将人工智能带入临床一线——UpToDate

欢迎回到《医疗创新播客》。

Welcome back to the Medical Innovation Podcast.

我是普兰哈·桑吉。

My name is Pranha Sangvi.

非常高兴能和霍莉·厄本医生对话。

Really excited to be speaking to Holly Urban, Doctor.

霍莉·厄本，原本是儿科医生，现在是沃尔特·克卢维尔的业务发展副总裁，也就是大家熟知的UpToDate的母公司。

Holly Urban, the a pediatrician turned vice president of business development at Wolters Kluwer, or as you may know it, the parent company of UpToDate.

霍莉，今天能请到你真是太好了。

Holly, it's so great to have you today.

非常感谢你邀请我。

Thank you so much for having me.

非常期待这次交流。

Really looking forward to the conversation.

是的。

Yeah.

我们先从了解你开始吧，Holly？

How about we get started by learning a little bit more about you, Holly?

能跟我们说说你的经历吗？为什么选择儿科，又是怎么走到今天的？

Can you tell us about your background, why you went into pediatrics, and how you got to where you are now?

是的。

Yeah.

当然可以。

Absolutely.

我是第三代医生，这听起来有点不可思议。

So, I'm a third generation physician, which is crazy.

我祖父是全科医生，他有两个儿子，也都当了医生。

My, grandfather was a GP, and then he had two sons who both became doctors.

我叔叔有七个孩子，五个成了医生；我爸爸有四个孩子，两个也走上了医路。

My uncle had seven kids, five of whom became doctors, And my dad had four kids, two of whom became doctors.

所以，这简直就像我们家的家业。

So it was sort of the family business.

我选择儿科是因为我喜欢与人打交道，当然，我也特别喜欢孩子。

I went into pediatrics because I I love the people, and, of course, I I love children.

我当时在一家大型医院负责质量管理，同时还在做儿科临床，那会儿联合委员会推出了用药核对流程，这在用药安全和患者安全方面是个巨大的进步。

And I was working in quality management at a large hospital when I was still practicing pediatrics, and this was about the time that Joint Commission rolled out medication reconciliation, which, you know, was a really a huge leap forward in terms of medication safety and patient safety.

这也是我第一次有机会思考：我们怎么用技术来提升患者安全？

And it was sort of, the first opportunity to say, how can we use technology to help with patient safety?

当时那家医院有自己的IT部门。

So at that hospital, they had their IT department.

他们设计了一整套用药核对的解决方案。

They designed a whole solution for doing med rec.

然后就把方案丢给了质控部门。

They threw it over the fence to the quality department.

他们对我说：‘乌尔班医生，你能去这家719张床位的教学医院把这套系统推行下去吗？’

They said, doctor Urban, can you go roll this out at this 719 bed academic hospital?

结果搞得一团糟。

And it was a disaster.

对吧？

Right?

完全没有临床人员的参与。

There was no clinical input.

我一直问，他们为什么没这么做？

I kept saying, why didn't they do this?

他们为什么没那样做？

Why didn't they do that?

我们当时有大量书面投诉。

You know, we had bitter complaints on paper.

但这让我真正明白了，如何帮助改善患者预后，协助医护人员做出更好的决策，甚至大规模提升患者安全。

So but it really led me to sort of understand how I could help impact, you know, patient outcomes and help clinicians make better decisions and even patient safety at scale.

就在那时，我转行进入了医疗信息技术领域，到现在已经做了大约二十年。

So it was at that point I then pivoted to working in health care IT, which I've been doing for about twenty years now.

我一开始在麦肯锡霍瑞森临床系统工作，那是一个电子病历系统，现在已经不存在了，当年和Cerner和Epic是竞争对手。

And I I started at McKesson Horizon Clinicals, which is an EHR that is really no longer around anymore that used to compete with Cerner and Epic.

现在那个系统早就被弃之不用了，成了EHR历史的尘埃。

It's now now in the the dustyard or the graveyard of EHR has gone by.

之后我去了Zynx和MCG工作，这两家公司都是赫斯特健康旗下的循证医学内容平台。

From there, I moved to working at Zynx and MCG, which are two evidence based content companies under Hearst Health.

后来我担任了Oracle Cerner的首席医学信息官。

And then I was CMIO at Oracle Cerner.

纵观我的职业生涯，我始终在思考：如何让这些技术真正帮助医护人员更好地照顾患者？

And and throughout my career, I've really been passionate around how can this technology actually help providers take better care of patients?

我们怎样才能大规模地改善治疗效果？

How can we impact outcomes at scale?

这正是我特别兴奋加入UpToDate的原因——它提供的临床决策支持和指导意见，正如你所知，是无可比拟的。

And so that's one reason why I've been so thrilled to work at UpToDate because, you know, the clinical decision support, the guidance that UpToDate provides clinicians is, you know, as you know, is unparalleled.

行业标杆。

Gold standard.

看到你如此积极地参与并推动医疗科技的发展，真的让人非常钦佩。

It's so great to see how active of a role you've taken in developing and contributing to health care technology.

我一直坚信，医生、临床人员在开发影响医患生活的医疗工具时，必须拥有发言权和决策权。

I've always been a big proponent of physicians and clinicians and providers having a say, having sort of decision making authority when it comes to developing the tools that are used for patient care that impact the lives of physicians and patients.

看到越来越多像你这样的医生积极参与到这些工具的开发中，真的特别棒。

So it's so great to see more and more physicians like yourself getting involved in actually developing these tools out.

是的。

Yeah.

不，普拉纳夫，你说得完全对。

No, Pranav, you're exactly right.

