentrates on things like healthcare, finance, and retail. In comparison, broad ML is about building a humanoid, giving it cognitive capabilities in artificial intelligence (AI) and the ability to emulate physical characteristics of human being. Now let us look at the broad ML application. Here, we are talking about creating programs algorithms and Robotics software that caters to generalized skill as opposed to specialized skill. It emulates general human behavior, and the purpose is to prove robotic capability equal to that of a human being. A recent example of such broad application of ML is the robot named Sophia that has gained citizenship in the Kingdom of Saudi Arabia due to its proven ability to emulate human conversation. As the technology advances we will see more robots being developed on broad ML applications. However, the current trend in the healthcare industry is to adopt robotics and its applications in a narrow way and to help emulate or replace experts in diagnosis of disease research of new drugs and other such areas. We can look at the difference in Table 2-1. Table 2-1. Narrow vs. Broad Machine Learning Application Applied Machine Learning Area of Application Focus Purpose Narrow Specialized skill Expert capability Emulate & exceed expert performance Broad Generalized skill General human behavioral capability Prove human-like capability urrent State of Healthcare Institutions C Around the World Now I would like to look at the big picture of the current state of the healthcare industry around the world. The turmoil that is going on in the healthcare world is depicted Figure 2-2. Note the two opposing forces: one that is the traditional healthcare institution that is generally comprised of wellness clinics, doctor clinics, and hospitals. A another new set of institutions that are coming up are based on robotics ML AI. In the international 16 Chapter 2 Key Technological advancements in Healthcare conference on best practices in healthcare management held in Bangalore in March 2018 at XIME, where I participated, this trend was clearly brought out. You can read more about the conference in the following url: http://xime.org/Health%20care%20 conference%20report. The traditional healthcare system derives its values from empathy, human touch, and healing through the doctor. As opposed to this, there is another set of institutions that are coming up rapidly. The values that these institutions bring forward are those of efficiency and accuracy of healthcare operations, better management of resources, and minimal human touch to avoid spread of communicable diseases. Both the systems target giving better care to the patient. In the traditional view the doctor is irreplacable and is the center of the healthcare institution. However, the new and modern view is that the doctor has a limited capacity of analysis and cannot analyze the big picture—hence, such machine algorithms and robots, which can do a better job. I have already discussed the narrow versus broad ML applications in this chapter. The reader should take note that institutions based on robotic ML and AI are trying to make headway into replacing the traditional healthcare system by targeting narrow ML applications first. Here the attempt is not to replace the doctor as a whole but to replace or emulate and then replace certain specialized functions of a doctor or healthcare professional. Figure 2-2. Opposing forces in the global healthcare industry 17 Chapter 2 Key Technological advancements in Healthcare One example of ML being used for narrow healthcare tasks comes from Siemens Company from the division healthineers. They have computer vision imaging on computer tomography and look at what the brain wiring looks like through an MRI scan. They have brain anatomy machines known as Tesla machines, which I used to do this task. The other application of ML by the same company is the CT scanner, which is used for parametric imaging or molecular imaging, and healthcare workers have applied it to show whether a tumor is benign or malignant. This research has been done based on applying AI to 250 curated images for the path lab machine. They have developed precise algorithms for positioning the patient using 3D cameras inside the Tesla machine, as this used to be a human-aided task, and every human had their different way of positioning the patient inside the machine, sometimes leading to poor quality of images. The ML algorithm has now enabled positioning of patients as quickly as possible to get better images. They have also developed deep learning algorithms for reading chest X-rays and to detect abnormality in the X-ray machine. This is an attempt to replace the specialized role of radiologist with numerous hours of expertise with all X-rays that are thrown before them, including an MRI and CT scan. On the same line, Siemens has developed an MRI image fingerprinting technique using deep learning to emulate what a radiologist does. It is also a pioneer in the field of lab robotic automation, using an electromagnetic levitation technique, which is used in speed trains around the world . I now bring to the reader