OpenEvidence首席营销官Travis Zack分享的洞见：如何通过AI即时获取临床证据

大家好，我是Andrew Perrinbap，Perra医疗手册播客的主持人之一。在Perra，我们从创意阶段到A轮融资全程陪伴创始人，并很高兴能分享医疗领域开拓者们如何从零到一构建医疗企业的故事。今天我们非常荣幸邀请到Travis Sack医生，他是UCSF的助理教授，同时也是被美国大量医生使用的AI医学搜索引擎Open Evidence的首席医疗官。

Hi, everyone. I'm Andrew Perrinbap, one of the hosts of Perra's Healthcare Playbook podcast. At Perra, we partner with founders from ideas to series a, and we're excited to share stories from trailblazing healthcare founders and leaders on how they built a healthcare business from zero to one. Today, we're so excited to be joined by doctor Travis Sack. He's the assistant professor at UCSF and CMO at Open Evidence, the AI medical search engine used by a huge share of US physicians.

Open Evidence近期宣布完成由GV和Kleiner Perkins领投的2.1亿美元融资，并与《新英格兰医学杂志》《美国医学会杂志》达成合作，还推出了名为Deep Console的新智能体，承诺能在数小时内完成博士级别的文献综述。我们将探讨这对临床决策、医疗安全及临床推理的未来意味着什么。Travis，非常感谢你今天能来参加节目。

Open Evidence recently announced a 210,000,000 raise led by GV and Kleiner Perkins, along with partnerships with NEJM, JAMA, and even came out with a new agent called Deep Console that promises PhD level synthesis in hours. We're unpapping what this means for point of care decisions, safety, and the future of clinical reasoning. Travis, thanks so much for joining me here today.

是的，非常感谢邀请。

Yeah, thanks so much for having me.

Travis，我想深入了解你的背景。你在肿瘤学与AI交叉领域有着令人惊叹的经历，最初是怎么走上这条道路的？

Travis, I want to like kind of dive into your background. You have this really awesome background sitting at the intersection between oncology and AI. How did you even get to that background in the first place?

说来也巧，并非刻意规划。我本科读物理时对医疗领域并无特别兴趣，研究凝聚态物质时主要开发成像技术，后来在生物物理研究生阶段也延续类似方向。读研期间家人遭遇健康问题，促使我转向能解决癌症难题的研究。

I don't know. It wasn't intentional necessarily. I I did my undergrad in physics without any interest necessarily in health care specifically. I was in condensed matter, kind of working on imaging techniques, and I went to graduate school for a lot of the same in biophysics. And during graduate school, my family had a couple health problems that made me wanna focus more on research that that could solve problems in cancer.

我的研究生项目很灵活，允许我采用计算方法和统计模型，将物理学的量化工具应用于医疗问题。后来我意识到，要真正深入这个领域需要医学博士学位，于是又返校攻读MD，始终立志成为肿瘤学家。接触临床数据流后，我发现许多机器学习技术能极大改进现有医疗实践。

So my graduate program was very, like, flexible and allowed me to do these kind of computational approaches, statistical models, and apply a lot of the quantitative tools I learned in physics to health problems. And then that kind of evolved. I was like, you know what? To really get into it, need to have an MD, so I went back and got my MD, always wanting to be an oncologist. And then having more exposure to the clinical data streams in medicine made me realize that a lot of ML techniques would be very valuable in what is currently being done.

大约八年前我开始转向真实世界数据的AI研究。那时医疗AI尤其是自然语言处理方向的研究者还很少，阴差阳错就走到了今天这个位置。

So moved to AI approaches in real world data about eight years ago. And back then, it was kind of a small community. It wasn't a ton of people who were doing medical AI, especially around natural language processing. So ended up that's how I ended up here, I guess.

确实。你比其他人更早地把握住了那波浪潮。你是怎么进入AI领域的，更不用说医疗AI领域了？

Yeah. You you were on that wave a lot earlier than everybody else. How did you even get into the, like, sphere of AI, let alone sphere of AI in medicine?

嗯，我认为关键在于不拘泥于使用什么工具，而是专注于你想解答的问题。早先我们做基因组学研究时，做了大量统计建模和基础生物信息学工作，后来对因果关系产生兴趣，转向贝叶斯推断。当我开始思考数据系统和临床结果预测建模时，很明显机器学习才是解决这类问题的正确方法——当然不是指大语言模型的深度学习，而是用ML技术整合多源数据来做预测。于是我想，好吧，看来我得学机器学习了。

Yeah. Well, it's all about I I think it's all about being agnostic about what tools you wanna use, and instead just figuring out what questions you want to answer. So back when we were doing genomics work, did a lot of statistical modeling, more kind of basic bioinformatics, got interested in kind of thinking about causality, and moved into Bayesian inference, and then once I moved to thinking about data systems and kind of clinical outcomes, predictive modeling, it was very clear that machine learning is the right approach for a lot of those things. Now that doesn't mean, you know, deep learning with large language models, but just the ML techniques are the right approach to combine multiple data streams to make a prediction. So I was like, okay, I guess I'm gonna learn machine learning.

这就是我的经历。我觉得在每个阶段都需要思考什么才是合适的工具，而不是固守过去学到的知识。这就是我走到今天的原因。

And so so that's that's how it was. I think it's just like for each step is just I have to question what is the right tool and not anchoring necessarily on what I had learned in the past. So that's how I ended up here. Yep.

我觉得这是适用于人生的好框架——信息变化如此之快，你不能总是固守旧知，必须持续跟进。对了，你是怎么认识Daniel和创始团队的？这背后有什么故事？

I think that's a really good framework for life in general that information changes so rapidly and you can't always anchor to it. You just gotta constantly keep up with it. And and how did you meet Daniel and the and the founding team? What what was the story behind that?