我刚入行的时候，做这类工作的医生少之又少。

Now I was one of the I mean, when I went into it, there were very few physicians who were doing this kind of work.

我几乎是自己蹚出一条路，所以看到年轻医生们对这件事充满热情，意识到他们能为临床解决方案带来实实在在的价值，我特别振奋。

I I kinda had to blaze my own trail, so it's really exciting to me to see younger physicians really get excited and understand how what they bring to the table is really material in terms of the solutions that are then deployed, for other clinicians.

我完全同意你的看法。

So I I agree with you.

这确实是个巨大的趋势和转变，而且是令人欣喜的变化。

It's it's been a it's been a big trend, a big change, but I think a welcome one.

我挺好奇的，到底是什么让你选择了 UpToDate？

I am curious what drew you to UpToDate specifically.

你在健康科技领域，甚至临床决策支持领域，应该也有不少机会吧。

You probably had your fair share of opportunities in the health tech space and even the clinical decision support space.

说实话，我一开始是作为用户接触到 UpToDate 的。

You know, I I came to UpToDate as a user.

它一直都被视为行业里的黄金标准。

You know, it was just always the de facto gold standard.

就像你说的，就算别的公司说想做个类似 UpToDate 的产品，他们也永远追不上 UpToDate。

Like, you know, even if other companies said, oh, we wanna do something like UpToDate, they can never catch UpToDate.

而且这个职位本身特别吸引我，是因为我在 UpToDate 的工作重点就是创新。

And the reason the the role itself was so fascinating is because my focus at UpToDate is on innovation.

所以作为医生，我主要跟我们的编辑团队一起工作，这个团队里有医生、药师和护士，他们负责撰写临床内容。

So as a physician, I work primarily with our editorial team that has physicians, pharmacists, nurses, all who write the clinical content.

所以我的主要工作就是跟他们一起探讨新方案，看看怎么利用 AI 或者整合临床决策支持系统，让 UpToDate 从一个单纯的参考网页变得更活跃、更智能。对我来说，这简直是我对临床决策支持的兴趣和我对电子病历工作流的了解这两者的完美结合，感觉就是我职业生涯中完美的下一步。

And so my primary job is working with them on new solutions on how we can get UpToDate, you know, from the world of a referential web page and get it more, you know, get it more activated using AI or even using integrated clinical decision So to me, was like a perfect combination of my interest in clinical decision support alongside my knowledge of EHR workflows that it just seemed like a a perfect way a perfect next step in my career.

这很有道理。

That makes a lot of sense.

把你已有的经验和知识用起来，去改变医疗服务的提供方式，以及医生和医护人员获取信息的方式。

Take what you already know, the experience you have, and use that information, that knowledge to make an impact in how care is delivered and how physicians and providers get their information.

这太棒了。

That's that's amazing.

但我很好奇，这在你日常工作中是怎么体现的？

But I'm curious, how does that translate into your day to day?

你平时的一天都是怎么过的？

What does a typical day look like for you?

天哪。

Oh my gosh.

呃，说实话。

So I oh, shit.

真的没有哪两天是一样的。

It it's one I would say no no two days look alike.

我跟一个临床团队紧密合作，他们主要负责信息学相关的工作。

I work very closely with so I have one clinical team that's focused on informatics work.

另一个团队专注开发集成临床决策支持的内容，还有一个团队负责我们的新AI产品——UpToDate Expert AI。

I have another clinical team who's focused on building content for integrated clinical decision support, and then another team that's focused on our new AI solution, which is UpToDate Expert AI.

这些临床团队的成员，正是这些项目的核心和灵魂。

So they're all the clinical fork folks who are sort of the heart and and the heart and soul of those opportunities.

所以现在有好多事情在同时进行。

So there's lots happening.

每天这三个领域都有新动态，特别有意思。

So every day, there's always something going on in all three of those areas, which is a lot of fun.

这肯定让工作充满新鲜感。

That keeps things interesting, I'm sure.

每天有点变化总是好事，能不断面对新问题，迎接新挑战。

Having a little bit of variety in the day to day is always great, so you're always challenged and tackling new problems and new run tiers.

我们听众大多对UpToDate很熟悉，但我还是想请你用自己的话，介绍一下你们的产品。

So most of our audience is pretty familiar with UpToDate, I'm sure, but I'd love for you to, in your own words, explain your your product offerings.

大家通常知道UpToDate是医疗界的‘维基百科’，但你们新推出的AI功能可能没那么多人了解。

I think there's a, you know, typical what UpToDate is known for sort of being the Wikipedia of health care, but you also have new AI features that may not be as widely known about.

所以我很想请你分享一下看法，帮我们的听众更深入地了解UpToDate的产品。

So I'd love for you to share your perspective and help our audience learn more about UpToDate's offerings.

是的。

Yeah.

我们传统的UpToDate主要为临床诊疗提供决策支持。

Well, so our legacy UpToDate is all about clinical decision support for the point of care.

我们有56位医生编辑负责内容审校。

We have 56 physician editors who edit the content.

还有7400多位领域专家参与内容撰写。

We work with 7,400 experts in the field who write the content.

内容每天工作日都会更新，非常及时。

It's updated every business day, so it's very, very current.

对很多人来说，它已经取代了翻教科书或查文献的需求。

And, you know, for a lot of people, it it's it replaced sort of the need to go to it back to a textbook or to do a literature search.

它的内容经过专家精心筛选和把关，确保随时可用。

It really has that curated content based on expert opinion and expert expertise to have it ready.

