2026年1月回顾：神经学中的人工智能话题

它会收集一系列标记，这些标记是以特定方式组合在一起的字母，然后在时空上将它们连接起来，并通过强大的超级计算重新整合所有内容。

It takes collections of tokens, which are letters in a certain way that are bundled together and connects them in space and time and fits them all together again through massive supercomputing.

所以这是从人类层面来说的。

So it's on a human level.

比如，如果你观察变换器的工作方式，它与人类阅读的行为完全无关。

Like, if you're looking at what a transformer does, it's totally unrelated to the act of human reading.

对。

Right.

而且因为你有更丰富的经验，你已经看到过许多由各种不同程序生成的产品。

And when you've looked at these because you have far more experience, you've seen several more products that have been generated from all the different programs.

你是一位写作专家。

And you're an expert in writing.

你看到什么模式了吗？

Do you see a pattern?

你有没有觉得有些东西很平淡？

Do you get the sense that something's bland?

你现在会读这些东西，然后想，有时候你是不是能猜出来？

Do you sort of read these things now and think maybe sometimes you're able to guess?

我并不是在贬低ChatGPT。

I'm not taking anything against ChatGPT.

我的意思是，它是一个很棒的程序。

I mean, it's a wonderful program.

如果你问它一些平庸的问题，它就会给出平庸的回答。

If you ask it banal questions, it gives banal answers.

我用Pseudowrite和Cohere为美国文学期刊LitHub写了一个短篇故事，我训练了机器人模仿作家的风格，比如纳博科夫、莎士比亚、狄更斯，然后让它整合这些作家的风格写一个爱情故事。

I used Pseudowrite and Cohere to write a short story for LitHub, a literary journal in America, where I trained bots to write like writers, like Nabokov, Shakespeare, Dickens, and then I had it write a love story integrating all of these talents into them.

这花了我很多功夫。

That took a lot of work.

要实现这一点需要大量的技术工作，但最终它能产生令人惊叹的、富有创造力且生动的语言成果。

It took a lot of technical work to get to be able to do, but then it creates these incredible, creative, vibrant linguistic results.

比如，如果你让它模仿狄更斯的风格，它就会做到。

Like, if you tell it to write like Dickens, it will.

这太惊人了。

That's an amazing thing.

我还在研究一些项目，我会给出提示故事，告诉它以某种方式写作，然后用模型进行训练，这样它每次都会讲出不同的故事。

I'm also working on things where I'm writing prompt stories where you tell it to write in certain ways, and then you train it on models, and then it will tell the same story differently every time.

本质上，你创造了一个无限再生的故事。

And, essentially, you create an infinitely regenerative story.

这只是即将发生的事情的一个小小例子，这就是为什么我认为我们还没有完全意识到即将到来的事物的严重性。

This is just a small example of what's coming, which is why I think we haven't quite grasped the stakes of what's coming.

我在加拿大，所以法语沉浸式教学在这里非常重要。

I'm in Canada, so teaching French immersion is a big deal here.

比如，所有孩子都学习法语，以便将来能当上总理。

Like, all the kids learn French so that they could be prime minister.

对。

And Right.

对。

Right.

你知道，尽管成年人都最终没有说法语。

You know, although none of the adults ever end up speaking French anyways.

但我有个朋友，他的孩子不喜欢学校里提供的任何法语沉浸式读物。

But a friend of mine, like his kid didn't like any of the French immersion books that were available for him in the school.

于是他直接去使用ChatGPT，让它按孩子的年级水平写一本关于他孩子最喜欢的超级英雄的法语书。

So he just went to chat GBT and had it write a French book at grade level about his kid's favorite superhero.

它真的写出来了，孩子读了之后非常喜欢。

And it did that, and the kid read it and loved it.

所以他现在不断使用ChatGPT来生成他想要存在的那些书。

So he's now going to ChatGPT to write the exact books that he wants to exist.

我的意思是，我真的不确定自己能否想象出这会带来什么后果。

I mean, like, I I actually literally I'm not sure I can conceive of the consequences of that.

对吧？

Right?

当消费者与语言创造者之间的界限彻底模糊、融为一体时，这究竟意味着什么？

Of what it means when essentially the line between a consumer and creator of language dissolves and and becomes this enormous blur.

别提这对儿童图书行业意味着什么了。

Like, never mind what it means for the children's books in industry.

但如果你想到一本书，可以随时让它为你定制，这又意味着什么？

But what does it mean when if you think of a book, can just have it made for you?

我的意思是，我认为这就是我们正在走向的方向。

And I mean, I think that is where we're going.

我觉得我之前的一些问题可能集中在如何识别这一点，以及如何避免任何形式的欺诈，无论我们是否需要定义或重新定义它。

I think maybe some of my questions have been focusing on how do we identify this and the concerning aspects avoiding any manner of fraud regardless of how we may have to define that or change the definition.

你认为这项技术最有希望的积极方面是什么？

What do you see as the most promising positive aspect of the technology?

你之前提到过，你让一个年轻学生对学习产生了极大的兴趣。

You touched on it a little bit in that you got a young student really excited to learn.

你还认为有哪些其他有希望的方面？

What other promising aspects do you see?

哦，我觉得这非常令人兴奋。

Oh, I think it's hugely exciting.

我的意思是，我在用它进行创意创作。

I mean, I'm using it creatively.

我认为这种创意爆发将会是巨大的。

I think the creative explosion of it is gonna be immense.

我觉得自己就像是电吉他的早期采用者。

I feel like I'm an early adopter of the electric guitar.

老实说，我就是这么感觉的。

That's honestly how I feel.

我写过一些东西，比如为《大西洋》杂志写的《Facebook让我们孤独吗？》。

I've written things like I wrote, is Facebook making us lonely for the Atlantic.