说来话长，就不细说了。但我觉得两个关键点是：永远愿意接听陌生来电，以及...运气够好。说到陌生来电，当时做医疗AI和自然语言处理的圈子很小。Open Evidence刚成立时基本都是研究生，其中一位联系我讨论临床推理、认知推理，以及用语言模型构建自然语言处理系统——那时候还是GPT-2的时代。

Yeah. It's a long story. I won't get into all of it, but I think the two important points of it are always be willing to take a cold call and be really lucky, I guess. But regarding the cold call, like I said, was a small community who's doing medical AI, natural language processing. Back when Open Evidence was was very new and mostly graduate students, one of them reached out about clinical reasoning, cognitive reasoning, and natural language processing systems using language models, though again, this is very It was like GPT two days, maybe.

更多是BERT。当时大家都在做BERT相关研究。这位来自外州的朋友的朋友说，如果能有个理解我们构建思路的医学博士提供视角就太好了。于是我们开始合作。

More BERT. Everyone was doing BERT stuff. And so this was from around across the country, kind of a friend of a friend. And they're like, well, you know, it'd be great to have an MD who understands kind of how we're trying to build these things to get a perspective on that. So we were collaborating.

那位学生也是Open Evidence的初创成员之一。除了合作外，他经常就他们开发的产品向我咨询各种问题。就这样建立了联系，直到后来他说：'好了，现在这是个正经公司了。'

That student was one of the kind of starting people in Open Evidence as well. So in addition to collaborating, he would often ask me a lot of questions about what they were building there, just kind of curbside questions. And so that's how I got connected. Then eventually, he was like, okay. This is a real company.

这就是我之后要去的地方。我一开始是支持这个项目的。本想继续当顾问，但随着事情变得越来越激动人心，他们逐渐把我拉得越来越深。所以

This is where I'm going afterwards. I was supportive of it. Tried to continue to be a consultant, and they kinda just dragged me in more and more as it got more and more exciting. So

嗯。所以这家公司是在2022年左右创立的。你是什么时候听说的，又是何时逐渐深入参与的？

Yeah. So so the com the company was started in, like, 2022. When did you hear about it, and when when did you kinda get more and more involved?

对。我记得埃里克大概在2022或2023年就提起过这事。那肯定是在GPT发布之前——这对许多相关事物来说显然是个重大转折点。具体时间我得问埃里克他什么时候联系的我，但大概就是那个时间段。

Yeah. I mean, Eric I believe Eric was talking about it around 2022 or 2023. It was definitely pre the GPT release, which was kind of kind of a big, obviously, watershed moment for a lot of these things. So somewhere around there. I'll have to ask Eric for specifics about when he reached out, but somewhere around.

嗯。

Yeah.

我也特别好奇，因为很多收听这档播客的听众都知道公开证据（open evidence），其中不少是临床医生。如果有人还不知道的话我会很惊讶。但为了那些可能不了解公开证据的听众，你能给我们简单解释下它是做什么的吗？

I'm so curious too because there's a lot of folks here that know about open evidence that listen to this podcast. Many many are clinicians, and I'd be surprised if if people don't know about it yet. But for those that may not be aware of open evidence, could you could you kind of break it down for us and and describe what it does?

好的。不过现在变得更复杂了，因为我们正在做越来越多的事情。但我可以告诉你最初的构想。最初的想法是：很多医学问题本质上都是信息检索问题。

Yeah. And it's gotten more complicated because we're doing so much more and more now. But I'll tell you how it was originally conceived. Right? The original conception was a lot of medicine boils down to an information retrieval problem.

正确的信息片段就存在于某处，只是极难找到。或者如果某个信息片段不存在，如何拼凑与之相关的信息来为自己找到答案。我认为信息检索中这两方面同等重要。并非每个患者都独一无二，这意味着我们遇到的很多问题在《新英格兰医学杂志》或某个随机对照试验里是找不到现成答案的。

The right piece of information exists somewhere, it just is extremely difficult to find that piece of information. Or if that piece of information doesn't exist somewhere, how to piece together the piece of information that relate to it to create an answer for yourself. Right? And I would say that there are equal parts of that when it comes to information retrieval. Not every patient's unique, which means that a lot of the questions we have are not in the New York, New England Journal of Medicine, right, or in a randomized controlled trial somewhere.

这个信息检索问题其实是我从实习时期就感兴趣的。那时候当然和我们现在做的完全不同，但核心理念一致——在医学领域实现这一点有两个关键：一是正确答案本身，二是向使用者展示答案的来源。因为作为医生，我不能仅凭一句‘这是正确答案’就治疗病人，对吧？

So this information retrieval problem was actually something I've been interested in since I was an intern. Back then, mean, it was nothing like what it was, what we're doing now, but same idea where you can't There are two pieces to getting that done right for medicine. One is the right answer, and one is showing the person where that right answer lives. Because if I'm a physician, I can't treat any patient on something that just says it's the right answer. Right?

即便在非AI医疗实践中，你也不该这样做。如果有人告诉你‘就该这么做’，你总会问：‘依据在哪？让我看看证据’，对吧？

Even you're not even supposed to do that when, you know, you're practicing outside of AI. Right? Like, if somebody tells you, oh, this is how to do it, you should always be, well, where's your source? Let me have the evidence for that. Right?

让我看到论文原文。我们每次都这么做吗？未必，但理应如此。所以任何面向临床医生的信息检索系统，都必须完美兼顾这两个方面。

Let me show me the paper. Now, do we do that every time? Probably not, but we should. Right? And so any information retrieval process that is meant to provide information to clinicians should have both those pieces and do it very well.

即确保答案永远准确，并确保来源是医生期望的权威渠道。这就是Open Evidence当初构建的理念——核心价值在于：一个能解答任何复杂度医学问题的平台，提供24小时内更新的正确答案，并附有可追溯验证的参考文献，实现‘信任但验证’。当然我们现在业务范围更广了。

That is, make sure it's always accurate, and make sure it's always pulling it from the right piece of, you know, where the physician would want that information to come from. So that was the idea, and that's kind of what what was what was being built at Open Evidence. And so the the core value ad, the core value proposition is a single place where you can ask any medical question regardless of complexity, and get the right answer up to date by the day, you know, within twenty four hours, and make sure that you have the references or citations for that answer that you can go verify, the trust and verify. So that is what we originally built. Obviously, we do a lot of other things now.