another example of an organization using ML applications to develop a solution for overcoming a social barrier in an innovative way. This company is a startup in India known as Niramai ; it was conceptualized by two women, Geetha Manjunath and Nidhi Mathur, who founded this startup. Nidhi presented in the XIME Healthcare conference all the solutions developed by her company for identification of breast cancer in women. In a country like India, where traditional beliefs are prevalent in the rural regions, the major hindrance to detecting breast cancer is that the traditional system requires a doctor to touch the patient’s breast to detect a lump that may become cancerous. The major method used even today is for the doctor to feel and use his/ her hands to see if there is a presence of a lump in the region of the body. To overcome this drawback, Manjunath and Nidhi looked at how technology could be used to help diagnose breast cancer without using touch or invasive procedures or applying pressure through mammography, which is painful. So they looked at a solution by using high- resolution, full-sensing thermal image with ML and use images to detect prevalence of cancer . By using a high-resolution thermal sensing device and artificial intelligence 18 Chapter 2 Key Technological advancements in Healthcare and ML, they are able to develop API, which is non-invasive, does not require any test, and does not cause any pain to the patient. They require permission to take off the patient’s clothes while the machine detects for the prevalence of cancer and whether it is malignant or benign, which matches any mammography test done manually. Over time I am sure that the algorithm will learn and improve by itself. Such innovative use of technology that focuses on overcoming social issues in healthcare are going to be adopted faster in countries where the population is high and there are social stigmas against medical help that are preventing it from spreading as a method of cure with the common population. Importance of Machine Learning in Healthcare The fact that sets healthcare apart from other fields like finance and retail is that healthcare deals with human life, and when we apply ML we need to have a gradual and careful way of adopting technology, as a lot is at stake here. Robust algorithms with prediction models with higher accuracy levels are required. This can be changed from a very simple example where we build a prediction model that predicts a particular type of cancer with an accuracy of 95 percent. In this case the prediction model will predict accurately for 25 patients and predict incorrectly for the other 5 patients. So the incorrectly predicted patients will still think they do not have cancer. This is the reason why application of ML in healthcare requires more testing before a model is deployed in production. Some of the key areas where healthcare has machine learning applications are: 1. Disease identification 2. Personalized medicine 3. Drug discovery 4. Drug manufacturing 5. Clinical trial research 6. Radiology 7. Digital health records 8. Epidemic outbreak prediction 9. Surgical robotics 19 Chapter 2 Key Technological advancements in Healthcare All of these areas in healthcare are core to the healthcare industry. Now we are going to look at the aforementioned areas of the healthcare industry and do the ML technology adoption process that I discussed in Figure 1-1 of Chapter 1. This mapping is going to help us in understanding where these areas stand with regard to the technology adoption process in the current scenario. By doing this we further look at what kind of advancement can happen in each of these particular areas. For an example on how to use this mapping information, let’s say that your hospital has implemented surgical robotics in the field of heart surgery. By knowing from this chart how advanced the robotic surgeries are with respect to the technology adoption process, we can look at what kind of technological advancement could come in the future for this surgical application. In order to have a current view of the global healthcare industry in the year 2018, I carried out a research study using the Delphi Method with 18 healthcare professionals. This is an independent study done by me and is not sponsored by any institution. I am also not directly connected with any healthcare institution on a regular basis, given to study a more independent perspective. The purpose of the study was to take expert opinion and to find out the current state of artificial intelligence and ML in the healthcare industry. I used the Delphi Method in research. We need to understand what the Delphi Method is and how it has helped us in this study. Let’s first look at the research methodology used in this study. Research Objective: The primary objective of this research is to use expert opinion in finding out and mapping two parameters of AI and ML: (1) the current technology maturity level of AI and ML in the key areas of the healthcare industry, and (2) the parameter of the technology adoption process inside the healthcare industry. There were initially 12 key areas identified by