AI醫療革命：GPT-4與未來

by Peter Lee, Carey Goldberg, Isaac Kohane

匯聚科學、醫學、人文三大領域權威
一場全人類、全臺灣
都應參與的關鍵對話
比爾．蓋茲 重磅推薦 山姆．奧特曼 專文盛讚

　　就在幾個月前，數以百萬計的人被ChatGPT的驚人能力——以及它詭異的幻覺所震驚。但那是2022年。GPT-4現在已經來了：更聰明、更準確、有更深層的技術知識。GPT-4及其競爭者和追隨者正處於改變醫學的邊緣。但由於生命受到威脅，你需要立即了解這些技術。

　　它們能做什麼？
　　它們還不能做什麼？
　　它們不應該做什麼？

　　要做出決定，就要親自體驗最先進的技術。加入三位已提前數月接觸GPT-4的業內人士，他們將揭示其巨大的潛力——改善診斷、總結病人就診情況、簡化流程、加速研究，以及更多。您將看到真實的GPT-4對話——未經排練和過濾，既精彩又冒失——所有這些都帶有寶貴的生成背景、坦率的評論、真正的風險洞察力和最新的情報。

　　本書帶您看見與思考：
　　● 通過真正的 AI 助手預覽醫生一天的生活。
　　● 了解 AI 如何增強床邊和其他地方的醫患交流。
　　● 了解現代 AI 的工作原理、為什麼會失敗，以及如何測試它以建立信任
　　● 賦予患者權力：改善可及性和公平性，填補護理和支持方面的空白，促成行為改變。
　　● 通過「及時工程」提出更好的問題並獲得更好的答案。
　　● 利用 AI 減少浪費、發現欺詐、簡化報銷並降低成本費用。
　　● 以 AI 作為研究合作者，優化臨床試驗並加速治愈。
　　● 找到正確的護欄並獲得對監管機構和政策制定者。
　　● 描繪可能的未來，思考接下來可能會實現什麼夢想？

Genres: Artificial Intelligence, Medicine, Science, Technology, Nonfiction, Health, Health Care

328 pages, Paperback

Published September 5, 2023

About the author

Librarian Note: There is more than one author in the Goodreads database with this name.

Reviews

[Jonathan · Rating 4 out of 5]

Reads a bit like a 280-page ad for OpenAI and GPT-4, but very good overview of this technology’s potential and current limitations in medicine, written by experienced veterans who have had a lot of time to think about this.

Some notes:

- GPT-4 and LLMs like it will have a far-ranging impact on medicine --> lots of amazing examples in the book of how it can already be used today
- A helpful analogy: mobile phones and the internet democratized access to information; GPT-4 is democratizing access to intelligence
- For the foreseeable future, GPT-4 can’t (and shouldn’t) be used in medicine without direct human oversight
- GPT-4 is already quite good at reviewing itself and can be used as a reviewer for its own responses
- Despite how good it is, it’s incredibly hard to systematically evaluate GPT-4 for medicine for a ton of reasons - e.g. because its domain of expertise is so broad and because its answers to the same prompt change every time you prompt it
- Clinicians shouldn’t use this technology in medical cases that lie outside their ability to solve or verify —> if you can’t verify, then don’t trust the results
- LLMs that are developed for, or adapted, modified, or directed specifically toward medical purposes are likely to qualify as medical devices (and they will likely be regulated as such)
- We’re currently in a “Wild West” in terms of AI regulation and LLMs; hopefully, there will be no incidents involving morbidity or mortality, and we’ll be able to learn from the “embarrassing incidents” (quoted from expert in book) that do occur
- There WILL be newer and more powerful tools; the pace of new AI model deployments will undoubtedly continue to accelerate --> whatever assumptions we have about the limitations of AI today are unlikely to hold up tomorrow

Favorite quotes:

“Let us not be seduced by our own creations, nor allow their allure to blind us to our moral compass.”

“[W]e must come to grips with the fact that GPT-4 represents a technological phase change. Previously, general intelligence was frozen inside human brains, and now it has melted into water and can flow everywhere.”

[ameya · Rating 3 out of 5]

some interesting use cases but others that would be just as efficient without an LLM. also little consideration of externalities.