我的意思是，我曾是社交媒体的早期怀疑者，但我觉得，社交媒体的灾难让我们对技术产生了极度的恐惧。

I mean, was a very early skeptic of social media, but I think in some ways, the disaster of social media has made us extremely afraid of technology.

对吧？

Right?

这几乎是我们的本能反应。

It's kind of our automatic response.

我的意思是，这怎么会摧毁社会呢？

Like, how is this going to destroy society?

我实际上认为这项技术很棒。

I actually think this technology is wonderful.

我不害怕它。

I'm not afraid of it.

尽管我有这些倾向，但我对它感到非常乐观。

Despite my tendencies, I feel quite hopeful about it.

我也认为，当你使用它时，语言并不会变得不那么重要。

I also think it's not going to be the kind of thing where language becomes less important when you use it.

它肯定不会取代任何人的工作。

It's certainly not going to replace anyone's jobs.

这是另一个担忧，但我真的认为这是一种肤浅的焦虑。

That is another worry, but I really think that's a facile anxiety.

这将类似于唱盘对嘻哈音乐的发明。

This is going to be like the invention of the turntable for hip hop.

就像这项新技术催生了所有这些新的艺术形式，它需要对以往音乐有极其系统的了解。

Like this new technology gave birth to all this new art, and it required a kind of hugely systematic knowledge of previous music.

它将要求我们真正理解语言及其运用，而我认为我们已经很久没有这种理解了。

It's going to require a real understanding of language and its manipulation that I think we kind of haven't had for a while.

我确实相信，事实核查将会像以前从未有过的那样，成为教育中不可或缺的一部分。

I actually believe that like fact checking is going to be integral to education in a way that it just wasn't before.

这不再是是否能生成新文本的问题。

Like, it's not gonna be a question of necessarily can you generate a new text?

而是你能否确定一段文本是否正确。

It's like, but can you establish that a text is actually correct?

在面对这项技术时，这种能力将具有最大的价值。

That is going to be a skill with maximum value in the face of this technology.

我的意思是，我真的相信这一点。

I mean, I I really believe that.

是的。

Yeah.

我明白你的意思。

I hear what you're saying.

在社交媒体上，这已经变得极其重要，并且会呈指数级增长。

It's already become incredibly valuable with social media, and it will just logarithmically.

是的。

Yeah.

比如说，如果我有个神经学方面的问题，我可以去问ChatGPT并得到一个答案。

I mean, for example, like if I had a neurological question, I could go to chat GPT and get an answer.

但我还是更愿意跟你聊聊。

I'm going to tell you I would rather talk to you.

你明白我的意思吗？

Do you know what I mean?

没错。

Like Right.

事实上，无论它给我什么答案，我都不会轻易相信，因为唯一能验证的方法就是我自己去查证——要么通过互联网，但互联网本身也不可靠；要么通过学术同行评审的资料，可我又看不懂那些。

In fact, whatever answer gave me, I really wouldn't believe because the only way I could do it was me to fact check it either with the Internet, which is, of course, unreliable in its own way or scholarly peer reviewed work, which I don't understand.

所以，要核实信息的真实性，获取真正可靠的信息，如今和以前一样困难。

So getting to the veracity of things, getting information that you can actually rely on is just as difficult now as before.

是的。

Yeah.

你给了我们太多值得思考和消化的内容。

You've given us so much to think about, to chew on.

我建议我们的听众阅读斯蒂芬在《大西洋》杂志上的文章。

I would encourage our listeners to read Stephen's article in The Atlantic.

文章标题是《大学申请文书已死》。

The title again is The College Essay is Dead.

这是一篇简短而精彩的阅读概述。

That's a really quick, great overview of a read.

当然，你也可以在社交媒体上关注斯蒂芬，了解更多他关于各种话题的观点。

And you can always, of course, follow Stephen across social media to hear more from him on a variety of topics.

今天能与你交谈、向你学习，真是一次令人耳目一新的愉快经历，斯蒂芬。

It has been an eye opening pleasure talking with you, Stephen, and learning from you today.

谢谢。

Thank you.

我的荣幸，斯蒂芬。

My pleasure, Stephen.

您好，我是维克森林大学的哈莉·亚历山大，今天我和尚多尔·贝尼斯基在一起。

Hello, this is Hallie Alexander from Wake Forest University, and I'm here today with Shandor Benisky.

他是丹麦奥胡斯大学的临床教授。

He's a clinical professor at Aarhus University in Denmark.

尚多尔是可穿戴设备和自动癫痫检测方面的专家。

Shandor is an expert on wearable devices and automated seizure detection.

今天我们将讨论他团队最近关于使用人工智能自动解读脑电图的研究。

And today we're discussing his group's recent work regarding automated EEG interpretation using artificial intelligence.

这项研究上个月发表在《JAMA神经学》上。

This was published in JAMA Neurology last month.

尚多尔，今天能邀请您做客播客，我感到非常荣幸。

Cheandor, it's such an honor to have you on the podcast today.

对我来说也是荣幸，海登。

Pleasure on my side, Hayden.

我想大家都很想知道，你们开发的用于解读脑电图的AI模型是什么？这个模型与过去我们见过的其他自动化脑电图检测方法有何不同？

I think everyone's really interested to know what is the AI model that you came up with for reading EEG and how does this model differ from some of the other automated EEG detections that we have seen in the past?

此前已发表的用于脑电图分析的AI模型仅关注有限的方面，例如是否存在癫痫样异常，或脑电图是否正常或异常。

So the previously published AI models for analysis of EEG addressed on limited aspects such as the presence absence of epileptiform abnormality or whether the trace is normal or abnormal.

我们实现的是对常规临床脑电图的更全面解读。

What we achieved is a more global interpretation of the routine clinical EEG.