the expert groups in the first iteration. These areas were then reiterated with the expert group to find out the most important areas that would evolve in the future. The expert group was able to identify nine areas in healthcare that would be important for the healthcare industry to advance further. The research study does not provide the results of this iterative selection of the key areas, but it starts from the point where the experts have selected these nine key areas. I have already discussed in this chapter those nine areas, starting from disease identification to surgical robotics. Research sample: The group of experts that was selected was from a total population of 232 experts. The expert group was comprised of healthcare professionals who had worked in the industry for more than 20 years at positions including 20 Chapter 2 Key Technological advancements in Healthcare patient care, a management expert in a healthcare institution as a director, chief executive officer of a major healthcare facility, and academic professors who had worked on research in the healthcare industry with accepted and published papers. I have covered all the experts from each of the areas in healthcare, such as patient management, drug research, surgeons, CEOs, and AI experts—to name just a few. A total of 18 such professionals were shortlisted for this study. There were no absentees or attrition in this study. Information needed: In order to make decisions and to support them, various secondary data like published papers on the state of ML and AI in healthcare were provided. An example is that of Siemens healthineer Emma Watson’s research in genome study and cancer detection. The required information in order to create a map between the two parameters mentioned earlier was based on the experts’ understanding of the current state of technology implementation in the nine areas, starting from disease diagnosis to clinical trial research. The decision making of the expert explanations on the levels of technological maturity and the phase-wise identification of technology was provided to them. Beyond that there was no other information provided, so care was taken not to create a bias in the minds of the experts. The information needed for this study included contextual, theoretical, and expert knowledge. The research also required for the experts to use their tacit or inherent knowledge, which they possess from being associated with the healthcare industry for so long. Research Design overview: The primary steps involved in this research are the following: 1. Define the objectives of the research. 2. Find experts who are willing to help in this research study. 3. Design questionnaires that gather information and involve less writing effort by the experts. 4. Administer the questionnaires to the experts. 5. Gather responses to the questionnaires and analyze them to see if consensus was achieved. 6. Iterate and administer more questionnaires until the experts reach a consensus on a particular key area. 21 Chapter 2 Key Technological advancements in Healthcare 7. Once a consensus is reached, move on to the next key area and iterate the questionnaire until consensus is reached. Until the time consensus is reached, provide more information based on the previous responses provided by the experts. 8. Analyze and create a map of the technical maturity levels and phases of adoption of AI and ML. Data Collection methods: Literature regarding healthcare was not data to be collected for this study. The test study that was conducted, which I mentioned earlier, was that of taking expert help in narrowing down from 12 to 8 key areas that are going to be important for the future of healthcare industry. This is important because in our study we are using expert judgment on what is going to be the focus of the healthcare industry based on their past experience. We have used the Delphi Method of study from a paper by Chittu Okoli and Suzanne De Poweski named “The Delphi Method” as a research tool and example of design considerations and applications . The questionnaire method was used for data collection from the experts through e-mail online administration of surveys and personally giving the questionnaire in the paper mode. Data analysis: During a particular iteration, when the data was collected, Microsoft Excel was used to record the experts’ responses in a tabular format. For any given key area a graph was made to check whether there was a consensus reached and if the graph sufficiently showed The Expert’s consensus. Then the iteration was stopped. So the data analysis was done manually with the help of computer software. The mapping of technology maturity and phases of technology adoption waere undertaken using Excel software to create a technology map graph. Ethical considerations: It is possible that bias could have slipped into the study had we not made sure that the results were the responses of the experts and were kept anonymous, not affecting the outcome of this study. So due care was taken in order to ensure that the experts were not known among each other. As I have already mentioned, there is in the