我们最令人兴奋的突破是利用人工智能，不再只是让人翻阅文章、查阅资料，而是提供更自然的对话式体验。

And so the the exciting thing that we've done is is to leverage AI, in in to get away from that sort of referential reading an article, context and getting more into that AI conversational experience.

我们对这项技术采取了谨慎推进的方式。

And we we've taken sort of a slow approach to it.

说实话，我很庆幸我们这么做了，因为技术发展到今天，我们的AI系统能够完全基于UpToDate的内容运行。

You know, we didn't and and I'm glad we did because the technology has evolved in a way that we're able to do it in a way that our AI solution is a 100% grounded on UpToDate only.

它不会去互联网上乱找参考资料，完全不依赖外部信息。

So it's not going out to the world wide web looking at gosh knows where for references.

它只专注于UpToDate的权威内容。

It's just focused on UpToDate.

还有一件特别有意思的事，我们没有简单地问：

The other important thing, which I think is super interesting, is instead of just saying, well, hey.

怎么用大模型基于UpToDate生成一个回答？

How can we use an LLM to create a generative response based on UpToDate content?

我们当时想，嗯。

We said, hey.

怎么用这项技术，让医生的使用体验更有引导性？

How can we use this technology to make it more of a guided experience with the clinicians.

不只是简单地回答一个问题，而是设计成能展开讨论的形式，比如：

So instead of just, I'm gonna answer a question, I'm gonna pad it in a way that we can talk about, hey.

有时候临床情况其实有很多细节是你需要留意的，或者：

Sometimes there's clinical nuances that you need to think about, or, hey.

有没有其他相关主题是你可能没考虑到，但其实应该关注的？

Are there other topics that are closely related to this that maybe you hadn't considered and you should?

这就像是你带着病例去请教主治医生时，他会提醒你：从临床角度看，还有哪些关键点是你可能忽略了的？

So so we it's more of a like like a like an attending when you're bringing a case to them saying, how what are the other features that are really critical from a clinical standpoint that you maybe should be thinking about and that you hadn't?

所以，从单纯生成回答，升级到这种有引导性的互动体验，我觉得特别有意思，客户反响也非常好。

So to move it not just to generate a response, but to this this sort of guided experience, I think, has been really exciting and very well received by our clients.

确实非常对。

That's very, very true.

这很有道理。

Makes a lot of sense.

如果你想想为什么UpToDate推出AI功能比一些新兴初创公司慢，从直觉上来看，背后其实有很多合理的考虑。

And if you think about why UpToDate was a little bit slower to release AI features compared to, you know, some of the newer startups out there, I think if you think about it intuitively, there's a lot of logic behind it.

UpToDate有自己的声誉要维护。

UpToDate has a reputation to maintain.

你们已经存在很久了。

You've been around for a long time.

你们已经深度融入了医疗系统和电子病历系统。

You're deeply integrated into health systems and EMRs.

所以你们必须非常谨慎，不能冒任何幻觉的风险。

So you have to be very careful, and you can't really risk any hallucination.

在理想情况下，绝对不能出任何差错。

You can't have any anything go wrong, really, in the ideal situation.

因此，你们必须比那些一无所有的初创公司更加谨慎，设置更多的安全机制。

So you have to be more careful, have more guardrails in place compared to a brand new startup, which has nothing to lose.

所以他们可以直接发布产品，进行测试，看看效果如何，并从错误中吸取教训。

So they can just launch, release their product, and test it and see how it goes and learn from their mistakes.

确实是这样。

Yeah.

我的意思是，这可以说是以安全为前提的创新速度。

I mean, you know, it's innovation at the speed of safety.

对吧？

Right?

我是说，对。

I mean Right.

硅谷那种‘快速行动，打破常规’的做法，并不一定总适合医疗行业，因为医疗行业监管严格、风险很高。

It's sort of the Silicon Valley move fast and break things doesn't necessarily always go well with health care, which is heavily regulated, high risk.

所以，这就是我们当初为什么进展得那么缓慢的原因之一。

So so that's one reason why we we why we did go very slowly with it.

这也是为什么三十年来，医生们一直使用 UpToDate 的原因，因为他们信任它。

And that again, that's the whole reason, you know, for thirty years, providers have used UpToDate is because I trust it.

所以我们想确保这份信任也能体现在我们的解决方案中。

So we wanted to make sure that that same trust would be reflected in the solution.

借助这些新技术，我们能确保方案深深扎根于 UpToDate，从而保持这份信任。

So, and with these new technologies, we're able to make sure it's so grounded on UpToDate that that trust is maintained.

对。

Right.

而且我觉得，拥有那么庞大的信息库也是个巨大的优势。那些文章由超过 75 到 100 位临床医生共同更新，这让 AI 或智能体能直接从中获取所需信息，而不是去网上漫无目的地搜索零散的文章。

And I think having that huge sort of repository of information, the articles that you're over 75, 100 clinicians have sort of updated is a huge benefit as well because you have all the information needed that the agents or the AI can pull from instead of just trying to search through random articles or online for information.

没错。

Absolutely.

有了这个背景，还有一点很重要，就是如果你只盯着零散的文献看，你就没法从整体上把握患者的具体情况。

And with that context, that's the other piece is that, if you're just looking at isolated pieces of literature, you're not broadly thinking about the overall context of the patient.

比如，对于一个肺炎患者，你得想想，他是不是免疫力低下，是不是孕妇，是不是小孩，或者是不是住院病人？

So, you know, for a patient with pneumonia, like saying, well, are they immunocompromised, or are they pregnant, or are they a child, or are they hospitalized?