我们撰写医保预授权函，与电子病历系统集成，能在患者诊疗全流程中实现信息检索，包括病历书写等。简化来说，这就像是给医生用的专业版Google。

We write prior authorization letters, we're integrated into EHR systems, you can do that same kind of retrieval across patient context, or writing notes, etcetera. But one for the for for in a overly reductionist view, it's doctor Google for doctors, more or less.

我觉得你指出了一个关键点：你们提供参考文献，不是简单说‘这就是答案’，而是让临床医生自行核查引用文献。这是否是你们区别于其他AI系统的特质？

And I think you hit on a really important point is that you provided references, and it wasn't necessarily like, hey. This is the answer, it's just the answer. But you as a clinician go and do your own reference check, click on the references, and do your own diligence. Yeah. Do you feel like that's what set you apart from a lot of these I

确实如此。在2024年之前——准确说是2024年末——我们是唯一这样做的。现在GPT也会提供参考文献，但你分不清是来自患者门户网站还是真正的PubMed。

actually yes. And so pre twenty twenty four ish, right, like, late twenty twenty four, that was we were still the only people doing that. Right? Now GPT will provide some references. You can't tell if it's from a patient facing site or an actual PubMed.

对吧？而且它有时两者兼顾。但从一开始，对我们来说最重要的是什么？就是不仅要检索到正确答案，还要清晰地展示得出答案的过程。

Right? And it sometimes does both. But from our outset, that was the most important thing for us. Right? Is that you're not just retrieving the right answer, you're making it very clear how you're getting to that answer.

我们依然认为这正是我们的独特之处——是的，现在其他人也在行内添加引用，但普遍存在'垃圾进垃圾出'的心态。而我们非常谨慎地筛选可用参考资料，严格控制模型能使用的参考内容。

And we still feel like that what's that's what sets us apart, is that, yes, now other people are adding citations in line, but there's still the garbage in, garbage out mentality where we are very careful about what you can use as a reference and what our model is allowed to use as a reference.

你们在期刊合作方面确实独树一帜。我甚至听说有其他几家AI公司试图通过砸钱与期刊建立合作，但由于它们的非营利性质，最终选择了与你们合作。

And you guys are really unique with this journal partnerships. I'm sure I actually even heard that there was a couple other AI focused companies that were trying to get these partnerships with journals, even even throwing money at them. But because of their, like, nonprofit status, they didn't end up going with them. They ended up going with you guys. Yeah.

那么从内部视角来看，你认为像《AnyJam》和《Jammin》这样的知名期刊为什么会选择开放证据（Open Evidence）呢？

It's like like, from your perspective, from the insider viewpoint, what why do you think prestigious journals like the AnyJam and and Jammin went with open evidence?

我认为是因为我们的质量。在正式接触前，他们的编辑就了解我们的水平——他们看到答案时会感叹'这些回答太棒了，完全符合标准'。所以他们不需要赌我们会妥善处理其内容，而是早已确信。

I think because of our quality. You know, it's it's we had the quality that they their editors knew us before we came to them, and they fit they're like, we love these answers. We feel like they're doing it right. So they didn't have to trust us that the That we would provide quality with their content. They knew in advance.

换句话说，他们不必依赖信任。他们亲眼所见：我们已经在妥善处理你们的内容，合作只是扩大传播范围的手段。如果非要简而言之，大概就是我们确实做得很好——人们总倾向与已证明实力者合作，而非听信空头承诺。

So I guess I should say they didn't have to rely on trust. They could see it with their own eyes. That we are already doing what we need to do with your content, and this is just a way to expand it and allow more people to see more of it. I think I think that is the the probably I mean, if I'm being very reductionist, that's probably what happened is, well, we just were very good at it, and people like working with people who are already good at it, not people that have to make a bunch of promises.

确实。深入想想，似乎用户群体早已建立起信任，这种信任可能源于同事间的口碑传播。我记得第一次知道开放证据是在ICU轮值时，看到同事用它查询ARDS等病症的最佳呼吸机参数——我就是这样了解到它的。

Yeah. I feel like, you know, tapping into that a little bit deeper, it seems like there was already a lot of trust with people using it, and and maybe that trust was an extension of their colleagues using this or or them seeing, like, this product, like, Evidence spreading through word-of-mouth. I remember the first time I actually learned about open evidence. I was on an ICU block, and I saw one of the fellows using open evidence to search up, you know, like, optimal vent settings for, like, ARDS and stuff. And and that's kind of, like, how I learned about it.

然后我去和其他住户及朋友交谈。当你们最初考虑时，有没有在营销上投入，还是纯粹靠口碑传播？

And then I go talk to my other residents and other friends. When you guys were, you know, first thinking about it, did you guys do any spend on marketing, or was it purely just word-of-mouth?

我想我们确实做过，但几乎完全依赖口碑。说我们做过可能不太准确——不能说完全没投过推特广告。但我们很快就停止了，因为发现相比口碑传播，广告效果微乎其微。所以主要还是靠口碑，但也要强调我们始终认真倾听用户并及时响应。比如我和Zach在住院医师群组做Reddit问答时，只要有人联系我们，我们都会立即发邮件解决问题。

I think we did, but it was almost entirely word-of-mouth. When I say we did, like, it would I think it would be inaccurate to say that there wasn't a Twitter ad. But I definitely know we basically stopped doing it pretty quickly because we realized it wasn't really making a blip versus the word-of-mouth. And so it's all been it's word-of-mouth, but also I would say there's a lot of it has also just been really carefully just always listening to users and always responding. Like, whenever you know, Zach and I did a Reddit AMA with the residency groups, you know, like, anytime someone talked to us on contact, we'd make sure that we emailed them immediately, just addressing their problem.