healthcare industry two groups of people: one group whose members like technology and the other group whose members do not like technology. We did not do an expert selection based on these specific criteria so this study could very well be biased on such grounds, and we have not tested for this. 22 Chapter 2 Key Technological advancements in Healthcare Limitations of the study: Qualitative research has as its biggest limitation that of not being able to exactly quantify the outcome of the future, and this is very much applicable to our study as well. However, by using categorical variables in our questionnaires we have tried to take the quantitative analysis of our outcome as well. Mapping of the technological adoption and understanding of the technological maturity is not something that a normal human being can do unless they have been associated with the industry, and that is why we chose experts to carry out the study. However, it is possible that some of the experts may not have had sufficient knowledge or exposure to the advances in AI and ML. We acknowledge that this could be a limitation to the study. We already know from Figure 1-1 in Chapter 1 that there are four phases of technology adoption. In Figure 2-3 we look at this mapping. Figure 2-3. Healthcare industry technology adoption phases In Figure 2-3 there are two axes. The x-axis represents the technology adoption phase as outlined in Figure 1-1, and the y-axis shows the technology maturity level. The technology maturity application level. The maturity application level is divided into Low, 23 Chapter 2 Key Technological advancements in Healthcare Medium, and High. Low means the technology is at a research stage and is not in production yet. Medium means the technology has been implemented in production with some hits and misses and needs more research to move mainstream production. High means the technology is well-researched and is ready to move into production or is being used in the production environment, such as hospitals, etc. Table 2-2 and Figure 2-4 present data with its analysis from the Delphi Method of research. Table 2-2. Data on the Delphi Method of Research Used in the Study Topic No of Healthcare Experts Delphi Method Invited Shortlisted Current Application of AI & ML in Healthcare 232 18 No of Iterations 4 We have already discussed this data in the methodology section of this chapter. Now we look at the data and its graphical representation regarding first parameter technology maturity level of AI and ML in healthcare. Figure 2-4. State of AI and ML in disease diagnosis In the area of disease diagnosis with regards to the first parameter of technology maturity levels of AI and ML in the healthcare industry, 56 percent of the experts felt that disease diagnosis had a medium level of maturity. The identification of disease diagnosis 24 Chapter 2 Key Technological advancements in Healthcare as a medium level of maturity means that the technology has been implemented in this area of disease diagnosis in production, but there are hits and misses and it needs more research to move in to mainstream production. A good example of this area would be Google’s deep learning for detection of diabetic eye disease . In this use of AI for disease detection in the traditional way, ophthalmologists use pictures of the back of the eye and also the computer vision retinopathy (CVR) to detect if there is a hint of a diabetic eye disease. CVR determines the types of lesions that are present in the image that show if there is any fluid leakage or if there is bleeding in the eye. In such a complicated analysis that is done by a retinopathy by an ophthalmologist, Google’s diabetic eye detector is able to create an AI-based system by the use of development data sets of 100 and 28,000 images given by 327 ophthalmologist . The deep neural network trained on these images of diabetic retinopathy, and when it was applied on more than 12,000 images it was able to match the majority decision of the panel of 728 US-board certified ophthalmologists. The algorithm’s AP scores compared to those scores done for disease detection manually by the ophthalmologist were identical, at a score of 9.5. Figure 2-5. State of AI and ML in digital health records Now let us look at another area: digital health records. Our experts conclude that this is at a medium state of technological maturity, with 61 percent of our experts in agreement of this opinion. We can see in Figure 2-5 that some of them (about 28 percent) 25 Chapter 2 Key Technological advancements in Healthcare also feel that the level of maturity is at a low level. Medium means that the technology is not being moved into mainstream production and has a few hits and misses here and there. However, a low state means that the research has not yet moved into production. To give you an application of use of AI in electronic health records, there is a company known as Savana (Madrid, Spain) that has successfully developed an application to re-use electronic health records with AI . The most notable feature of this system is that it uses the natural language in the form of free text written