你知道的，所有这些临床因素都会影响你的判断。比如，你会问：那我该怎么治疗这个肺炎患者呢？

You know, all of those clinical factors are gonna change what if you ask, you know, well, how do I treat this patient with pneumonia?

你得考虑到所有这些细节，而我们正是通过内容把这一切都体现出来了。

You need to be thinking about all those nuances, and we know that because we reflect all of that in the content.

这很有道理。

That makes a lot of sense.

你能再详细说说专家AI工具后台是怎么运作的吗？

Could you go into a little bit more detail about what's going on behind the scenes in the Expert AI tool?

假设我正在接诊一位76岁男性患者，有高血压和糖尿病病史，现在因胸痛前来就诊。

Let's say I'm seeing a patient, a 76 year old male with history of hypertension, diabetes, who is presenting for chest pain.

如果我把这些信息输入到专家AI工具里，后台具体是怎么处理的？它是怎么在UpToDate里找到相关文献，然后迅速生成直接帮到我临床决策的建议的？

If I type that into the Expert AI tool, what's going on behind the scenes to process the information, find the relevant documents or articles within UpToDate, and produce output that's directly relevant to my clinical decision making in a moment?

是的。

Yeah.

这是个非常好的问题。

That's a really good question.

我们确实利用了一些新技术，令人兴奋的是，我们现在能结合多种不同的技术手段——不过我不会称它们为‘智能体’，因为它们并不是独立运作的。

So we've we've and the other thing that's been exciting about some of the new technologies is we're able to sort of use multiple different, I won't call them agents because that's not quite right because they're not acting independently.

但我们有不同的模块，比如先解析你的问题，弄清楚你真正想问的是什么。

But we have different components that will say, one, we'll sort of break down the question to figure out, you know, what you're really asking.

举个例子，如果你输入 HTN 代表高血压，系统会自动把它转成完整的‘高血压’，再交给大模型处理。

You're you know, like, I'm gonna say if you put an HTN for hypertension, I'll translate that to hypertension for the LLM.

这部分是其中一步。

So there's that piece.

然后还有另一个小模块，会扫描所有可能相关的主题内容。

And then there's another little component that goes through and looks at all of the different topics that may be relevant.

接着第三个模块会从中筛选出和你问题最相关的部分。

And then there's a third piece that sort of pulls out the ones that are relevant to the question.

系统内还内置了处理机制，一旦检测到患者身份信息，会自动清除。

We also have embedded in the system if there's PHI, we strip it out.

当然，我们也不会特意提醒你别输入患者信息，那太生硬了。

So, you know, we obviously don't, you know, say don't put in PHI.

你第一次登录时，系统会显示相关免责声明。

We have disclaimer when you first log in.

我们的网页上也有免责声明。

We have disclaimers on the web page.

但如果有人不小心输入了，系统也会自动把它过滤掉。

But if people do inadvertently do that, we strip that all out.

所以后台其实有好几个环节在协同工作，因为我们希望整个过程能像跟医生随便问个问题那么简单。

So, so there's multiple sort of pieces happening in the back because one thing, we wanted it to be as easy as a question and an answer with a clinician.

我们不希望医生还得想着怎么写提示词。

We didn't want the clinician to say, I have to think like a prompt engineer.

年轻人可能还行，能当提示工程师，但像我这样的老年人，根本搞不定这个。

You know, a lot of younger folks, you probably could be a prompt engineer, but, like, old ladies like me could not be a prompt engineer.

所以我们想确保他们只需要像平常问医生那样提问题就行，完全不用琢磨该怎么问才对。

So, we wanted to make sure that that they could just do that, you know, ask a question and not feel like they had to understand the best way to ask the question.

他们怎么想就怎么问，完全不用改。

They could just ask it the way they would ask any other clinician.

对。

Right.

说实话，绝大多数医生根本不懂AI，也没时间去钻研怎么写提示词。

I mean, the vast majority of clinicians don't have expertise in AI or not spend the hours trying to learn how to pump engineer.

他们就只想把问题输进去，然后得到答案。

They just want to Right.

就像问普通医生一样，问什么就得到什么，这样才合理。

Put in their question and get the answer as if they're asking what you so that makes sense.

我之前在Health Planner AppToDate上看过你们的Expert AI演示。

And I did get to see a demo of your Expert AI back at the Health Planner AppToDate.

真的很不错。

It was really cool.

让我印象最深的是那个‘假设说明’框。

So one of the things that stood out to me was the assumptions box.

我输入问题后，它会告诉我系统默认做了哪些假设，比如我可能默认患者是男性而不是女性，或者有胸痛病史之类的。

So if I put in my prompt, it tells me what assumptions are actually being made so that I I can double check, you know, I was assuming that, I don't know, a male instead of a female, or they have chest some history or something.

能清楚看到系统是怎么从我的问题一步步推导出答案的，这种透明度太重要了——我觉得所有做医疗AI的人都明白，医生根本不信任AI这种黑箱。

So the fact that you can double check that and see how it's actually thinking through the whole prompting going from your question to the actual answer is a great move in transparency that I think everyone, I guess everyone who's building healthcare has realized that clinicians don't trust the black box that AI typically is.

他们希望看到每一步的推理过程，明白系统是怎么得出这个结论的。

They wanna see the step by step process and understand how it came to the conclusion it did.

否则，他们不敢用在病人身上，毕竟关系到病人的安全。

Otherwise, they won't use it for patient care because it's a risk for their patients.

我想听听你对这个问题的看法。

So I'd be curious about your thoughts there.

你是怎么考虑透明度的？怎样才能让医生真正信任它？

Like, how do you think about transparency and sort of making it trusted by clinicians?

是的。

Yeah.