我们现在仍保持这个传统。很多人发邮件时以为会收到AI自动回复，但其实是我或团队成员亲自回复——经常是我。我们认为这非常重要。

We still do that. You know, it's like we still have somebody somebody a lot of people email and expect like an AI bot to respond. No, it's me or someone else on the team. Often me. And we we feel like that's super important.

这传递出一种感觉：我们不是试图推销产品的庞然大物，而是一群真心觉得这个工具有用才创造它，并持续完善的人。

And it it provides a sense of like, this is not just some monolith company who's trying to sell us something. This is someone who like, a group who actually, like, built a tool because we felt like it would be useful, and we continue to do that.

我很欣赏你们的理念，始终倾听信任公司的用户实在太棒了。你们实现了惊人增长——年增长率2000%，去年35万次咨询，今年850万次？等等，是每月850万次？

I I love to hear your perspective on this because because I think that's so awesome to be always listening to your users and like, the people that trust your company. And you guys have had such incredible growth. I think that's, like, 2000%, like, year over year growth rate, and it was, like, 350,000 consultations last year, now, like, 8,500,000 consultations this year. I'm sure there's so many 8.5 a month. 8.5 a month.

实际上我们上个月完成了1300万次。现在每天处理约65万次美国医疗提供者的咨询问题。

We actually did 13,000,000 last month. Yeah. So so Incredible. That is We do about we do about 650,000 questions a day with health care providers in The United States.

这太了不起了。我突然想到——作为临床医生我们都清楚，医学知识正以惊人速度增长。你们曾提到过，每73天医学研究的引用量就会翻倍对吧？

That's incredible. You know what? One thing that would be really awesome, I'm just thinking, putting on, like, my clinician hat too, who I think both of us acknowledge and and many other physicians acknowledge that the rate of medical knowledge is just continuously increasing at a very, very rapid rate. I think you guys also quoted that what is it? Like, every 73 days, like, the number of citations in medical research, like, doubles.

是的。我想说我们引用别人的话，那是别人说的，我们只是重复。

Yeah. I would say that we quote we say that. That's somebody else said that. We repeat

是啊。

Yeah.

我不确定。你知道，我不太喜欢他们使用的重复统计数据。但我想说的是，我认为没人会否认它在加速发展。是的。

I'm not sure. You know, I I don't love the doubling statistics they use. So but I what I would say is I don't think anyone would argue that it's accelerating. Yeah.

而且跟上所有这些知识太难了，连最新的指南都跟不上。我觉得我刚毕业不久，但在医学院学的一些东西已经过时了。对吧。

It and and it's just so hard to keep up with all that knowledge, even keeping up with the latest guidelines. I feel like that I've learned and I just recently graduated. Right? And I feel like some of the things that I learned in med school is still outdated. Right.

是啊。还没...

Yeah. It hasn't been

那样。实际上这是个双面问题。我在做教育讲座时经常提到——我们正面临双重挑战，对吧？

that way. And it's actually a two part thing. I talk about this when I'm giving educational things. We're we're fighting an uphill battle in two ways. Right?

首先，是医学证据更新速度太快。但其实还有另一方面，就是每条知识适用的患者群体越来越细分。比如一百年前，治疗糖尿病就一种方法。

So one, it's the pace in which evidence change and medicine changes. Right? But actually, there's another piece as well, which is each individual piece of knowledge is applying to a smaller and smaller subset of patients. Right? Like, a hundred years ago, there was like, well, this is how you treat diabetes.

大概就是不用吃药，只要戒糖就行。如果运气好的话，仅此而已，对吧？但现在你得考虑：他们有哪些并发症？他们用过哪些一线药物？

And it was like, probably no drugs, you know, just stop eating sugar. If if you're lucky, then that's it. Right? Now you have to be like, okay, what are their comorbidities? What first line agents did they get?

他们用过哪些二线药物？有肾脏疾病吗？所以当你学到某个知识点时——比如针对慢性肾病患者或糖尿病合并慢性肾病患者的二线治疗方案——这些信息的适用人群会越来越窄，而且有效期越来越短，很快就会被淘汰。对吧？必须明确的是，这对患者而言都是好事。

What are, you know, what second line agents get? Do they have kidney disease? And so when you've when you learn a piece of knowledge, like here's second line treatment for chronic kidney disease, or for diabetics with chronic kidney disease, or whatever it happens to be, that piece of information applies to a smaller and smaller group of people, and it applies for less long period of time before it becomes outdated. Right? So and and those and to be, like, very clear, those are both great things for patients.

对吧？这对患者非常有利，我们可以针对每个病例高度个性化地制定方案，而且支持这种做法的证据越来越多，医疗水平也在持续进步。如果情况没有改善，我们也不会不断更新治疗方案。

Right? Those are great things for patients that we get to hyper personalize exactly what we should do for each case, and there's evidence for exactly how to do this more and more. And things keep getting better. Right? If if they weren't getting better, we wouldn't be we wouldn't be updating things.

所以这两点对患者都很有利，但对需要保持知识更新、确保为患者做出正确决策的医生来说却是巨大挑战。这两方面因素都在给我们制造障碍。而Open Evidence这类AI系统的目标，就是确保决策能基于昨天刚发布的最新数据。

So those are both great things for patients, but they're extreme challenges for the physician who's trying to stay up to date and make sure they're making the right decision for the patient. So, you know, both of those things are working against us. And that's, you know, the goal is using AI systems like Open Evidence to try and make sure those decisions are being done with the data that came out yesterday.

我非常赞同你强调的这点。我们在UCSF和斯坦福工作，能接触到顶尖专家和最新知识，但并非人人都有这种条件。那些在偏远地区行医的医生——可能是方圆几十英里内唯一的医疗提供者——就处于这种劣势。Open Evidence对他们帮助巨大。你觉得这类人群的使用率是否明显高于城市或学术中心？还是参差不齐？

I I love that you highlighted that. And and I feel like from both of us, you know, you working at UCSF and and me being at Stanford, we feel like we have access to, like, top tier specialists or or the latest knowledge, but not everyone has that. And I feel like people, like, in very rural parts of the country that are practicing and maybe the only set of providers or specialists within, you know, ten, twenty, 50 mile radius, they're kind of at this disadvantage. I'm sure open evidence does such a huge benefit to them. Do you feel like there's a huge uptake in using open evidence in those populations compared to, like, urban populations or, like, academic centers, or is it kind of mixed?