by medical practitioners in the electronic health records or in the digital health records to analyze real-time information that is generated by a doctor. The Savannah system performs immediate statistical analysis of all patients seen in the records of its software and offers results relevant to the input variable provided by the user. It uses natural language processing to accomplish this goal in the background. It uses supervised ML in classifying a written text by a doctor into background information or diagnosis information. The unsupervised ML techniques use cases for determining the semantic content of words as the algorithm learns autonomously without any predefined semantic or grammar relations. For example, engineers and Parkinson’s have similar meaning or different meaning of an example naproxen oh and Ibuprofen or asymmetrical is similar to just give you an idea of how it is done practically. Now let us look at the Figure 2-6 the state of AI & ML in Digital Personalized Medicine. Figure 2-6. State of AI and ML in personalized medicine 26 Chapter 2 Key Technological advancements in Healthcare Now we look at the state of AI and ML in the area of personalized medicine. After the four iterations our experts have told us that the technology adoption maturity is at a very low level. Eighty-three percent of our experts tell us with certainty that this is the case. The area of personalized medicine is also known as precision medicine. Here the treatment of disease, and its prevention for a patient is based on their personal variability in genes, the environment in which they live, and the lifestyle that each person follows. It is like building your own custom treatment. It is clear that this profession is not possible without the use of AI, which runs on super computers in order to learn using deep learning and to develop cognitively. It is similar to that of physicians—the computers need high processing power. They use deep learning algorithms and they need the specialized diagnosis knowledge in the area that they wish to do diagnosis, such as physicians in cardiology, dermatology, and oncology—to name just a few. Now we look at another key area that of epidemic outbreak prediction in Figure 2-7. Figure 2-7. State of AI and ML in epidemic outbreak prediction In Figure 2-7 we can clearly see that our experts tell us that the technology maturity is at a low level. Sixty-seven percent of our experts feel this after three rounds of iteration, when a consensus was reached. One good example of this is the Google flu trends , where Google was able to predict the spread of flu across many countries; however, it is no longer publishing the results. The way it used to work is that Google 27 Chapter 2 Key Technological advancements in Healthcare would analyze search queries to identify the world regions from which such large numbers of queries were coming, and it would automatically predict that those regions were going to be affected by flu. This project was started in 2008, and it was shut down over concerns of privacy issues raised by various agencies. However, in the background, Google is going to provide such data to different public health institutions to help them see and analyze the trends. This technology exists, but it needs to take care of the privacy issues before it can become mainstream. We now look at an interesting application of AI and ML in the area of radiology in Figure 2-8. Figure 2-8. State of AI and ML in radiology Our experts tell us that after four iterations, the technological maturity in this area is high, and 72 percent of our experts concluded this. A notable accomplishment by the Siemens healthineers involved application of neural networks on imaging like X-rays and converting such images into data and then analyzing them like a radiologist does. This uses deep learning and, more specifically, artificial neural networks (ANNS), which replicate the human neurons. They are reported to have detected through chest radiographs with 97 percent sensitivity and the results of cases of lung tuberculosis with 100 percent specificity  (https://www.healthcare.siemens.com/magazine/ mso-artificial-intelligence-in-radiology.html). 28 Chapter 2 Key Technological advancements in Healthcare More and more such applications are going to arise that may be integrated into the medical devices, making them automated independent robotic functions. This is the future of radiology as it stands today. We now look at Figure 2-9 for state of AI & ML in surgery below. Figure 2-9. State of AI and ML in surgery As we can see in Figure 2-9, in the field of surgery our experts see that use of AI and ML is at a very high maturity level, and 56 percent of the experts shared this opinion after our iterations. In the task of flesh cutting, which is a basic skill for a surgeon, a robot is able to make precise cuts with much less tissue damage as compared to human surgeons [8; https://spectrum.ieee.org/the-human-os/biomedical/devices/in- fleshcutting-task-autonomous-robot-surgeon-beats-human-surgeons]. There is a robot named STAR (Smart Tissue Autonomous Robot) that hovers over a patient and then, based on the algorithm, makes precise cuts that expert surgeons make, but STAR makes less damage by surrounding the flesh. This STAR system is able to sew back flesh that has been cut in surgery, and such stitches have been found to be more regular and leak-resistant than those of experienced surgeons. So this clearly shows that the use of robots in the field of surgery is indeed at an advanced stage. In Figure 2-10 we look at the State of AI & ML in Drug Discovery. 