我们的模型首先将正常和异常记录区分开来，然后将异常记录进一步细分为四种对临床决策至关重要的亚类。

So what our model does is first separates normal from abnormal recordings, and then it subclassifies the abnormal recordings into the four most important subcategories which are necessary for a clinical decision making.

这些类别包括局灶性癫痫样、全面性癫痫样、局灶性非癫痫样和全面性非癫痫样。

And these categories are focal epileptiform, generalised epileptiform, focal non epileptiform and generalised non epileptiform.

这些主要是慢波。

These are mainly slowings.

这些是最关键的分类，能够帮助临床医生做出治疗决策或决定是否需要进一步检查患者。

So these are the most important categories which can help clinicians then to make a therapeutic decision or decision on further workup of the patients.

我们算法的另一个重要创新在于，我们对其进行了全自动模式的验证。

And the other important novelty about our algorithm is that we validated it for a fully automated mode.

因此，大多数先前发布的模型都需要人为干预，通常是因为它们的特异性很低，需要人类专家确认结果。

So most of the previously published models need a human intervention, typically because their specificity is very low, so human experts need to confirm the findings.

现在，在测试数据集中，我们表明，无需任何额外的人工干预，该模型的性能与专家相当。

Now in the test data set, we showed that without any additional human intervention, the model achieves the same performance as the experts.

该模型的名称是SCORE AI，SCORE是脑电图标准化计算机辅助结构化报告的缩写。

And the name of the model is SCORE AI and SCORE is the acronym for standardized computer based organized reporting of EEG.

十年前，我们将其作为一种标准化方法，用于从脑电图中提取临床相关特征，并据此生成标准化报告。

Now we published this ten years ago as a standardized way of extracting the clinically relevant features from the EEG and then use this to generate a standardized report.

但后来我们发现，这种标准化的标注是训练AI模型的理想输入。

But then it turned out that the standardized labeling is the ideal input for training an AI model.

因此，我们将我们的模型命名为ScoreAI。

So that's why we call our model ScoreAI.

非常好。

Excellent.

所以我完全可以理解，这将填补临床护理中的许多空白。

So I can certainly see how this would fill a lot of gaps in clinical care.

但您如何看待将其融入临床实践？

But how do you envision this being incorporated into clinical practice?

嗯，这取决于临床实践的地点。

Well, this depends on where the clinical practice is.

在资源有限的地区，基本上没有脑电图专家。

So in the resource limited areas, basically there are no EEG experts.

而在有专家的地区，SCORE AI 可以减轻工作负担。

Now in areas where there are experts, Score AI could ease the workload.

因此，该模型还会在脑电图记录中突出显示脑电图异常。

So EEG abnormalities are also highlighted by the model in the EEG recording.

因此，专家们首先可能根本不需要打开被算法判定为正常的记录。

So the experts, first of all, would perhaps not need to even to open a recording which is interpreted as normal by the algorithm.

然后，如果模型发现异常，它还会在波形中突出显示这些异常。

And then if the model finds abnormalities, then the model also highlights the abnormalities within the trace.

因此，专家可以进去检查模型的标记。

So the experts could go in and then check the markings of the model.

是的，这太令人兴奋了，因为到目前为止，我们还没有任何类似的东西能真正做到这一点来减轻我们的工作负担。

Yeah, that's so exciting because as of yet, we don't have anything that is like that yet that can really do that to ease our workload.

你们现在在临床实践中使用它了吗？

Is it something that you are using now in your practice?

还没有，因为我们正在等待监管机构的批准，但一旦获得批准，我会非常非常乐意将其纳入我们的临床流程。

Not yet because we are waiting for the approval from the regulatory bodies, but as soon as we will have the approvals, I would be very, very happy to put it into our clinical pipeline.

也许我可以提一下，它有望在明年年初登陆Natus Neuroworks平台，但需在获得监管批准之后。

And perhaps I may mention that it will be available on Natus Neuroworks hopefully early next year, but after the regulatory approval.

我也想感谢我的合作者Jesper Suede和Harald Orlian，他们对AI模型的开发做出了最大贡献。

And I would also like to give the credit to my collaborators, Jesper Suede and Harald Orlian, who contributed most to the development of the AI model.

是的。

Yes.

另外，我想像在阅读你们的文章时，有时会遇到一些难以判断的脑电图，两种解读都有可能。

And another way, I guess, I envisioned using it when I was reading your article was just sometimes you get a challenging EEG that could be called either way.

找同事当裁判挺好的，但你知道，大家都很忙。

And it's nice to ask a colleague like a tiebreaker, but, you know, everyone's busy.

所以你有没有想过以这种方式使用它？

So do you ever envision using it in that way?

比如当你处于模糊地带，犹豫该不该判定为是或否时，可以借助软件来帮忙判断？

Like when you're in a gray area wondering if you should call something, yes or no, and then you could use the software to weigh in?

这确实是一个非常好的观点。

This is indeed an excellent point.

我们论文中的一个发现实际上表明，你可以用它来实现这个目的，因为我们发现AI模型的评分比单个专家的平均评分更接近11位专家的共识平均值。

And one of our findings in the paper actually could indicate that you could use this for that purpose because we showed that the scorings of the AI model are closer to the average consensus of 11 experts than the average of the individual experts.

因此，通过Score AI自动实现，你可以获得一个稳健的专家多数意见。

So you could get a robust majority expert second opinion by doing this automatically with Score AI.

是的。

Yeah.

随着这项技术的进步，每个人，尤其是所有解读脑电图的人，都会想知道：当我们可以开发出如此先进、几乎无需人类参与的AI时，人类专家的角色将会如何变化？

So as this technology improves, then the question on everyone's mind, or at least everyone that interprets EEGs is what is going to be the role of the human expert as we develop more advanced AI like this that can really operate without us?

这个问题我经常被问到。

Oh, get that question often.