不，应该说两类人群都在使用，但方式不同。我们很快会发表相关研究——很有意思的现象。学术中心采用得更快，因为他们接触新事物的渠道更多。

No. Well, I would say that there's uptake on both, but they use it differently. And this is something we're actually gonna publish pretty soon. But but it's very it's been very interesting, you know. I I I would say that the academic centers are taking it out more because they just have access and knowledge about new things faster.

对吧？学术中心对新事物的接纳总是更快。而偏远地区的使用更多是出于必要性，实际产生的价值也更大。斯坦福重症监护室的医生可以钻研PEEP参数之类的问题固然很好，但偏远地区存在一个被很多人虚构的伪命题——初级保健医生总能转诊给专科医生。

Right? Like, just everything gets uptaken faster in an academic center. And then I think the uptake in rural places are more because it's just more necessary, and is actually providing more value. I think that it's great that people in the critical care unit at Stanford can, you know, geek out about PEEP or whatever it happens to be. But but I think in the rural places, I there's this false dichotomy that a lot of people are pretending to exist, is a PCP can refer to a specialist.

或者说，他们可以转诊给专科医生这个想法。但现实是，很多时候要看内分泌科医生，得开车三小时去，还要等六个月。到了那里才发现检查报告不全，三个月后还得再来复诊，因为所需材料不齐。这就是许多专科医疗的现状。因此，初级保健医生（PCP）必须决定在这种情况下该怎么办，他们不愿让病人苦等六个月才能得到治疗。

Or like there's this idea that they can refer to a specialist. And the truth is, a lot of times to get an endocrinologist, it's like a three hour drive away, and it's a six month wait. And when they get there, they realize they don't have the right labs, and they have to come back three months later for a follow-up because they didn't have everything they needed. And that's the reality for a lot of specialty care. And so what that means is the PCP has to decide, you know, what to do in that situation where they don't wanna wait six months potentially for their person to be treated.

很明显，他们一直在利用我们的服务填补这些空白，至少确保当病人见到肿瘤科或内分泌科等专科医生时，PCP已为他们做好准备，能一次性获得所需诊疗。理想情况下，本该一次就诊完成的事，就不需要多次往返。但遗憾的是，我认为很多情况下确实别无选择，PCP只能尽力而为——他们不可能每次都恰好读过美国胃肠病学会的最新指南，对吧？

And so they've been using it's very clear that, you know, we are just a way of of filling in some of those gaps, or at least making sure that when the person sees a specialist like an oncologist, endocrinologist, they are the PCP has them teed up to to get exactly what they need. And so there's no additional, you know, multiple visits to get what should have been a single visit thing. That is the ideal situation, I would say. Unfortunately, I do, I think there's a lot of cases where there's just literally no options, and the PCP is trying to do their very best, and they don't happen to have read the American Gastroenterology Association's latest guidelines every single time. Right?

这让医生能在12分钟的接诊中（当他们要快速处理20个病人时）迅速掌握最新情况，并确保所有诊疗行为都遵循循证指南。

And so this allows them to get up to, you know, get up to speed in a twelve minute visit when they have 20 patients as fast as possible, and provide evidence based, you know, accordance with evidence based guidelines for for everything they do.

听起来真是太棒了。我还想请教的是，你们帮助这么多临床医生为患者做出正确决策，但你们是否追踪过因你们工具或平台而改变的患者管理方案？比如上周末我值班时，遇到个罕见的DORB心脏病例，完全没头绪。

That's really awesome to hear. And, you know, the other thing that I wanted to ask you guys about is you guys help so many clinicians kinda make the right kind of decisions for patients. But do you guys ever hear of or ever track a, like, change in patient care management because of your tool or your platform? The reason is yeah. You know, I was on a P shift literally this past weekend, and I had this rare DORB heart condition that I had no clue about.

患者因呼吸窘迫就诊，该用什么药？当时我的主治医师正在处理镇静和复位，我就直接问了OpenEvidence。结果系统列出禁用药清单——我差点就要用其中一种，真是惊险。

I was like, okay. What what the patient came in with, like, respiratory distress. What what meds do I use? And I literally asked open evidence because my attending was away dealing with the sedation and the reduction. And it had a list of meds that I can't use, and I was like, wow.

这工具算是帮我躲过一劫。不知道你们未来是否会追踪这类案例？

I was gonna use one of those meds. Yeah. It kinda saved me from this from that encounter. I'm kinda wondering, do you guys hope gonna track that in the future?

我们确实在追踪。戴上真实世界数据这顶帽子来说，目前问题咨询、答案生成与实际执行建议之间还存在脱节。因为直到最近，我们还未实现与电子健康记录（EHR）系统的整合。

We do. So, you know, that putting on my real world data hat. Right? Like, we do have a bit of a disconnect between the questions being asked, the answers being generated, and the actual implementation of those recommendations, right? Because up until very recently, we have not been EHR integrating.

对吧？这意味着我们能提供很好的建议。我们能够判断哪些患者有IgG4或其他情况，但缺乏直接证据说明实际发生了什么。唯一的方法就是进行前瞻性试验，或者通过像CMS这样的大规模回顾性数据来探索。所以我们正在筹备前瞻性试验，对此我非常感兴趣。

Right? So that means we can provide a great recommendation. We can tell, you know, which patients have IgG4, or whatever it happens to be, but we don't have direct evidence about what's happening. And the only way to do that is prospective trials, one, or large scale retrospective data through things like CMS, which is what we're exploring. So we are setting up a prospective trial, and I'm very interested in that.