29 Chapter 2 Key Technological advancements in Healthcare Figure 2-10. State of AI and ML in drug discovery Our experts tell us that the application of AI and ML in the drug discovery area is at a medium technological maturity level, which means that although there is technology there are a lot of hits and misses, due to which this technology has not moved into mainstream production. The entire drug discovery area is closely linked to clinical trials, but in practice the drug discovery happens long before a clinical trial happens for any drugs. The drug discovery process requires pharmaceutical many tests to check and it is carried on many different drug compounds, which could help in eliminating or limiting the spread of a particular disease. So this discovery shows a particular compound works on this disease in the lab for the tests done for toxicity and other things such as the absorption in the human body, the motor metabolism rate, and so on. Once the compounds show results in these early lab tests, then they are moved to clinical trials to get government approvals. The largest pharmaceutical companies are using AI. Companies like 50 Shades SK are using AI and ML to find new compounds for potential drugs. They are also building models to predict how well potential drugs are going to do in the testing phase. Discovery drugs and their combinations are being developed using AI for combinational treatments and AI is creating personalized medicine based on genetic codes of the patients [9; https://emerj.com/ai-sector-overviews/machinelearning-drug-discovery-applications-pfizer-roche-gsk/]. In the Figure 2-11 below we look at State of AI & ML in Drug Manufacturing. 30 Chapter 2 Key Technological advancements in Healthcare Figure 2-11. State of AI and ML in drug manufacturing Now we look at the state of AI and ML in the area of drug manufacturing. This is an interesting case where our experts, at the end of four iterations, were able to conclusively tell us that drug manufacturing was at all three levels of maturity (i.e., low, medium, and high). So this means that in the area of drug manufacturing, the technology is at research stage and some of it has not moved into production. There is also technology that has been implemented to be tested in production but it has not yet moved into mainstream production environment, such as hospitals. And the experts also tell us that there is technology in drug manufacturing that is well-researched and is ready to move into production or it is already being used in a production environment. In truck manufacturing, robots and AI are being used in the pharmaceutical factories for automated inspection and packaging, leading to efficiency and saving workers from hazardous and repetitive tasks. The robots used in pharmaceutical manufacturing facilities are Cartesian Parallel and Selective Compliance Assembly Robot Arm (SCARA) [10; http://www.pharmtech.com/usingrobotics-pharmaceutical-manufacturing]. Merck is using a robot in its bottling line to place dispenser caps onto bottled allergy medications, and this is providing efficiency to its operations. Robots at Enclave are increasingly being used for vial-filling applications, inspections, and packaging and in various kinds of drug assemblies’ inspections—to name just a few applications. 31 Chapter 2 Key Technological advancements in Healthcare Now we look at the last and the ninth area, which is that of clinical trial research in Figure 2-12 below. Figure 2-12. State of AI and ML in clinical trial research From Figure 2-12, we can see that our experts tell us that it is at a stage of medium technological maturity and the AI and ML state is at that level. After four iterations, 67 percent of the experts in this group were able to conclude this. So this means that there is lot of research happening in this field; however, there are hits and misses, and it needs more careful research in order to move technology into mainstream production. There is research by an MIT robot laboratory that was performed in September 2014 [11; https://news.mit.edu/2014/mit-robot-may-accelerate-trials-for-strokemedications-0211]. There has also been research done by KPMG voice in 2016 [12; https://www. forbes.com/sites/kpmg/2016/12/21/using-smart-robots-to-run-clinical-drug- trials/#749ef31f36d2]. This promises to implement automated clinical trials in order to demonstrate the effectiveness of drug treatment in hundreds of patients. Since the laboratories are still working on this technology, there is little application of it in the real world, but there are lot of experiments ongoing to see how it can help minimize the clinical trials stage of drug discovery. If this technology moves into production, then it will reduce the cost of drug discovery and clinical trials, which can take up to 2 million dollars and is an extremely costly and time-consuming affair. 