这方面有几个要点。

So I so so a couple things there.

我之前提到过系统处理内容的不同步骤。

One is I mentioned those different sequences of of how it looks at the content.

作为用户，你可以打开解决方案，看到每一个步骤的具体内容，清楚它参考了哪些信息、分析了哪些部分。

You as a user can open up the solution and actually see every single piece of that that sequence, so you know exactly where it went, what it looked at, what content it reviewed.

最后，我们会在回答末尾引用UpToDate的文章。

And then, obviously, we cite the UpToDate articles at the end of the answer.

但你可以看到，比如我输入了一个问题，系统会分析76份不同资料，这些内容你都能在系统里清清楚楚地看到。

But you can see, like, you know, I have I put in a prompt, and they'll look at 76 different pieces of content, and you can see all that in the solution itself.

所以我们不想让它变成一个黑箱。

So we we we wanted it to not be black box.

我们希望用户能一目了然地理解整个过程。

We wanted it to be totally apparent to users.

至于假设部分，我觉得这些假设特别重要，比如考虑一个门诊中复杂的尿路感染病例。

And then regarding the assumptions, I think the assumptions are really powerful because, you know, you think of an example like a a complicated UTI in the outpatient setting of care.

这个病人该怎么治？

How do I treat this patient?

但如果患者是孩子，答案就完全不同了。

Well, that answer is totally different if it's a child.

如果患者怀孕了，情况也不一样——如果是非孕妇、普通成年人，通常会用氟喹诺酮类药物。

If the person is pregnant, you know, like because it just if it's a nonpregnant person, if it's an adult, you typically do a fluoroquinolone.

但如果患者是孕妇，那答案就完全不对了。

But if that patient's pregnant, that's the wrong answer.

所以我们非常清楚这一点，这是基于内容正文中的上下文信息。

So that's why we we want and we know this because of the context within the the body of the content.

正因如此，我们才能非常有把握地判断哪些假设是关键因素。如果某个假设无关紧要，我们就不会把它提出来。

So that's how we're able to very confidently know what assumptions matter and what you know, we won't if an assumption doesn't matter, you know, then we don't we don't put it forward.

但如果某个假设会直接影响最终答案，我们就必须确保患者所处的具体情况与所给出的解答完全对应。

But if it's relevant in terms of it's gonna change how that answer is given, we wanna make sure that what the patient you're thinking of is the same thing that's being answered.

顺着这个话题，大家提到医疗AI时，经常会担心或问到一个问题，那就是偏见。

Going off of that, one sort of concern or sort of one question that's brought up a lot when it comes to health care AI is the idea of bias.

比如，你看大部分医学文献的研究对象，大多是白人或者肤色较浅的男性。

For example, like, if you look at vast majority of medical literature is done on, for example, Caucasian or light skinned males for the most part.

女性代表性不足，有色人种的代表性也不足。

Females are underrepresented, people of color are underrepresented.

我敢肯定，UpToDate的文章里也难免存在这种情况，不管我们愿不愿意接受。

I'm sure that carries through to, you know, up to date articles whether or not we like it or not.

由于这个原因，总会有一些影响，AI的输出里也会体现出来。

There's always some impact cut because of that, and that comes through in the AI output as well.

那你怎么判断，在某些情况下，种族、民族或者其他因素该不该纳入分析，又该在哪些情况下排除？你们又是怎么尽量减少这种偏见的呢？

So how do you think about whether or not, you know, some ethnicity or race or different factors should be included in the analysis or, you know, left out in certain cases, and how do you try to mitigate the bias as much as possible?

这个主要是在编辑层面来处理的。

Well, that that's something that's really done at the editorial level.

我说的编辑，指的是那些撰写临床内容的医生作者和医生编辑。

And and by editorial, I mean by the the physician contributors and the physician editors who write the clinical content.

他们心里始终装着那些被忽视的群体，努力确保核心内容能真实反映这些人群，做得更好。

So they have as top of mind, if there are underrepresented underrepresented populations, they wanna make sure they're reflecting that in the core content and and do better.

举个具体的例子，几年前我们有一部分皮肤科内容，里面会展示皮损的图片，比如黑色素瘤或者皮疹的照片。

So as a as a very concrete example, a few years ago so we have some dermatology content where we'll have pictures of lesions, so pictures of a melanoma or pictures of a rash.

我们发现大多数图片都是白人皮肤的，于是有意识地增加了不同肤色人群的图片，包括不同深浅的肤色，就是为了尽量纠正这种偏差。

We realized most of them were on white skinned individuals, so we have made a conscious effort to include pictures that reflect people of color, different levels of melatonin, excuse me, so that so that so that you so that to try and address that bias.

这些调整都是在内容制作初期就完成的。

So that's all done upfront in the curated content itself.

所以，这也是为什么我觉得基于这些精心整理内容的AI更可靠，因为在内容创作阶段，我们就已经在努力减少偏见了。

So, again, that's another reason why I feel like AI that's grounded on this curated content is better because we're trying to mitigate bias at the human level when the content is written.

这个例子让我特别有共鸣，因为当年在医学院的时候，我曾经做过一个项目，收集了数百种疾病——包括黑色素瘤——在不同肤色人群身上的表现，从浅色皮肤到深色皮肤都有。

That example actually very closely resonates with me because back in medical school, I actually created a resource where we compiled representations of hundreds of medical conditions, including melanoma, on various physiotherapist skin types, so light skinned to dark skinned individuals.

那是个非常棒的举措。

And it was a great initiative.