[alexander · Rating 3 out of 5]

This is the best account I’ve found so far of what will really change in medicine, authored by experienced veterans who’ve had the time to really think about this.

In short they really buy into it, but I continue to have a hard time actually piercing through the illusion of self that Large Language Models seem to have. What actually is “I” in their responses, particularly as we don’t know the data it is trained on? The authors try to get GPT-4 to say what it could be used for, and so of course it generates plausible text of which there is a lot n here, which I slightly tuned out of towards the end, and some confusing nomenclature around davinci, chatGPT and gpt-4 (which one should a patient use? If you Google it they all come out) which is further padded with the tendency to recommend double checking GPT-4 by running another session in parallel, which seems very non-Steve Jobsian.

[Nilesh Jasani · Rating 4 out of 5]

The AI Revolution is an engaging and thought-provoking read, offering insightful examples that highlight the potential of artificial intelligence. The authors eloquently capture the essence of their argument in the middle of the book:

"Medicine traditionally refers to a sacred relationship between a doctor and a patient — a twosome, a dyad. "And I'm proposing that now we move to a triad," he said, with an AI entity like GPT-4 as the third leg of that triangle.

Today's LLMs are likely to appear elementary in a few years. As impressive as their feats are, as shown in the book, they still have much to demonstrate in order to surpass the expertise of our finest medical professionals indefinitely. Even if they claim to outperform the average practitioner, it is natural for many of us to harbor reservations and doubt their abilities, regardless of the irrationality behind such sentiments. Nevertheless, the book masterfully showcases the incredible potential of integrating GPT as the third agent in the doctor-patient dynamic. From aiding in diagnosis, documentation, and explanations to serving as an error handler, facilitating patient-doctor communication, optimizing planning, and enhancing overall efficiency—the possibilities are vast. Furthermore, the book hints at the future prospects of LLMs as long-term record-keepers and even contributors to drug discovery, further emphasizing their potential value.

It is pretty likely that healthcare and pharmaceuticals emerge as generative AI's most significant application sectors over time.

[Andrei Gavrila · Rating 4 out of 5]

I really liked it. Not only does it shed light on applications of GPT in medicine but it does it in a smooth, easy to read and understand way.

This book is great for all audiences and people that never heard of GPT or only know very little about it will be able to understand how impactful it can be for us all.

[Jonathan · Rating 3 out of 5]

Finished in one sitting. Wanted to love it but disappointed by lack of explaining llm and gpt until last 1/3 of book. No credence or respect to anyone other than Microsoft for creating ai in a vacuum. Spent 60% of the book displaying gpt responses.
Good nuggets inside it.

[Bryan · Rating 4 out of 5]

Caveat: I work with the author. I also work in the field of artificial intelligence.

This book describes the problems and decisions that medical professionals will have with these advanced generative models. The models will make mistakes. That will likely lead to some bad outcomes. The models will also get a lot of things correct.

One of the thought exercises in the book is about people who have no access to health care professionals. A person could get answers from a generative model, that could actually save their life. The models will most likely save more lives than it costs. But where is the line of how many mistakes it can make. So, should we let people use them? Some of these questions are very difficult to answer. Do we feel better if someone dies because a human doctor made a mistake, than if an AI model makes a mistake? If the model saves 99/100 people and the human saves 97/100, why does is still seem wrong to use the model because the one person may not have died. This book will make you think.

The authors make a pretty compelling case for these models to be accessible to people who would otherwise not have any access to medical advice. But it is scary to implement something like that.

[Aluniuha · Rating 4 out of 5]

Zdecydowanie ważna pozycja dla osób z branży medycznej. Zawiera dużo informacji jak można korzystać z GPT-4 i jakie są jego możliwości i ograniczenia. Przestrzega przed opieraniem się na odpowiedziach podanych przez AI i przypomina, że to my, jako eksperci, powinniśmy weryfikować jego odpowiedzi. Dobrze się czyta, miejscami tro je się przeciąga, ale to wynika z moich preferencji i zainteresowań.

[Sophia · Rating 5 out of 5]

Fascinating book on AI and what potential it holds for our lives especially in healthcare. I love sci-fi and many of the concepts are happening for real in our lives now. Also a bit scary if not used for the betterment of humanity.