32 Chapter 2 20 Key Technological advancements in Healthcare Phases of Technology adoption in Healthcare 2018 18 16 14 12 10 8 6 Phase 4 4 Phase 3 2 Phase 2 0 Phase 1 Figure 2-13. Phases of technology adoption in the healthcare industry, 2018 gives us a quick view of the technology adoption process as reported by the healthcare survey in this book. Now we move on to the second parameter of our research study, which is that of phases of technology adoption in healthcare, which we have already discussed in Chapter 1 Figure 1-1 Machine Learning Technology Adoption Process. As you will recall from the previous chapter that there are four phases Phase 1: Quick Applications phase, Phase 2: Early Applications phase, Phase 3: Assisted Applications phase and Phase 4: Independant Operations. Here we have the quick applications in Phase 2 the independent operations in Phase 4 analysis. We can see from Figure 2-13 that the disease diagnosis is at an early application Phase 2 status, as told to us by our expert group. The digital health records are at Phase 2 status, with 56 percent of our experts concluding this after three iterations. Personalized medicine is at Phase 1, which is quick application status, and 83 percent of our experts concluded this after four iterations. The epidemic outbreak prediction area is at Phase 2 for AI and ML 33 Chapter 2 Key Technological advancements in Healthcare applications, and this is concluded by 67 percent of our experts. Our experts also concluded that 56 percent say radiology is at a Phase 3 assisted application stage, and we have seen in technology maturity level that there are various applications by companies like Siemens for use in their Tesla machines. For surgery, our experts conclude after three iterations that 50 percent say AI and ML is at 50 percent for Phase 3, which is assisted application stage. I have seen the application of robotic surgery in mainstream applications like surgery to correct heart disease; it’s a trap. In drug discovery, our experts concluded that the application of AI and ML is and 56 percent of them concluded that it is at Phase 2 early application level. For drug manufacturing, 56 percent of our experts concluded that the use of AI and ML is at a Phase 3 level. For clinical trial research, the use of AI and ML was concluded by our experts to be at a Phase 2 early application level. With this, we end the presentation of the study, which took me more than 3 months to implement along with the experts from the healthcare industry. I do hope that it will provide the reader with a concise view of where the healthcare industry stands with respect to its applications and adaptation of AI and ML. E nd Notes  Study Suggests Medical Errors Now Third Leading Cause of Death in the U.S., May 3 2016, https://www.hopkinsmedicine.org/ news/media/releases/study_suggests_medical_errors_now_ third_leading_cause_of_death_in_the_us  Google Flu Public Data: https://www.google.com/ publicdata/explore?ds=z3bsqef7ki44ac_#!ctype=m&strail =false&bcs=d&nselm=s&met_s=flu_index&scale_s=lin&ind_ s=false&ifdim=region&hl=en_US&dl=en_US&ind=false  Siemens Healthineers: https://www.healthcare.siemens.com/ about  Niramai: http://niramai.com/ 34 Chapter 2 Key Technological advancements in Healthcare  “Google’s AI program can detect diabetic eye disease,” Jul 17 2017, https://research.googleblog.com/2016/11/deeplearningfor-detection-of-diabetic.html. Chapter 2 Key Technological advancements in Healthcare 35  The Delphi Method, https://www.academia.edu/399894/ The_Delphi_Method_As_a_Research_Tool_An_Example_Design_ Considerations_and_Applications  Medrano, I. H., J. T. Guijarro, C. Belda, A. Ureña, I. Salcedo, L. Espinosa-Anke, and H. Saggion “Savana: Re-using Electronic Health Records with Artificial Intelligence,” 27 March 2017, International Journal of Interactive Multimedia and Artificial Intelligence 4(7):8-12. http://www.ijimai.org/journal/sites/ default/files/files/2017/03/ijimai_4_7_1_pdf_22755.pdf  Eliza Strickland, 13 Oct 2017, 19:30 GMT, Robot named STAR (Smart Tissue Autonomous Robot), https://spectrum.ieee. org/the-human-os/biomedical/devices/in-fleshcutting-taskautonomous-robot-surgeon-beats-human-surgeons  Machine Learning Drug Discovery Applications – Pfizer, Roche, GSK, and More, Last updated on November 29, 2018, published on October 12, 2017 by Jon Walker, https://emerj.com/ ai-sector-overviews/machine-learning-drug-discoveryapplications-pfizer-roche-gsk/  Cartesian Parallel and Selective Compliance Assembly Robot Arm (SCARA), By Jennifer Markarian, Nov 19, 2014, http:// www.pharmtech.com/using-robotics-pharmaceuticalmanufacturing  https://news.mit.edu/2014/mit-robot-may-acceleratetrials-for-strokemedications-0211  Using Smart Robots to Run Clinical Drug Trials, Ashraf Shehata, Dec 21, 2016, 11:45 am, https://www.forbes.com/sites/ kpmg/2016/12/21/using-smart-robots-to-run-clinical-drugtrials/#3d61d05336d2 35 CHAPTER 3 How to