我们当时就是想提升医学教育中的多样性，让医学生、医生、住院医师这些人都能接触到更全面的案例。

We were trying to increase diversity in medical education for medical students, physicians, residents, and beyond.

所以你在这方面做出这样的努力，真的让我特别印象深刻。

So it really stands out to me that you made that effort or that you you're making that effort.

是的。

Yeah.

说实话，我觉得人为的修正比AI自己去调整更有效。

I I mean, I I honestly think a human fix is that better than the AI would fix.

所以

So

AI的表现完全取决于它所训练的数据，这说得通。

AI is only as good as the data is trained on, so that makes sense.

确实非常对。

Very, very true.

我在考虑如何在临床中使用AI时，还经常想到速度和准确性的平衡问题。

I think one other thing that I've thought a lot about when it comes to using AI in the clinic in the clinical space is speed versus accuracy.

比如在病人床边时，你当然希望响应越快越好，总不能等三分钟才得到结果吧。

Like, when you're at the bedside, you obviously want things to be fast because you don't wanna wait three minutes for a response.

你希望输入问题后能立刻得到答案，但同时又不能牺牲准确性或输出质量。

You wanna be able to put in your answer and get it as quickly as possible, but at the same time, don't wanna sacrifice accuracy or the quality of the output.

我想知道，你和UpToDate是怎么考虑这个问题的？怎样才算找到合适的平衡点？

I'm curious, how did you and UpToDate think about that and what's the right balance?

是的。

Yeah.

我们最初推出UpToDate Expert AI的时候——这其实没多久，就在去年夏天——我们的唯一重点就是准确性。

So when we first launched UpToDate and this isn't that long ago, UpToDate Expert AI, excuse me, last summer, our 100% focus was on accuracy.

我们花了很多精力做对抗测试，制定评估标准，持续验证输出结果，还让真人去检查、实际测试系统。

So we did a ton of work red teaming the solution, writing rubrics to validate the output and to do that continuously, to have humans get in there and look at it and kick the tires.

所以一开始，我们完全把重点放在准确性上。

So we really focused up front on accuracy.

等我们越来越有信心，也收到使用该系统的客户反馈，确认准确性没问题后，才把注意力转向提升速度。

And and then as we got more confident and got back from our from our clients who are using the solution, that the accuracy was there, then we sort of turned our focus on making sure that the speed was there.

一开始，说实话，速度有点慢，用户都忍不住说：‘别磨蹭了。’

Because initially, frankly, it it was a little bit slow, you know, slower than people sort of like, come on.

快点吧。

Let's get on with it.

我还有事要忙呢。

I got stuff to do.

但这一点，我自己也亲眼看到了变化。

But that is something that that I've seen personally over time.

现在速度已经提升太多了。

It has gotten so much better in terms of speed.

所以现在真的特别棒了。

So so so it's so it's it's fantastic now.

实际上，现在速度非常快，这太好了。

Actually, it's very fast now, which is great.

但关于速度，还有一点很重要，就是我们推出新产品或新功能的速度有多快。

But the other thing about speed is how quickly we re release products or additional features to the solution.

比如药物剂量的计算。

So, we're looking at things like drug dosing.

药物剂量这个问题，其实还是关乎患者安全。

And, you know, drug dosing is back to talking about patient safety.

这可是个巨大的风险。

That's the that's a, you know, that's a huge risk.

正因为患者安全的考虑，我们到现在还没敢上线这个功能。

So that's something that we haven't felt comfortable even releasing yet just because of the patient safety implications.

所以我们必须确保药物剂量绝对准确，才会正式发布。

So we're having to make sure that we're gonna get that drug dosing right before we release it generally.

是的

Yeah.

这说得通

That makes sense.

我认为在医疗领域，安全永远是最重要的

I think safety is the number one priority when it comes to anything in health care.

患者永远是第一位的

So patient comes first toe wings.

你们是怎么考虑测试的？比如在药物剂量这方面，你们有没有一组临床医生或医生帮你们测试？是先推给内部测试用户，还是有专门的团队负责测试？

How do you think about sort of the testing that, like, do you have a group of clinicians or physicians that are testing it for you when you like, talking about drug dosing, for example, is it, like, releasing beta users at first or do you have internal teams that are more to testing?

你们怎么确保测试的准确性，以及如何收集反馈？

How do you think about testing in for accuracy and just getting feedback?

这里涉及多个层面

So so there's multiple dimensions here.

首先是内部团队

So the first is we have internal.

我们有54位全职医生负责所谓的红队测试。

So we have 54 employed physicians who do what we've called red teaming.

这些红队测试有时会故意问一些有偏见的问题，比如‘北卡罗来纳州的首府是哪里？’

And some of that red teaming is even, you know, trying to ask biased questions or ask, you know, what is the who's the you know, what's the capital of of North Carolina?

像这种无关紧要、不相关的问题。

Like, non important, you know, nonrelevant questions.

就是故意找茬，看看系统会不会出错，测试就从这里开始。

Like, just kinda break it, right, so that so that it starts there.

同一组人，加上我们那7400位专家贡献者，一起制定了评估标准，明确指出针对某个问题，答案里必须涵盖哪些关键因素。

That same group along with our expert contributors, that network of 7,400 have written rubrics where the what the rubric says is for a question, these are all the factors that need to be reflected in the answers.

比如说，如果要治疗社区获得性肺炎，答案就得提到哪些抗生素，得涉及哪些检测，诸如此类。

You know, if I say if I'm gonna do an antibiotic for CAP for community acquired pneumonia, it needs to reference, you know, these different antibiotics, it needs to reference testing, you know, whatever the case may be.