我认为主要目标不仅是改变管理方式，还包括更好地遵循指南。但我特别关注AI或语言模型如何通过权重进行判断——本质上就是‘幻觉’。对吧？而我们医生也时刻依赖着权重判断。对吧？

I think the primary objectives there is not just changes in management, you know, the better concordance with guidelines. But I'm very interested in this idea of how the determine in AI, right, or or language models is pulling from weights, which is basically hallucinations. Right? And we physicians pull from weights constantly. Right?

我们经常这样：医学院学过的内容可能是对的，就直接用于患者。临床决策中90%的情况都是如此。剩下10%不确定时，我们会怀疑：这些权重可信吗？我的训练可靠吗？

We are constantly like, this is what I learned in medical school, it's probably right, this is what I'm gonna give the patient. Right? And we do that 90% of the time we make a clinical decision. And then the 10% of the time we don't, we're like, well, I don't know if my weights, I should trust my weights. I don't know if I should trust my training.

这时就需要查阅资料。但这是个很高的行动门槛。我们想研究的是：降低这个门槛后，医生是否会减少依赖权重？他们是否不再那么信任隐性知识？是否会更多地去寻找循证依据？

I need to go look it up. And then, you know, that is a very high activation barrier. And what we're interested in looking at is, now that we've lowered that activation barrier, do physicians pull from weights less, basically? Do they not trust their implicit knowledge as much? And do they actually go and find the evidence based information for the the question more?

这说明他们本该一直这么做，唯一阻碍是时间有限。如果每个问题都要花30分钟查PubMed或20分钟翻UpToDate，还不一定能找到答案，你不可能对每个患者都这么做。但如果只需5-10秒，人们就可能减少依赖经验，更多转向证据。

Which suggests that they should have been doing it more all along, and the only thing that the only reason they haven't is there's only so much time in the day. Right? If every single question requires thirty minutes on PubMed or or twenty minutes on UpToDate, and you may not find the answer, you're not gonna do that three times per patient. You cannot do it. If each one takes five to ten seconds, you may pull from weights less and go to the evidence more.

以上就是我们这个前瞻性试验要验证的主要假设。

And so those are the primary kind of objectives we're looking at for this prospective trial.

这让我非常感兴趣，因为作为临床医生，遵循相似的思维框架会让人安心。遇到典型胸痛患者时，你很清楚该怎么做。但偶尔会觉得‘这次有点不对劲’，又因为研究起来太耗时而不愿调整方案。是啊，我迫不及待想看到研究成果。

I'm so interested by that because I feel like it's comforting as a physician and as a clinician to kind of operate from similar schema and frameworks. Right? You see a typical chest pain patient, you know exactly what to do because as you've seen it a thousand times. But if there's, like, that one time where you're like, something's a little off, and I don't wanna change it because it's gonna take too long to research it, That's Yeah. I can't wait to see that.

务必把那个发过来。你看，我们在这个话题上已经讨论了很多内容，我真的很感谢你的坦诚和见解。我刚才问的是关于未来的问题。毕竟你在这个领域既在运营又在建设，那么五年后你如何看待开放证据的运作方式？以及你认为临床医生将如何提供医疗服务？

Definitely send that over. You know, I'd love to kind of we we've talked about so many things on this topic, and and really appreciate your candidness and and and your insights. I kinda asked, like, kind of the future. You know, you're you're operating in this space, you're building this space. You know, five years from now, how do you kinda see open evidence operating, and then how do you see kind of clinicians kind of providing care?

是的。我们有一个愿景，在阐述这个愿景之前，我想强调一个非常重要的观点：当我们思考医生与AI协作时，必须明确五年后的目标。对我来说，这个目标就是确保我们仍然有值得信赖的医生掌握主导权。这意味着我们在创建医生-AI协作模式时，不是要外包认知处理，而是要强化和训练医生内部的认知处理能力。坦白说，我提到这点是因为从去年开始，实习医生从一开始就接触开放证据系统。

Yeah. So we have a vision, and the thing that I preface that vision, and something I feel very important is that when we think about physician AI collaborations, I think it's very important to think about the goal in five years. And the goal in five years for me is to make sure that we still have physicians that we can trust in the driver's seat. And so that means that when we're creating a physician AI collaboration, we're doing it in such a way that we're not offshoring cognitive processing, but instead we are actually reinforcing and training cognitive, the internal cognitive processing of physicians. And the reason I would, just to be very transparent, the reason I'm bringing this up is, trainees, starting, you know, as of last year, have had open evidence since onset.

对吧？从他们实习阶段就开始使用。目前系统的设置非常适合能力提升和学习，因为它采用拉取请求模式。你必须主动整合并提炼出针对患者的核心问题，明白吗？

Right? Since they started as an intern. And so and right now, the way it's set up is great for augmenting and learning, because it is a pull request. You have to actually, like, aggregate and synthesize what is the important question for this patient. Right?

你需要把问题输入开放证据系统，当获得输出时还得思考：这个答案如何应用到我的患者身上？这两个都是能真正提升医生能力的主动过程。经历这些需要大量思考和信息处理的认知过程，才能真正掌握这些知识。

You have to put that in open evidence, and then when you get an output, have to say, how does that answer, like, how do I apply that to my patient? Right? So both of those are very active processes that actually improve you as a physician. Right? Going through those things are cognitive processes that require a lot of thinking and and and processing of information, such that you actually will retain that information.

对吧？至少这是正确的使用方式，也是我推荐的方式。虽然我们要向推送模式发展，但...

Right? And at least that's how you should be using it. And that's how I I recommend using it. So I will say we are moving towards more push. Right?

我认为目标应该是：你应该问哪些问题？你遗漏了什么？无论问诊时间是四十秒还是三分钟，系统都能在你完成记录时提示你本该询问的问题，这样你就能及时补问。

I think the goal should be what questions should you be asking? What are the things you missed? You talk to the patient for forty seconds or, like, three minutes, whatever happens to be. What were the questions you should have asked that we can tell you at the point that you finish your transcription so you can go back and ask those questions before it's too late? Right?