我们是不是该停止培训临床神经生理学家了？AI会取代我们吗？

Shall we stop training clinical neurophysiologists or will AI replace us?

我听过一个非常好的回答，我想把这份功劳归于卡蒂琳，我完全同意她的观点。

I've heard a very good answer and I would like to give the credit to Katyalin who told me that I fully agree with her.

她说，AI不会取代我们，但会使用AI的人会取代我们。

So she said that AI will not replace us, but humans using AI will.

如果我们忽视技术的进步，忽视AI所蕴含的可能性和潜力，那么那些善于利用AI作为高效工具的聪明人，很可能会取代我们。

So if we ignore the advances in technology and we ignore the possibilities, the potentials which are in the AI, then the smart humans who will take this as a tool, as a very efficient tool, will probably replace us.

再次强调，我们应该把这看作一个既有优势、也有局限的工具。

Now, again, we should think of this as a tool which has advantages and of course limitations.

从长远来看，我确信它将帮助我们发现人类目前还无法察觉的脑电图异常。

And on the long run, I'm sure that it would help us detect abnormalities in the EEG which we humans cannot see yet.

它还能帮助我们解读更多的脑电图数据。

And it also would help us interpret a higher amount of EEGs.

我只是想澄清一下，你们开发的技术目前仅应用于常规脑电图记录，尚未包括持续脑电图和重症监护病房的脑电图。

I just want to clarify that the technology that you've developed was applied to routine EEG recordings specifically continuous EEG, ICU EEG were not included in this yet.

但我很好奇，是否有计划将这项技术扩展到这些类型的脑电图研究中？

But I'm wondering, there plans for developing technology to apply to those types of EEG studies as well?

是的，当然有。

Yes, definitely.

我们已经为长期监测开发了适配版本，并正在测试中。

So we already have an adaptation for long term monitoring and we are currently testing it.

我们目前正处于长期监测版本的验证阶段。

We are in the validation phase of the LTM version.

接下来的步骤将是训练该技术用于新生儿脑电图和重症监护病房的脑电图。

And then the next steps will be training it also for neonatal EEG and also for the ICU.

所以你提到你们与一些合作者合作了，当然，这样一个项目似乎需要大量的不同意见输入。

So you mentioned that you worked with some collaborators, of course, and a project like this seems like it would take lots of different input.

所以还有谁参与了

So who else was kind

在

of on the

帮助实现这一成果的团队中还有谁？

team that helped to bring this to fruition?

除了学术参与者外，对模型开发起到决定性影响的公司是Holberg EG。

Besides the academic participants, decisive influence and the company who developed the model was Holberg EG.

这是一家位于挪威卑尔根的公司。

They are a company based in Bergen in Norway.

我想再回到你提到的一点，即这项技术甚至可能被开发用于检测人类无法察觉的事物。

And then I want to circle back to something you said about maybe it could even be developed to detect things that humans cannot detect.

所以我很想知道这将如何实现，因为我对人工智能的理解非常浅薄，只知道你可以根据输入数据对其进行训练。

So just curious how that would work because my understanding of AI, which is rudimentary at best, is that you can kind of train it, based on inputs.

一旦训练完成，它就能独立地检测事物。

And then once it's trained, then it can go on to detect things independently.

那么，你们如何训练它来检测我们目前可能都无法察觉的事物呢？

So how do you train it to detect things that we maybe can't detect already?

这又是怎么做到的呢？

How does that work?

实际上，我们在Score AI的训练阶段就实现了这一点，因为我们有一个想法，就是给AI模型更多的空间和自由度。

So actually we built that also in the training phase of Score AI because we had this idea of giving more space, more freedom for the AI model.

之前的尝试严格限制了模型应该寻找的脑电图异常痕迹、异常类型或异常点。

The previous attempts really limited what trace or what kind of abnormality or what point on the EEG abnormality the model should find.

例如，它们只限定模型寻找尖峰的峰值。

For example, they restricted it to find the peak of a spike.

但我们认为，脑电图中包含的信息远不止尖峰的峰值。

But we thought that there is much more information in the EEG than just the peak of the spike.

因此，我们训练模型在16秒的脑电片段内寻找异常。

So we trained the model to find abnormalities within sixteen second epochs.

我们并没有将它限定在尖峰峰值出现的那一刻。

So we didn't restrict it to that point in time where the peak of the spike was.

我们的初步结果（尚未发表）表明，该模型使用的特征与人类完全不同。

And our preliminary still unpublished results suggest that the model uses completely different features than we humans.

所以下一个令人兴奋的步骤将是进行某种逆向工程，看看模型究竟发现了什么。

So the next exciting step will be to do a kind of reverse engineering and then see what exactly the model finds.

因为模型确实能检测到尖峰，但它是基于与我们人类专家所用的不同特征。

Because the model indeed finds the spikes, but based on different features than what we human experts use for it.

这被称为可解释人工智能（XAI）。

And this is called x AI, this explainable AI.

这是我们接下来的项目之一：对我们开发的模型进行逆向工程。

And this is one of our next projects to do reverse engineering of the model we developed.

现在我们已经在讨论AI甚至可能检测出人类无法发现的东西。

Now we're already talking about the AI maybe even detecting things that humans can't.

因此，我们进入了这样一个领域——我相信你以前也听说过，关于科幻中可能出错或失控的各种情况。

So we're getting into the realm where I'm sure you've heard before of thinking about science fiction and all the things that can kind of go wrong or how it can maybe get out of hand.

本着一个忠实的科幻迷的精神，我必须问一下：对于使用这种AI进行脑电图解读，您对用户有什么关于其局限性或风险的忠告吗？

So in the spirit of just being a dutiful sci fi fan, I need to ask, any words of caution for users about the limitations or risks of using this AI for EEG interpretation?

当然有。

Oh, absolutely.