我们积累了大量这样的评估标准，都是内部逐步完善出来的，然后持续用它们来检验模型更新后的准确性。

And so we have this large, large body of rubrics that have been internally developed to inch and then we run that continuously as we update the model to make sure that that accuracy is there.

我们还拿自己的系统跟一些外部基准做了对比测试。

We also have benchmarked ourselves against some of the external benchmarking.

所以你可能知道，那些像USMLE的题目，其实并不是真正的USMLE题目。

So you might you know, the USMLE like questions, they aren't USMLE.

它们只是模仿USMLE的风格。

They're sort of model on USMLE.

我们也拿这些题目做了基准测试，结果表现得非常好。

So we benchmarked against all those as well, and it and it's performed really well against those.

最后，也是我前面说的，最重要的是我们的终端用户。

And then finally, and like I said, probably most importantly is our end users.

当临床医生在实际使用时，每次提问都可以给我们点赞或点踩。

So as clinicians are using it at the point of care, they have an opportunity with every question to give us a thumbs up or thumbs down.

如果他们点了踩，并且指出严重问题，我们每天都会密切关注。

And if they give us a thumbs down and they notice a critical concern, we watch that every single day.

我们会立刻处理，因为你知道，安全永远是第一位的，尤其是在临床决策中。

We address it immediately because we know that, you know, safety is paramount, especially when you're talking about clinical care.

幸运的是，这类情况我们遇到得非常少。

And and, fortunately, we've had very few of those.

我们遇到的那一个，发现是底层内容有问题，需要修正，这点挺有意思。

And the the the one we did have, we recognized it was sort of an issue with the underlying content that need to be addressed, which is interesting.

没错。

Right.

这是我们非常重视的一件事。

So we it's something we're passionate about.

我们很快会发布一篇白皮书，详细聊聊我们是怎么用评分标准来评估安全性和测试效果的。

And we'll we'll have a white paper coming out shortly where we're gonna be talking more about the rubrics approach just to share with everyone kind of how we've looked at it from a safety and testing standpoint.

这读起来应该挺有意思的。

That seems like an interesting read off.

留意一下这篇论文吧。

Keep an eye out for that paper.

但你提得非常好，归根结底，大语言模型或AI的输出并不是完全可预测的。

But you brought up a really good point that at the end of the day, LLMs or AI's outputs is not fully predictable.

即使两次输入完全一样的数据，输出也可能会有细微差别。

Even if you give the exact same data two different times, you'll get two slightly different outputs.

它永远不可能完全可预测。

It's never fully predictable.

我很欣赏你们设立了这么严格的评估标准，确保输出质量稳定，同时密切关注用户反馈，一旦出问题能迅速修正，防止再犯。

So I really like that you have these strict rubrics in place to make sure that there's a certain quality of output and that you're keeping an eye on the user feedback so that if something does go wrong, you can quickly fix it and make sure it doesn't happen again.

毕竟，说到底，总会有某些地方出错，或者产生意想不到的输出。

Because at the end of the day, unfortunately, something will eventually go wrong or something will will be outputted as not expected.

所以必须快速响应，从中吸取教训。

So you have to react fast and learn from it.

是的，没错。

Well, yeah.

我经常跟首席营销官和医疗系统的负责人聊到这一点，他们总说，这些AI工具不是装上就完事了。

I mean, and this is something I talk a lot about with CMOs and health system leaders like to say, you know, these AI tools are not set it and forget it.

你知道的。

You know?

它不像心电图机，插上电就能一直用，直到机器报废为止。

It's not an EKG machine that you plug into the wall, and you're gonna get EKGs for as long as the life of the EKG machine.

你知道，模型可能会出现偏差。

You know, these there's risk of model drift.

你知道，这本来就是概率性的，所以它总会变化。

You know, there's it's it's probabilistic, so it's always gonna change.

所以我们才需要持续运行这些评估标准和其他测试，不断监控，确保……

So that's why we have this continual process of running these rubrics and other tests that have they're ongoing to make sure that that

安全性和质量始终得到保障。

safety and that the quality remains intact.

而且技术本身也在飞速发展。

And the technology sort of innovates so quickly as well.

现在几乎每周都有新模型发布。

There's new models coming out pretty much every week these days.

所以要不断测试新东西，看看哪些效果更好。

So make sure you're constantly testing new things and seeing what works better.

我们该怎么调整模型，或者改一改温度参数，还能做些什么，让输出更可靠、更稳定？

How can we fine tune it differently or change the temperature or just what can we do to make the output even more reliable and more consistent over time?

没错。

That's right.

完全正确。

That's exactly right.

是的。

Yeah.

最让人兴奋的是，我们能利用这项技术飞速发展的优势，最终帮助临床医生更好地照顾患者——这始终是我最关心的底线。

It is exciting that, just to be able to leverage how fast this technology is evolving, but in a way that, again, ultimately is gonna help clinicians take better care of patients, that's always bottom line for me.

对。

Right.

我想换个话题，聊聊实际使用它的人。

I think I'd love to sort of shift gears towards people actually using it.

首先，我想听听，只要能分享的话，现在到底有哪些人在用它？

First, I'd love to hear, much as you're allowed to share, who is actually using it?

UpToDate Expert AI 是否已经部署在所有使用 UpToDate 的医院里？

Is UpToDate Expert AI released at every hospital that has UpToDate?

是分批次推出的吗？

Was it done in waves?

大致来说，上线过程是怎样的？

Sort of, how did the rollout process look?

情况并不统一。

Well, we it it it has been variable.