[sara rimer · Rating 5 out of 5]

Engaging, informative, easily understandable, must-read

To borrow a term from Carey Goldberg’s smart, lively prologue about GPT-4’s impact on a fictional but totally plausible physician, this book makes me feel…augmented, at least when it comes to learning about AI and medicine. It’s a nuanced guide to the possibilities, practicalities and potential pitfalls of the revolution that is already upon us. Can GPT-4 really empathize w humans? Read this book and find out.

[Scott Pearson · Rating 5 out of 5]

Artificial Intelligence is changing the way information is handled worldwide. The advances pose particular opportunities for medicine, where descriptive texts are the norm, research expands knowledge exponentially, and paperwork is a main product. Of course, new dangers uncover themselves, too. Will AI merely exacerbate existing health inequities, or will it provide better quality care for anyone with a smartphone worldwide? These subjects need to be thought through in order to secure positive outcomes. These authors, with a foot in medicine and a foot in computing, pen a fascinating early take on these dilemmas for us to ponder.

Early on, businesses have capitalized on AI opportunities the most, with AI-driven chat bots popping up everywhere. Because Chat GPT-4 has been trained extensively on medical research, they actually can serve as great healthcare assistants. They are wrong sometimes – and sometimes widely wrong. Yet Chat GPT-4 passes medical licensing exams better than most physicians, so it might already outperform most doctors, especially with rare diseases. This book discusses these ethical dilemmas so that everyone can access this discussion. It also glances at how doctors could improve the quality of patient experience by getting AI to do paperwork.

Inspired and curious, I typed in my medical history about a condition that I’m exploring with my physician. The workup I was given was far more thorough than anything even my excellent physician could offer. It had me, a biomedical software developer, dreaming of how to leverage the technology for the ultimate good of patients. It can even modify the sentence structure and language to address those at lower literacy levels instead of science geeks like myself.

As with any new technology, challenges await, yet at this juncture, the opportunities seem great. Workers in the healthcare sector and healthcare users in the general public might be able to leverage AI for good if it’s addressed properly. Frankly, it’s a matter of survival for both parties. This book can help readers think through these thorny issues at an early gate so that they – we – won’t be left behind. AI can prove to be an effective helper, but like anything in life, wisdom comes from sorting out the good from the bad. This book definitely equipped me to do just that.

[Daniel Mosby · Rating 4 out of 5]

Given this is an AI book, I used AI to help me summarise my learning points from the highlights I made as I was going along. This book is just under 2 years old, but you can still tell when reading it that things are advancing so quickly even this book is out of date already, albeit only in some limited areas. I'd say it's still worth a read for anyone interested in the intersection of AI and medicine.

Medicine demands a partnership between humans and AI rather than AI replacing medical professionals. AI should support healthcare providers, not act autonomously in critical decision-making.

The medical training of AI systems matters, just like the training of human doctors.
The scope of AI’s domain expertise is difficult to fully evaluate, creating risks in medical decision-making.
AI should embody principles like transparency, accountability, and collaboration to be used responsibly.

AI can help alleviate the paperwork overload that consumes nearly half of doctors’ and nurses’ time, contributing to burnout.

AI-generated medical information must be verified before use. A separate AI session can be used for error-checking, but final validation should always involve a human. Human-in-the-loop oversight is mandatory for high-risk applications.

GPT-4 is better at reviewing text than creating it. It can translate between health data standards, advancing interoperability in healthcare. AI is prone to hallucinations and subtle errors, making it unreliable for autonomous medical decision-making.

AI lacks true moral reasoning—it cannot reliably navigate ethical dilemmas. AI systems trained solely on language models struggle with decision-making in morally loaded situations.

Low-risk AI uses (e.g., writing insurance letters) require minimal oversight. High-risk AI applications affecting patient care must have mandated human review and liability.

AI should be viewed as a “cognitive forklift”—lifting the heavy mental workload so humans can focus on complex, uniquely human tasks. AI will not be risk-free, but neither are common medical tools like aspirin.

[Brad · Rating 3 out of 5]

A pretty quick read recommended by someone at work as we pilot use of GPT and brainstorm ways to appropriately use an LLM in medicine. So far it's looking helpful for non-clinicians as we (understandably) haven't let any LLM loose on patient data. I'm more interested in the clinical side and look forward to seeing what we can do as augmented clinicians.