首先，一般来说，要确保你使用的是经过验证的AI模型，并且仅将该模型用于其经过验证的用途。

So the first general thing would be that make sure that you use a validated AI model and then use the model exactly for that thing that it was validated for.

在查阅验证研究时，要确保该AI模型具有泛化能力，即它没有在相同数据集上进行过拟合、训练和测试。

And then when you check the validation studies, make sure that the AI model is generalizable so it was not overfit or trained and tested in the same data set.

考虑到这些限制，我认为AI为我们带来了许多新的可能性。

Now, with these limitations in mind, I think AI offers us lots of new possibilities.

因此，我们不应对使用这些新工具感到焦虑。

So we should not have an anxiety using these new tools.

我认为这是一项令人兴奋的新进展。

I think this is rather an exciting new development.

我同意，确实如此。

I agree, definitely.

非常令人兴奋，我迫不及待想看到它明年在临床中应用。

Very exciting and I can't wait to see it coming out in clinical use next year.

因此，今天我与尚多尔·贝尼茨基讨论了他团队近期发表在《JAMA神经学》上的论文《使用人工智能自动解释临床脑电图》。

So again, I've been speaking today with Shandor Benitsky about his group's recent publication in JAMA Neurology titled Automated Interpretation of Clinical Electroencephalograms Using Artificial Intelligence.

您可以在2023年8月刊的《JAMA神经病学》上找到全文，或通过节目说明中的链接获取。

You can find the full article in the August 2023 issue of JAMA Neurology or by the link in the show notes.

Shandor，非常感谢您今天参与我们的节目。

Shandor, thank you so much for being with us today.

这也是我的荣幸。

Pleasure on my side.

感谢您的邀请，Haley。

Thank you for the invitation, Haley.

你好。

Hi.

我是来自弗吉尼亚州里士满弗吉尼亚联邦大学的Trey Bateman。

This is Trey Bateman from Virginia Commonwealth University in Richmond, Virginia.

今天，我很荣幸采访来自明尼苏达州罗切斯特梅奥诊所的行为神经学家Dave Jones，他同时也是神经科人工智能项目主任。

Today, I have the pleasure of interviewing Dave Jones, who's a behavioral neurologist at the Mayo Clinic in Rochester, Minnesota, as well as the director of the Department of Neurology Artificial Intelligence Program.

Dave是2025年6月27日在线发表于《神经病学》的一篇论文的资深作者，论文标题为《基于FDG PET的机器学习框架以支持阿尔茨海默病及相关疾病的神经学决策》。

Dave is a senior author on a paper published online in neurology on 06/27/2025 titled An FDG PET Based Machine Learning Framework to support neurologic decision making in Alzheimer's disease and related disorders.

戴夫，正如你所知，我非常推崇FDG PET成像在认知评估中的强大作用。

Dave, as you know, I'm a huge fan of the power of FDG PET imaging in cognitive evaluation.

因此，我非常期待今天能和你讨论这篇论文以及State Viewer。

So I'm really excited to talk to you today about this paper and State Viewer.

感谢你今天加入我们。

Thanks for joining me today.

谢谢，特雷。

Thanks, Trey.

很高兴能来到这里。

It's great to be here.

这项工作主要围绕State Viewer展开，这是一个临床决策支持系统框架，有助于解读FDG PET扫描图像，以支持对影响认知和行为的神经退行性疾病的鉴别诊断。

So this work really centers around State Viewer, a clinical decision support system framework that helps with the interpretation of FDG PET scans in the evaluation and interpretation of those scans in the service of differential diagnosis of neurodegenerative diseases that affect cognition and behavior.

我认为大多数人对FDG PET扫描都不陌生，但过去在培训期间，很多人可能听过类似的说法：它主要用于区分阿尔茨海默病和额颞叶痴呆，大致如此。

I think most will be familiar with what FDG PET scans are, but many might have been told something like, well, it's really only useful to differentiate between Alzheimer's disease and frontotemporal dementia, something along those lines, in the past during their training.

在进入论文之前，你能为我们听众简要说明一下，为什么这种说法可能并不完全正确？以及开发State Viewer这样的临床决策支持系统的初衷是什么？

Before we jump into the paper, can you give our listeners a brief overview of why that might not be entirely correct and what the impetus was for developing a clinical decision support system like State Viewer?

我认为，FDG PET 常被局限地视为仅用于区分阿尔茨海默病和额颞叶痴呆，因为额颞叶痴呆患者的大脑前部通常受累更严重，而阿尔茨海默病则相反。

I think FDG PET is often pigeonholed as just a way to tell Alzheimer's disease from FTD because of the fact that the brain is usually more affected in the front versus the back in FTD and vice versa in AD.

但这实际上低估了它的价值，也误解了与患者症状相关的功能神经解剖学的细微差别。

But that really undersells its value and mischaracterizes the nuance of functional neuroanatomy that's relevant to the symptoms our patients are experiencing.

FDG PET 实际上是脑网络功能的一种反映。

FDG PET really is a readout of brain network function.

它捕捉的是认知的底层生理机制，而不仅仅是蛋白质。

It captures the underlying physiology of cognition, not just proteins.

因此，当正确解读时，它可以区分更广泛的神经退行性综合征，因为每种综合征都会以独特的方式破坏脑功能和脑网络。

So when interpreted correctly, it can differentiate a much broader range of neurodegenerative syndromes because each one disrupts brain function and brain networks in distinct ways.

它在神经科核心实践中的定位诊断中起着关键作用，而患者前来就诊的知觉、认知、情感和行为症状往往使定位变得非常复杂。

It's a crucial aid in the core neurologic practice of localization, which can be very complex for perceptual, cognitive, emotional, and behavioral symptoms that our patients are coming to us, for help with.