这些虽然是推送功能，但我们设计时仍确保医生保持实质性的主导权，认知处理的重任始终在他们肩上。比如现在新版OE问诊系统会在你书写病历笔记时，针对你的诊疗方案逐项询问证据依据。未来几个月还会提示你可能遗漏的病史询问要点和病历中值得关注的潜在问题——不仅因为它们存在于病历中，更因为有临床证据支持需要关注。

And so these are push things, but we're doing it in such a way that we still are keeping the physicians in the driver's seat in a meaningful way that the cognitive processing is still, you know, still on their shoulders. So examples of that that, you know, is actually now, right? This is the now, the new OE visits, is we ask all the questions at your visit. So when you write your note, when we do an ambience scribe, you write your note, we will ask all the questions about your assessment and plan, and what the evidence is for each of the things you decided to do. In the next few months, we will also be pulling out what are the things you should have asked your patient, what are potential things you missed in the chart that you can also identify, but not just because they're there in the chart, because they're there in the chart, and we feel like there's evidence to get it done.

对吧？我认为这些病历摘要方法存在很多问题，往往是因为医生未处理病历内容有充分理由。可能信息就在那里，但内部流程会让人想'听着，我只有12分钟，没法面面俱到'。当前摘要工具的核心问题在于它缺乏依据来声明'我不管你有忙'。

Right? I I think a lot of the problems with a lot of these chart summarization approaches is there are often very good reasons a physician didn't address what was in the chart. You know, maybe it was there, but there's an internal process to be like, look, I only have twelve minutes. I can't address everything. And a lot of the problems with what's currently out there for summarization tasks is it doesn't have the evidence to say, I don't care how busy you are.

这实际上是个关键问题，我们可能认为你疏忽了。但这并非因为你太忙故意忽视或蓄意不做。

This is actually a critical thing that we probably think you missed. It wasn't because you were too busy and ignore and didn't do it on purpose.

这对未来真是令人兴奋。我感觉你们不仅是高风险临床决策支持工具，更像是针对各阶段医生的内部培训工具。因为我现在就得到这种反馈——主治医师会问'你问过这个吗？'

That's exciting things for the future. I I I feel like you're not just like a high risk clinical decision support tool. You're like an internal training tool for the physician at any stage. Because I get that feedback right now. I mean, you the attending's like, did you ask this?

'你问过那个吗？做过这个吗？'而我在想，等我当上主治医师后，可能就得不到这种反馈了。

Did you ask that? Did you do this? And I'm like, when I'm an attending, I probably won't get that feedback. You

确实如此。正因如此，我认为为医生设计的聊天或搜索引擎至关重要。当《纽约客》或《纽约时报》谈论聊天机器人诊断时，他们忽略了很多要点。作为医生，创建鉴别诊断虽稍有帮助——或许能想到几匹'斑马'——但更重要的是，AI工具应在诊断、治疗和管理中帮你思考可能遗漏的事项，并确定最佳后续步骤。

know Exactly. And actually, that's that's why I think it's so important to have, like, a chat or a search engine that is designed for physicians. And I think a lot of the things that like The New Yorker or, you know, The New York Times miss when they talk about chatbots and diagnosis. As a physician, you don't yes, creating a differential helps a little bit, maybe there are a couple zebras you didn't think about. But what's more important, in an AI tool that's working with you for diagnosis, treatment, management, is to help you make a decision about what are potential things you weren't thinking about, and what are the potential best next steps.

比如，我需要哪些信息才能将六项鉴别诊断缩减到一项？这是个迭代过程，不是'我给你所有信息，立即输出最可能的诊断'。而是利用全部医学知识和训练有素的诊断引擎，逐步确定最佳下一步检查，从而更好地区分可能性与非可能性。对吧？

Like, what are the pieces of information I need to actually take this differential diagnosis from six things to one thing? And that's an iterative process that is not, I gave you all the information, spit out the differential with the most likely run. It's an iterative process to be like, look, we have the entirety of medical knowledge as well as our own trained diagnosis engine. What is the best next diagnostic step you can do such that to better bifurcate what it is versus what it isn't. Right?

因为你可能无法完成潜在清单上的所有项目。什么才是基于价值的最佳下一步？这才是对临床医生真正有用的东西，而不只是《纽约时报》喜欢展示的噱头。

Because you might not get to do everything on the potential list. What is the best next step that is, like, value based? And and that is something I think that, you know, it's just what is useful for a clinician versus what is exciting to, like, show in the New York Times.

Travis，有两个快速问题。第一，我很想听听公开证据团队追踪的指标是什么？我知道你们追踪很多指标，但有没有哪个指标能体现临床医生或用户价值，同时又能激励团队？

Travis, couple two quick questions. One, I'd love to kinda hear in the open evidence team what's, like, a metric you guys track? I'm sure you guys track a lot of metrics, but what's a metric that you guys track that reflects, like, clinician or user value and then keeps you guys motivated?

是的。我认为最重要的两个指标当然是提问数量。对吧？我们知道这些问题不是来自周五晚上泡酒吧的普通消费者。

Yeah. So I think the two ones are, of course, questions asked. Right? Like, we know that these these aren't coming from, you know, consumers that are at the bar on Friday night. Right?

这些问题是来自正在工作的医生们。对吧？所以问题越多——就像你说的，这都是自然增长。这些问题不是因为我们在推特上投广告才产生的。

These are coming from physicians who are trying to do their job. Right? So the more questions and and this isn't this is, like you said, all organic growth. These questions aren't being asked because we're doing an ad spend on Twitter or whatever it happens to be. Right?

这是因为人们确实觉得我们很有价值。这是其一。其次我们有个内部指标是问题获得点赞与未点赞的比例。当然就像客服数据一样，人们更倾向于点踩而非点赞。

This is because people find us really valuable. So that's one. And then we have an internal metric that is, you know, how often were questions liked versus not liked. And of course, just like customer support, people are more likely to put a down vote than an up vote. You know, they're more likely to say they did like something than just ignore it if they liked it.