Fully agree that--at least in current state of LLM and other AI--there needs to be a human in the loop to monitor and sanity check outputs. Unfortunately this very factor of not being able to fully trust the tool is one of the things that most hampers it in this field.

Lee and his colleagues are pretty rosy about how much AI can and will do. Also, they seem really impressed by things that seem dumb to me--the analysis of how the doctor did in a visit and where they could improve? That's 1st year med school stuff. I agree more with the "will" side than the "can" side, but perhaps that's because we haven't been able to do any clinical augmenting yet? Several parts of the book also read like an advertisement for GPT-4 rather than LLMs more broadly, which makes sense given that at least one of the authors works for Microsoft.

They anthropomorphized the heck out of GPT-4. I guess it kind of makes sense given the conversational nature of interfacing with GPT, but it seemed weird. It's a machine--it doesn't need pleases and thank yous. *eye roll*

Major concerns about some of the errors that GPT-4 made and that LLMs in general can make. Sometimes doing math wrong, confabulations / hallucinations. Needs a lot of work and and solving before this is ready for the big time. I do agree with what the authors state near the end, that we need to engage *now* in figuring out how to use and regulate this tool, but it's moving quickly.

Rating: G, maybe PG because of some discussions of health issues?

[Larry Mastin · Rating 5 out of 5]

Despite the title, this book is less about AI in Medicine than about GPT-4 and how it could be a medical game-changer. The first chapters illustrate the intelligence and foibles, of GPT-4 through a series of illustrative conversations with it. Its mind-blowing sophistication comes through when posed with moral questions, such as “is it okay to rob a bank if you're poor?, or require empathy, such as “what would you say to a scared girl with a strep infection?" GPT-4’s ability to infer from context, adjust tone of voice, and speak at a level appropriate to the conversation is light years beyond regurgitating facts. AI models like GPT-4 can ace multiple-choice knowledge tests such as the U.S. Medical Licensing Exam (USMLE, P. 19). But can they understand? Common knowledge has long said that no machine can truly understand. Essay questions are often used to judge an individual’s ability to synthesize knowledge and demonstrate understanding. But LL models now ace these tests, too. So how now do we define “understanding”? As one author notes, “tests from the latest scientific research fail to prove that GPT-4 lacks understanding”.

The medical applications are legion, but the book focuses on a few; diagnosis, patient advice, math and coding, reducing paperwork, accelerating discovery of drugs and treatments, and ethics. As a medical consumer, I thought the most most empowering application was the possibility for anyone to obtain cusomized, sophisticated medical advice. Doctors can dispense this, but in practice, our access to doctors is limited. For many people, it's non-existant. And everyone could use a second opinion.

Throughout these chapters, the possibilities and pitfalls are illustrated through Q & A with the LL model, interspersed with commentary. I found this format surprisingly engaging. GPT-4’s answers always start with caveats, and are generally couched as suggestions or thoughts, more than recommendations or peremptory statements.

There are several weaknesses of GPT-4: it can't ingest new information without being taken offline; it can't store information from one session to another; the amount of information a user can enter is limited; it occasionally fabricates or hallucinates, and it can make simple arithmetic errors. More broadly, LL models are only as reliable, fair, and equitable as their trainers make them. OpenAI doesn't currently disclose the traing data it uses.

Unlike some of the dark books about AI, this one is hopeful. The objective in Medicine is to help people thrive, and this book tells us how this tool moves us toward that goal. I am happy to recommend it.

[Richard · Rating 3 out of 5]

The authors of this book should really be editors, since much of the text is taken verbatim from interactions with GPT-4. That's not necessarily a criticism -- a subject like this begs for specific examples. A book that describes how to use Excel, say, or some programming language, will consist mainly of text directly output from the computer.

But they go a little too far when, for example, in Chapter 9 ("Safety First") they simply create two personas and ask GPT-4 to generate a discussion. The results are interesting and provocative, but I left feeling, well, like why did I need the authors for this? Wouldn't I be better off just asking GPT-4 these questions myself?