问题是，解读这些定位模式需要高水平的专业知识，以及对数千张图像的接触，再加上详细的临床评估和对功能神经解剖学的深入了解。

The problem is that interpreting those localizing patterns requires a high level of expertise and exposure to thousands of images coupled with detailed clinical assessments and knowledge of functional neuroanatomy.

即便如此，专家们在识别数千次扫描中这些分布式模式的细微差异时，其一致性仍然有限。

And even then, experts are limited in how consistently they can recognize those subtle variations in distributed patterns for thousands of scans.

因此，State Bureau 正是为填补这一空白而诞生的。

And State Bureau was really born from that gap.

我们一直将 FDG PET 用于常规临床护理，以支持对患者的复杂决策，但我们意识到，借助人工智能，我们可以构建一个工具，捕捉这些模式的全部复杂性，标准化解读，并为临床医生提供定量证据，说明扫描结果最接近哪种综合征，即使临床医生在自己的实践中尚未接触过所有这些模式。

We had been using FDG PET as part of routine clinical care to support complex decision making for our patients, But we realized that with AI, we could build a tool that captures the full complexity of those patterns, standardizes the interpretations, and gives clinicians quantitative evidence for what syndromes a scan most resembles, even if the clinician had not yet been exposed to all those patterns in their own practice.

我个人的经历源于多年与那些无法被简单归类的患者共事，并看到即使是经验丰富的临床医生也会产生分歧。

Personally, it came from years of working with patients who didn't fit neatly into categories and seeing how even experienced clinicians could disagree.

因此，我希望打造一个可扩展、透明、客观的工具，真正帮助应对这些现实中的灰色地带病例。

So I wanted to build something that was scalable, transparent, objective, and actually help with those real world gray zone cases.

这是我个人在临床实践中非常喜爱使用的工具，它能帮助我提供最好的诊疗。

It's something I love using in my own practice to ride the best care that I can.

缺乏专业知识的情况下解读扫描结果，有时会导致临床实践中出现非常混乱的结果。

The interpretation of scans without that expertise can sometimes lead to really confusing results in clinical practice.

所以听起来，这正是试图解决这一问题的尝试，为我们提供更有用的解读，从而更精确地揭示我们所见的潜在疾病和综合征。

So it sounds like this is something that starts to try to help get around that problem and give us some more useful reads that may give more precise information about the underlying diseases and syndromes we're seeing.

现在让我们深入探讨一下这篇论文。

Let's dive into the paper a bit now.

因此，我理解这项研究中，你们的研究参与者来自多个来源，包括梅奥诊所衰老研究、阿尔茨海默病研究中心，以及梅奥诊所的神经退行性疾病研究小组，还有神经科门诊的患者。

So my understanding of this work is that you drew research participants from several sources, The Mayo Clinic Study of Aging, the Alzheimer's Disease Research Center, and the Neurodegenerative Research Group there at Mayo, as well as patients seen in the neurology clinic.

在这些群体中，你们既有具有明确临床综合征或表型的患者，也有从门诊随机挑选但未调取其健康记录数据的患者。

Within each of those groups, you had patients who had clear clinical syndromes or phenotypes, as well as those randomly chosen from the clinic, but without retrieving their health record data.

其中一部分患者拥有神经病理学信息，可作为潜在病因的金标准进行对比。

A subset had neuropathological information to compare against as the gold standard of underlying etiology.

重要的是，每位患者都必须接受过FDG PET扫描，因为这是State Viewer的主要输入数据。

And importantly, each of these patients had to have an FDG PET scan since that's the main input for State Viewer.

经过质量控制并剔除扫描质量不足的病例后，你们最终获得了三千六百多名患者。

And you ended up with just over three thousand six hundred patients after quality control and everything, got rid of some that didn't have adequate scans.

接下来的部分非常新颖，也是这项工作的核心引擎，即用于分析这些FDG PET图像的机器学习模型。

Next part is really novel and what's the engine behind this work, and that's the machine learning model used to analyze those FDG PET images.

我认为机器学习可能让人难以理解。

I think machine learning can be intimidating to try and understand.

你能为我们简单解释一下这个模型是做什么的，以及它在临床中是如何为你服务的吗？

Can you give us a simple explanation for what this model does and how this works for you in clinic?

输出结果会是什么样子？它如何帮助您为患者做出更准确的诊断？

What would the output look like and how does it help you arrive at better diagnoses for patients?

这个模型只需要一张FDG PET扫描图像。

All this model needs is an FDG PET scan.

这是唯一的输入。

That's the only input.

它不需要任何其他数据。

It doesn't require any other piece of data.

然后，该模型将这位患者的扫描图像与数千张已知诊断和预后的特征明确的标注扫描图像进行比较。

And then the model compares this particular patient scan to thousands of other well characterized labeled scans with known diagnoses and outcomes.

它通过计算机处理FDG PET图像，找出最相似的病例。

It finds ones that are most similar using some computer processing of the FDG PET.

然后，它为每种疾病返回一个相似性得分。

And then it returns a similarity score for each disease.

我有时会告诉人们，它在我们的数据库中找到了一例与您大脑相似的病例。

I sometimes tell people it finds a brain like yours in our database.

因此，我们知道过去是否见过症状或大脑模式与您相似的患者。

So we know if we've seen patients with similar symptoms or brain patterns to yours in the past.

所以有时我们会把这种算法称为‘像我这样的大脑’算法。

So sometimes we'll all call this a brain like mine algorithm.

例如，它可能会告诉您，这个扫描结果与被标记为后皮质萎缩的扫描有85%的相似性，而与DLB的相似性约为10%。

For example, it might tell you this scan looks eighty five percent similar to scans labeled as posterior cortical atrophy and maybe ten percent similar to DLB.

这有助于重新定义我们对诊断的思考方式——不是作为二元标签，而是作为疾病谱系内的相似性。

And that helps reframe how we think about diagnoses, not as binary labels, but as similarity within a disease spectrum.