但刚开始时我们的点赞率只有80%多，现在每次模型更新都会仔细监测确保保持在99%以上。这就是我们在宏观层面追踪的数量与质量指标。当然内部还有很多关于准确性和引用的指标，但那些都不如真实的用户故事重要。

But, you know, when we first started, we were in the eighties, and now we like our very we look very carefully anytime we change a model to make sure we stay above 99. And so that's that's, I guess, the quantity versus quality aspects that we track from broad levels. Obviously, we have a lot of internal metrics that are there are things about accuracy and citations, but those are not those are less than, like you said, the real user stories that we're getting.

太棒了。最后一个问题：你作为临床医生转型者，对想进入初创公司或风投领域的同行有什么建议？

Awesome. And then last one is that, you know, you have this clinician hat. There's a lot of clinicians that are trying to get into start up roles or venture roles. What's one piece of advice you have for clinicians that are trying to be able to do both or even transition into the venture world?

天啊，这难倒我了。让我想想...我在这方面其实也很新手，你可能有更好的见解。

Oh, man. I don't know. I'm I'm trying to think here. I I you probably have way better things. Like, this is all very new stuff to me.

对吧？直到六个月前，我基本上还是个学者，主要就是帮帮朋友的忙。虽然我也接触过几家初创公司，但我觉得谨慎选择很重要。我曾花时间与不少初创团队共事，因为他们人很好，虽然我看不出他们的价值主张在哪里，但就是想帮这群好人。我从不后悔，毕竟他们都是好人。

Right? So I was really an academic all up until, like, six months ago where I was just helping out friends more than anything else. I I but I I have kind of interacted with a couple other startups, and I think it's it is important to choose carefully. I've spent some time with a lot of startups because they were a nice set of guys, and I was like, yeah, I don't really see what the value prop is here, but I wanna help this nice group of guys. And I don't regret any of those, cause they're nice group of guys.

我喜欢帮助别人。但你知道，时间是有限的资源，对吧？而且投资初创公司是高风险的事情，对吧？

I like helping people. But, you know, time your time is a limited resource. Right? And your the the bet on a startup is is, you know, high risk. Right?

或者说，成功的几率很低。作为一个临床医生，如果你每周已经工作70小时，我认为谨慎选择确实非常重要。但我也觉得全力以赴同样关键。所以我遇到了这群人。

Or, you know, it's it's it's low chance of success. Right? I mean that so as a clinician, you know, if you're already working seventy hours a week, I think it is it is really important to choose carefully. But then I I also think it is important to go all in. And so I just I met these guys.

我当时就觉得，这是我见过最聪明的机器学习科学家。尽管两年前我们的模型还很糟糕，根本不能用——那还是发布前的阶段，就像我们正在搭建的时候。

I was like, these are the smartest ML scientists I've ever met. I even though our model, like, two and half years ago was terrible, I'm like, I can't use this for anything. This was, you know, pre release. Right? Just like when we were building.

我就觉得，这件事必须做成。这将会改变我的生活和工作方式。而且我坚信这群人绝对能做到。

I was like, this is something that needs to happen. Right? Like, this is something that needs to happen. I very much see this as, like, gonna change my life, how I practice. And I think these guys can definitely do it.

所以我全力以赴。所谓全力以赴，我并没有辞职，但所有空闲时间都在帮他们，思考如何协助，一有想法就大声说出来。我想扎克和丹尼尔确实很看重这点——毕竟这关系到他们要找谁来全职加入。

And so I went all in. Right? So that by going all in, what I meant was, like, well, I didn't quit my job. But, like, every spare minute, I was helping them out, trying to figure out how I can help them out, making sure if I had something I thought about, I, like, may I was I was loud. And I think that was I I mean, Zach and Daniel, I think, clearly valued that, right, when it came to, you know, who they would look for to kinda take on full time.

所以我从全身心投入中获益良多。就像我说的，我长期作为志愿者，并不追求即时咨询报酬，只是想为他们做些什么。

So I I benefited from, like, just going going all in, and I didn't necessarily like I said, I was a volunteer for a long period of time where I was just not necessarily looking for the immediate, like, cash consulting grab, just trying to do something for them. Yeah.

是的，我真的很喜欢这一点。我认为，组建一支顶尖团队可能是职业发展最好的加速器之一，见识成功和世界级团队的模样非常重要。而全身心投入，我认为也极为关键。虽然有时很难做到，有时难以从临床实践中抽身。

Yeah. I really love that. I think, you know, picking an a plus team is probably one of the best accelerators for your career, and seeing what success and an incredible world class team looks like is really important. And and leaning in, I think, is so important. And and it's sometimes hard to do, and it's sometimes hard to kind of go away from that clinical practice.

但如果你对此充满热情，对临床医生来说，多花时间参与风投或与初创公司合作会让他们感觉非常轻松。这让事情变得简单。

But I think if it's really passionate for you, and then I think for clinicians, it it feels so easy for them to spend that extra time doing that venture or working with that startup. It makes things easy.

没错。莱尔，我觉得这要看情况。如果你只是想要每小时赚点小钱，可以同时为五家公司提供咨询。但他们也能看出差别，对吧？

Yeah. And I think, Lyle, like I mean, I guess it depends. If you just want, like, a few bucks an hour or whatever, you can consult for five different companies. But they're they recognize difference too. Right?

如果你全身心投入，明确表示'这就是我想合作的对象'，人们会感受到你的诚意，而不是那种'我偶尔抽几小时帮忙，同时还在其他四个团队兼职'的态度。我想他们也能察觉到这种区别。最终取决于你想要从中获得什么。

If you go all in and be like, this is who I wanna work with, people see that versus, you know, I'll give you a few hours here and there and I'm doing it with four other groups. Right. I I think that they kinda recognize that too. And it all depends on, you know, what you want out of it.

特拉维斯，你能来参加节目真是太棒了。再次感谢你来到这里并分享这些精彩见解，我们非常感激。

Well, Travis, this is so awesome to have you on the show. Thank you again for being here and providing our incredible insights. We really appreciate you.

不客气，很开心。见到你真好。

No problem. Was fun. It's nice seeing you.