我们一开始只在个别机构的测试用户中小范围发布，不是用于临床，只是为了评估它的质量和准确性，以及对用户的价值。

We we started early on just releasing to sort of test users at individual institutions just, again, to get a sense and not in clinical care yet, just to get a sense of of quality and accuracy and and value to their to their use.

当反馈普遍非常积极后，我们就和每一个企业客户沟通，问他们希望怎么推广这个工具。

And and as as that began to, really come back very, very positive, what we did is we worked with every one of our enterprise customers to say, how do you wanna roll out this tool?

你是想先在信息部门选五个人试用？

So do you wanna start with a group of five in your informatics department?

还是先让心内科医生使用？

Do you wanna start with just your cardiologist?

还是直接全院推广？

Do you wanna roll it out system wide?

所以这完全取决于各个医疗系统和企业客户各自的选择，决策差异很大。

So it's really been a very variable decision, a 100% driven by how the health care system, the enterprise client wants to handle it.

另外，我们也意识到，现在许多系统都已经建立了相当完善的AI治理机制，所以我们从一开始就主动配合这些流程，确保客户对使用和全面部署这套系统感到安心。

Because the other piece that we've also done is we've recognized that so many systems now have pretty robust AI governance systems in place, And we've wanted to make sure up front we're going through those governance processes to make sure that they would have the comfort level that they needed to use the system, deploy the system system wide.

据我最近了解，已经有五十家医疗系统在全机构范围内使用它了，这让我觉得特别振奋。

And as of last I heard, we have 50 systems, 50 health care systems that are using it all across the system, which to me is just fantastically exciting.

正如你提到的，光靠一个人点头是不够的。

As you alluded to, you don't need just one person to approve it.

你得让一个医疗系统内的多个部门都同意，比如首席信息官或首席技术官得批准它接入UpToDate，还得让一线临床医生真正信任UpToDate，愿意用它来辅助诊疗。

You have to get approval from several different parties within a single health system, whether it's a chief informatics officer or chief technology officer who's allowing it to even be a part of their UpToDate, and all the way down to the actual end users, the clinicians who need to trust UpToDate enough to rely on the AI or use it for their patient care.

所以这一步真没那么容易。

So it's not an easy obstacle or easy step to take.

是的。

No.

没错。

That's right.

我们还做了大量说明工作，告诉大家，这就是目前的实际情况。

Then there's been a lot of education to say, you know, this is you know what to date.

这是UpToDate正式推出的功能。

This is granted on UpToDate.

但说实话，我亲自演示过，也跟很多首席医学信息官、首席医疗官和医疗管理者聊过，他们真的特别兴奋。

But, honestly, like, I've demoed it, and I've talked to so many CMIOs and CMOs and health care leaders, and they're just excited.

他们其实早就盼着UpToDate能做这件事了。

They're they've, you know, like, we've been waiting for UpToDate to do this.

所以，别拖了。

So, like, come on.

赶紧把东西交到我们手上吧。

Get it get it in our hands already.

我在医疗IT行业干了二十年，这点我跟你说过。

So they it's I've I've been in health care IT, I told you, for twenty years.

我从来没遇见过哪个产品像Expert AI这样受到如此热烈的欢迎。

I have never had a product that's been as warmly welcomed as Expert AI.

我是真心这么觉得的。

I mean, I say that totally honestly.

真的让人特别兴奋。

It's been really exciting.

不是的。

No.

就连我也没想到，医生和整个医疗系统会这么快就接受人工智能，毕竟传统医疗行业一向以行动迟缓、抗拒新技术著称。

Even I I was surprised how quickly physicians and health systems as a whole adopted AI because traditional health care is known for moving very slowly, adopting technology very slowly.

你想想，我们当初用电子病历系统花了多久才普及起来？

It's it's how long did you even go start using EMRs in the first place.

确实如此，一点不假。

So it's Very, very true.

说实话，真是个意外的惊喜。

Pleasant surprise, honestly.

是的。

Yeah.

有点不一样。

Something different.

而且，这可能反映出一种未被满足的需求——人们之前得不到及时的答案，而像Expert AI这样的工具正好能填补这些以前没被满足的需求。

And, you know, maybe it's a signal of an unmet need, you know, that that people weren't getting what they needed in terms of quick answers or, and and so, you know, a a tool like Expert AI is is able to meet some of those needs that hadn't been met previously.

我不想读一大篇内容。

I don't wanna read a long topic.

我就想要个快速的答案，这正是它能提供的解决方案。

I want a quick answer, and and this is the solution with that.

没错。

Right.

所以我能赶紧做完工作，回家陪家人，不用整天待在医院里。

So I can finish my work quickly, go home and spend time with my family, or not be in the hospital twenty four seven to pretty much.

那最终用户，也就是医生们，第一次使用UpToDate Expert AI时，反应怎么样？

And how did the end users, clinicians, the physicians actually react to UpToDate Expert AI the first time they used it?

他们当时是怎么想的？

You know, what were their thoughts?

他们有什么问题或顾虑吗？

What questions or concerns do they have?

各种各样的都有。

It's been a wide variety.

有些人，嗯，正面的反馈主要是认可我们之前提到的那些假设。

You know, some people so so I'd say on the good side, people really like the assumptions that we talked about already.

对于那些临床提示，大家的反应都很积极。

A really good response for some of the clinical nudges.

我们每次回答结束时，都会提示用户：你还有没有其他需要考虑的？

So we finish every question with a nudge to say, are there other things you need to think about?

我常举的例子是，有人问：今天发现身上有蜱虫，

So the example I always use is if you put in a question about patient found a tick today.

该不该查莱姆病抗体？

Should I check for Lyme serologies?

答案其实是否定的。

The answer is actually no.