That question is a good microcosm of the whole issue of using GPT-4 in medicine. Did I really need these authors to go to the trouble of making a book? That's eerily similar to the question of do I really need a doctor to tell me that these symptoms are congenital adrenal hyperplasia (one of the GPT-4 generated cases they present). How much will I really need doctors now that I have GPT4? Editors after all are useful, to help direct our thoughts in a particular, useful direction. Maybe in the future we'll think the same of doctors.

If you want a good summary of how GPT-4 will change medicine, a more concise summary is in the one hour YouTube talk by Peter Lee: Emergence of General AI for Medicine.

[Josh · Rating 1 out of 5]

This is techno-optimism at best, and at worst is a poorly written advertisement for OpenAI.

To their credit, they do acknowledge some of the shortcomings of LLMs, even if they downplay them, and even if they claim that GPT-4 is not "just a large language model". But even their discussion of GPT's shortcomings sounds like "AI makes dangerous and concerning mistakes in medicine. On that point, we queried GPT-4: ..."

This book also suffers from having several authors. These authors have a wide range of opinions and stances on the topic discussed in this book, and it is hard to keep track of all the different voices. Peter's chapters, for example, seem to argue that GPT-4 is a conscious being that is capable of understanding and reasoning. He is excited about his "relationship" with this language model, and sings its praises profusely. I suggest you skip Peter's chapters, by the way. But "Zak" is much more reserved, and even makes tongue-in-cheek comments about Peter's enthusiasm. All of this causes this book to be inconsistent and confused.

All around, I might have given this two stars if it weren't for the continual recurrence of chapters authored by Peter. Anyone who has been on the internet in the past five years has seen random people get excited about AI. We've seen people post their conversations with a chatbot, and yes, we agree that it's cool. But if you don't have anything else to add to the conversation, why write a book?

[TK · Rating 4 out of 5]

It depicts insightful ideas on how technologies like GPT-4, and similar ones, can revolutionize the healthcare and medicine ecosystem. It also provides important negative implications, especially in a consequent field such as healthcare. Even though it felt like an ad for chatgpt, I understand how this kind of technology can have a significant impact on people's health. It would feel incomplete if I had not read Deep Medicine by Eric Topol. Both together picture a pretty concrete overview of what we have ahead of us for the healthcare ecosystem going forward. There was a tiny chapter on the use of this technology in science, but I missed more ideas on this, especially in the intersection of life science, medicine, and AI, and how it's and can contribute in the future.

[Asya · Rating 1 out of 5]

I don’t get a single high rating for this book. It doesn’t properly introduce or really explain what LLMs are, and how GPT4 fits into the landscape. Nevertheless it is extremely focused on GPT4 and a bunch of Microsoft products. The chapters written by individual authors lack cohesion, are often redundant and random on insights. And on top of it most of the book is composed of prompts and GPT4 responses to them. Ok, you can get some ideas into how to engineer medically relevant prompts… and some basic thinking on applications and ‘dangers’ of GPT4 in medicine but this is not a book to get an objective introduction into this exciting topic.

[Richard · Rating 3 out of 5]

A VP for Research at Microsoft, a Harvard Medical School Bioinformatics guru, and a medical journalist were given early access to what we now call ChatGPT-4. The result is cautious amazement at what is possible. There are long sections showing ChatGPT output, which is a bit tiresome, but you'd have to include it if no-one had seen this before. Will this technology revolutionise health care? Well, what a future we might have if (if) it can, but no-one knows how to tell which output to trust and which not to. But, in the other corner, if you're in that 50% of the planet with no health care provision, doesn't this make a huge difference?

[Martyn Perry · Rating 4 out of 5]

The many examples of GPT-4 interactions which dominate this book really do bring to life the amazing potential and risks of AI in healthcare.

At times they’re a little repetitive or long-winded, certainly towards the latter stages of the book, but they’re all necessary to demonstrate just what can be achieved by the AI of today, and to anticipate what the AI of tomorrow could feasibly achieve.

Recommended?:
This is a fantastic book, readable and fascinating, and very up to date with the latest features of GPT-4. It’s a significant achievement, and a useful reference for those working in digital health to understand the art of the possible, and the scale of the risks.