在临床实践中，我会在患者病历中调出这个结果。

Clinically, I pull it up in the patient chart.

它会显示代谢变化的热力图、相似疾病类别的排名列表，甚至允许您直观地浏览参考数据库中最相似的病例，以便判断这些病例是否真的与您的情况相似。

It shows a heat map of metabolic changes, a ranked list of similar disease classes, and even lets you visually explore the most similar cases in the reference database so you can see if those cases really look like yours.

它并不是一个黑箱，因为它提供了高度个体化的高质量数据，帮助临床医生真正围绕当前患者进行思考。

It's not a black box because it gives you high quality data individualized to the current patient to help the clinician actually think about the current patient.

当您将这个模型应用于一组具有多种潜在诊断的患者时，它在分类上的表现如何？

When you apply this model to a group with a bunch of different potential diagnoses, how well did it do in classifying those?

它在这些诊断中的表现是否都一样好，还是有些诊断对模型来说更困难一些？

And did it do equally well across those, or were there some diagnoses that were more difficult than others for the model?

这正是其中一个令人惊叹的地方。

That's one of the amazing things.

我们研究了九种不同的表型，FDG PET中确实包含了这些不同临床疾病的独特信息，而且模型在各个方面都表现得非常好。

We looked at nine different phenotypes, and there's just distinct information about all these different clinical disorders in the FDG PET, and it did really well across the board.

这包括阿尔茨海默病的典型和非典型表现、额颞叶变性谱系以及DLB。

This included typical and atypical presentations of Alzheimer's disease, the FTD spectrum, and DLB.

我们还能够证明，这些预测与现有的病理数据高度一致，这一点非常重要。

And we're also able to show that these predictions aligned well with available pathology data, which is important.

因此，它不仅仅是重复临床直觉，而是真正追踪了这些疾病的核心生物学多维度指标。

So it's not just replicating clinical navels, it's really tracking with multiple measures of core biology to these diseases.

对于像后皮质萎缩、DLB、语义变异型和原发性进行性失语这类更明确的综合征，模型的表现最为出色。

And the performance was strongest for more distinct syndromes like posterior cortical atrophy, DLB, semantic variant, primary progressive aphasia.

事实上，令人惊讶的是，它能够准确区分所有主要的PPA亚型：语义型、非流利型和词元型，而仅凭临床语言特征有时很难将它们区分开来。

In fact, pretty amazingly, it separates all the major PPA classes: semantic, nonfluent, and logopenic, which can sometimes be difficult to tease apart just based on the clinical language features.

该模型在面对具有不同潜在病理机制的临床综合征时，表现稍差一些。

The model had a little bit more trouble with clinical syndromes that have very different underlying pathologies.

例如皮质基底节综合征，它有时可能由潜在的阿尔茨海默病引起，但也可能由皮质基底节变性引起。

Things like corticobasal syndrome, which oftentimes can be due to underlying Alzheimer's disease, but may also be due to underlying corticobasal degeneration.

然而，当你看到数据的呈现方式时，大多数临床医生都能理解并自行得出正确的结论。

However, when you see the way the data is presented, most clinicians will understand this and come to the right conclusions for themselves.

因此，你们还在论文的放射科阅片研究部分对模型进行了测试，其中我们经常遇到的一个临床难题是根据扫描表现区分后皮质萎缩与路易体痴呆。

So you also put the model to a test in the radiologic reader study portion of the paper where a really common clinical problem that we see a lot is differentiating between posterior cortical atrophy and dementia with Lewy bodies based on the scan presentation.

你可以看到，这两种疾病有一些重叠的症状，因此在临床上区分它们可能很困难。

And you can see there's some overlapping symptoms, and so this can be a difficult clinical distinction to make.

两者在FDG-PET上都表现出明显的枕叶皮层后部改变。

Both have prominent posterior changes on FDG PET in the occipital cortex.

我本人曾多次看到报告中对某病例更符合后皮质萎缩还是路易体痴呆存在分歧。

And I've personally seen a lot of disagreement in reports on whether something is more consistent with posterior cortical atrophy or dementia with Lewy bodies.

你们将一种包含定量皮层表面投影的常规工作流程，与仅使用State Viewer生成报告的工作流程进行了比较。

You compared a pretty typical workflow that included quantitative cortical surface projections against a workflow that only included reports generated by State Viewer.

State Viewer 表现如何？

How did State Viewer do?

我们是不是都要失业了？

Are we all getting ready to be out of a job yet?

我认为在可预见的未来，我们所有人都还会继续工作，但这项研究中最令人满意的部分之一就是验证了这个工具是否真的能帮助临床医生解决当前的实际问题，尤其是像区分路易体痴呆与后皮质萎缩这样极具挑战性的视觉读片任务。

So I think we'll still all be in a job for a long time to come, but this really was one of the most satisfying parts of the study because it tested whether the tool can actually help real clinicians with a current problem, which is especially hard visual reading tasks like distinguishing dementias Lewy bodies from posterior cortical atrophy.

它们看起来可能非常相似。

They can look pretty similar.

因此，我们给放射科医生提供了两种工作流程：一种是传统的皮层投影，另一种仅使用 State Viewer 的报告。

So we gave radiologists two workflows, one using a traditional cortical projections and another using just state viewer reports.

我们发现，State Viewer 显著提高了做出正确诊断的可能性，尤其是在经验较少的阅读者中。

And what we found is that state bureau significantly increased the odds of reaching the correct diagnosis, especially among less experienced readers.

平均而言，它使正确诊断的几率提高了约3.3倍，并且阅读速度提升了约50%。

On average, it increased the odds of the correct read by about 3.3 times, and they could do it about 50% faster.