Implement Machine Learning in Healthcare We now look at the areas of healthcare areas that hold huge potential. For this we need to carefully examine the technology mapping graph in Figure 2-2 of Chapter 2. There are certain areas in the graph that lie in Phase 1 and are low in technological maturity level. Although these hold potential, they do not give us a huge area of opportunity, as the technology is not currently supporting developments in these areas. For example, personalized medicine is very new and there is huge amount of research that must happen, including use of AI, to enable it to move to the next phase. Such research must be linked to the healthcare industry very closely so that the adoption happens faster. Next is the Phase 2 area of epidemic outbreak prediction, which has a few hits and misses and needs to address privacy issues in order to move to Phase 3. The real potential lies in the Phase 3 column of areas, where the technology has moved into the assisted applications stage. reas of Healthcare Research Where There is A Huge Potential I am not going to discuss all the technology used in the Phase 3 column of Figure 2-2, but I am going to discuss the ones that hold huge potential. The three most promising areas in healthcare are: 1. Digital health records 2. Disease diagnosis 3. Radiology © Puneet Mathur 2019 P. Mathur, Machine Learning Applications Using Python, https://doi.org/10.1007/978-1-4842-3787-8_3 37 Chapter 3 How to Implement Machine Learning in Healthcare Digital health records hold huge potential for three reasons: first is that the number of patients for a doctor goes up as the population around the world increases; second, the amount of data per patient is also going to go up; and finally doctors can make good decisions for treatment only if they have the complete case history of the patient. There is another reason why digital health records hold huge potential. The advancement of machine learning and artificial intelligence allows analysis of data in a quick and efficient manner. The hospital or healthcare facility of the future will access digital health records or patient records that will be digitized and accessible to all the healthcare facilities when a patient has authorized the facilities to see its records. In essence, we would have connected hospitals, which would access common electronic patient health records in addition to storing private data of its patients. In the US, there are electronic medical records that are mandatorily maintained by any practicing doctor. However, this data is not available and shared among hospitals, the government, or any other agency or individual who wants to treat the patient. So in the future there is going to be a change to a universal database of patients, which will enable any healthcare facility authorized by the patient and the government of the country to access common case records. The major challenge that could stand in the way of such a data-sharing mechanism involves concerns around data privacy. However, if proper authorization and security of data can be guaranteed by the central nodal agencies and governments, then it would be possible very soon to house such a case history database, and it could lead to building global digital health record systems. In the near future we could see better record-sharing mechanisms being developed in technology, leaving concerns such as data privacy and security behind. As a result, instead of hospitals working in silos, they would start to work in a connected manner. In a hypothetical scenario, there could be a patient who, after being treated by a particular hospital and not being satisfied with the line of treatment given, would like to change their hospital or healthcare facility. In such a case, they would just have to give their consent to sharing all the healthcare diagnostic reports with the new hospital, and once this is done, the data would automatically be accessible by the new hospital to which the patient is about to be admitted. There is another challenge that currently does not allow the concept of connected hospitals, and that is of disconnected diagnostic centers. The diagnostic centers are the labs that cater to the patients’ needs of getting their lab tests done, such as blood or urine tests. They give out reports based on their findings from the lab tests. So they generate a lot of data after the lab tests are done, in the form 38 Chapter 3 How to Implement Machine Learning in Healthcare of measures in the values in the reports. This data that is generated in each of the reports for the patient is currently done by printing the reports in the form of PDF files or through printing reports on paper. There is no central mechanism where the diagnostic centers can upload the patient reports, and there is no current benchmark in place in the healthcare industry that can enable them to use the same standards for publishing their reports. All this will have to be standardized, including the format of the lab test reports and the central database where the electronic record of the patient will get updated. In order to move from Phase 3 to Phase 4