[Marcus Goncalves · Rating 4 out of 5]

The book offers a solid overview of how Large Language Models like ChatGPT might influence healthcare, focusing on enhancing access, reducing administrative tasks, and providing real-time assistance and information to medical professionals and patients. While it explores the potential benefits extensively, it doesn't delve into the technical or philosophical aspects deeply. A significant portion, including hypothetical dialogues on legislation and usage, was generated by ChatGPT, which detracts from its depth. Despite this, it effectively highlights the need for prompt and careful legislative action to harness these technologies' benefits in healthcare.

[Andrew Halterman · Rating 4 out of 5]

A thought provoking book on the benefits and risks using generative AI in medicine. It’s not too hard to imagine inviting generative AI to the table for tasks that present comparatively low risk to patients. It’s the high-risk tasks that are truly intriguing: diagnosis, medical discovery, patient and provider mental and emotional support.

This book contains the best summary of generative AI’s potential contributions to medicine. It provides a very optimistic viewpoint that I think is influenced by a great deal of expertise and also by Peter Lee’s position at Microsoft. Recognizing this, I recommend this book to anyone interested in the intersection of medicine and generative AI.

[Lincoln]

This book will quickly become out of date given how fast the field is changing. That said, I really appreciated a more comprehensive review of current state as well as considerations that will impact the future than I have been able to find anywhere else. Many in my network have read and recommended this book and I consider it required reading if you are involved in data and analytics in healthcare.

[Christine · Rating 1 out of 5]

At the core of medicine is MD-PATIENT dynamic. With dispensing of advice/treatment/procedures, patients do their own processing and healing. Without AI taking any liability for this practice, it will NOT revolutionize how patients are treated at the bedside. This book is Microsoft peddling a product that NO ONE understands completely without taking responsibility for potential security, reliability and privacy issues. It is interesting to see how tech companies want to solve medical problems with tech products....... and brand the skeptics as dinosaurs.

[逸文 · Rating 3 out of 5]

I read it in Chinese version.
The book delves into the potential applications of AI in healthcare, from improving diagnoses and summarizing patient visits to streamlining processes and accelerating research.

One of the standout aspects of the book is its balanced approach to discussing both the incredible potential and the ethical considerations of AI in medicine. The authors, with their diverse backgrounds in computer science, medicine, and journalism, provide a comprehensive view of how AI can democratize access to intelligence, much like how mobile phones and the internet democratized access to information1.

The book also includes real GPT-4 dialogues, annotated with context and commentary, which offer readers a glimpse into the technology's capabilities and limitations. This hands-on approach helps readers understand the practical implications of AI in a medical setting.

Overall, "The AI Revolution in Medicine: GPT-4 and Beyond" is a thought-provoking read that encourages readers to consider the opportunities and responsibilities that come with the integration of AI into healthcare. It's a must-read for anyone interested in the future of medicine and the role of technology in improving patient care.

[Shreyas · Rating 3 out of 5]

This is a great book that dives into the capabilities and potential of GPT-4 in the healthcare workspace, but becomes repetitive at times. The variance in language and technicality is noticeable between the authors and leads to an uneven reading experience. The AI prompts throughout the book offer great insight on what GPT is capable of but can be tiring to read at times. Overall, still a solid read about a new and emerging technology that’s an exciting read.

[Wes Boland · Rating 4 out of 5]

Very solid short read on potential applications of AI in medicine in the future safety/ethical dilemmas within incorporating LLMs/AI in medicine. Would recommend reading to get excited about future use cases to better understand the technology.
Most of the book and applicability will become outdated within a year .

[Sree Gayathri · Rating 5 out of 5]

Overall great summary of areas where LLM can be useful to advance science and medicine. Parts of it felt like marketing material for GPT4 with some emphasis on limitations. Although book did a good job of highlighting how LLM work but still unsure what data is used to train these systems and can it extend already worse misinformation crisis we are facing is a mystery.

[Claire Floyd · Rating 5 out of 5]

Finally! Embrace, collaborate, and relief! AI holds great promise of cognitive burden relief and creates opportunities to enable clinicians get back to bedside and do what we are trained to do! I’m excited and cautious for what this partnership holds!

[Walter D Dixon]

Highly recommended to any health care provider!

I had just finished a book general AI when I discovered this book. First two thirds was very interesting but the end fell flat. Overall, worth reading, but.I would skip last couple of chapters, and good directly to epilogue.