因此，它并没有真正取代专家的判断，而是增强了判断，为读者提供了一个一致的定量分析框架，并在从新手到世界级专家的不同水平之间实现了公平性。

So, it doesn't really replace expert judgment, but it did augment it, giving readers a consistent quantitative framework, and it leveled the playing field across a range of expertise from novice to kind of world expert level.

表现最好的实际上是我们的经验最丰富的阅读者，配合StateViewer使用，其表现甚至优于模型本身。

And the best performer was actually our most experienced reader in conjunction with StateViewer, which did even better than the model itself.

所以很明显，我们离失业还差得远，但希望借助这类工具，我们能做得越来越好。

So it's pretty clear we're not out of a job by any stretch of the imagination, but hopefully we're getting better at it with these types of tools.

如果我没记错的话，这种优势在经验最少的人身上最为明显，也就是那些新手放射科医生。

If I'm remembering correctly, the benefit also seems to be largest for those with the least amount of experience, so the more novice radiologists.

我记的是对的吗？

Am I remembering that correctly?

完全正确。

That's absolutely right.

处于培训阶段的人还没有太多读片经验，但给他们这个工具后，他们的表现可以超过专家水平。

So people at trainee level haven't had a lot of experience reading scans yet, but giving them the tool, can perform better than expert level.

这似乎是一项非常了不起的成就，我认为这能够真正传递这种专家诊断信息，甚至在一定程度上实现民主化。

That seems like a really impressive feat and something that I think could really bring this type of expert diagnostic information and maybe democratize it a little bit.

这看起来确实是一项非常强大的成果。

That seems like something that would be really powerful to see.

在我们剩下的时间里，你能为我总结一下你在这里工作的关键要点吗？

In the remaining seconds that we have here, could you summarize for me the key take home message that you want to remember about your work here?

我认为关键在于State Bureau为FDG PET解读带来了定量、客观、一致且可扩展的分析能力。

I think the key here is that State Bureau brings a quantitative objective consistency and scale to FDG PET interpretation.

这并不是要取代专家的直觉，而是通过透明、数据驱动的洞察来增强它。

It's not about replacing expert intuition, it's about enhancing it with transparent, data driven insights.

在治疗决策、预后评估和患者咨询越来越依赖准确诊断的时代，这类工具对于确保患者获得最准确的诊断，最终获得最佳护理至关重要。

In an era where treatment decisions, prognostication, and patient counseling increasingly depend on getting the diagnosis right, tools like this are essential to ensure patients receive the most accurate diagnosis and ultimately the best possible care.

谢谢你，Dave，与我们分享你的研究成果。

Thank you, Dave, for sharing your findings with us.

我知道我非常希望看到State Viewer被更广泛的受众使用，包括我自己的临床工作流程，这似乎是我们迈向这一目标的重要一步。

I know I'd love to see State Viewer in use to a wider audience, including my own clinic workflow, and it seems like this is a really important step towards that eventuality.

我们开发State Viewer的初衷，是为了让诊断过程更加精准和清晰，不仅适用于学术中心，也适用于日常临床实践。

We built State Viewer to bring more precision and clarity to the diagnostic process, not just for academic centers, but for everyday clinical practice.

我们的希望是，通过让这类先进工具更易获取，我们可以提高诊断准确性、减少不确定性，最终帮助更多患者更早、更有信心地获得所需的治疗。

Our hope is that by making advanced tools like this more accessible, we can improve diagnostic accuracy, reduce uncertainty, and ultimately help more patients get the care they need earlier and with greater confidence.

我刚刚与戴夫·琼斯进行了交谈，他是论文《基于FDG PET的机器学习框架以支持阿尔茨海默病及相关疾病的神经学决策》的通讯作者，我鼓励大家在《神经病学》期刊上阅读这篇论文。

I've been speaking with Dave Jones, senior author on the paper titled An FDG PET Based Machine Learning Framework to Support Neurologic Decision Making in Alzheimer's Disease and Related Disorders, which I would encourage everyone to read in neurology.

大家好。

Hello, everyone.

我是弗吉尼亚大学的安迪·萨瑟兰，今天非常高兴能与波士顿贝丝·以色列迪肯斯医学中心内科的亚当·罗德曼对话。

This is Andy Sutherland from the University of Virginia, and I'm very excited today to be speaking with Adam Rodman from the Department of Internal Medicine at Beth Israel Dickenas Medical Center in Boston.

本周的神经病学播客，我们将讨论一篇最近作为预印本发表的非常令人兴奋的文章，标题为《大型语言模型在医生推理任务中的超人表现》。

For this week's neurology podcast, we're gonna be discussing, I think, in a very exciting article that was recently published as a preprint, titled Superhuman Performance of a Large Language Model on the Reasoning Tasks of a physician.

在这项研究中，罗德曼医生及其同事评估了由OpenAI于2024年9月推出的O1模型，这是一个思维链大型语言模型。

In the study, doctor Rodman and colleagues evaluated the o one model, which is a chain of thought large language model introduced by OpenAI in September '24.

我们将在播客中进一步讨论这一点。

We're gonna talk about this a little bit more in the podcast.

他们评估了我们所有医生都熟悉的复杂临床诊断推理能力，例如鉴别诊断、管理推理、概率推理等。

And they evaluated in the performance of complex clinical diagnostic reasoning skills familiar to all of us, such as differential diagnosis, management reasoning, probabilistic reasoning, etcetera.

他们将这一结果与之前其他大型语言模型的实验进行了比较，这些模型可能对我们的听众更熟悉，例如基于GPT-4的模型如ChatGPT，甚至包括一些历史人类对照组。

And they compared this to prior experiments with other large language models that will be perhaps more familiar to our listeners, such as GPT-four based models like ChatGPT and even some historical human control.

所以，我相信到目前为止，这个富有争议的引言已经成功引起了我们听众的兴趣。

So I'm sure at this point, this provocative introduction has hopefully got our listeners' interest piqued.

那么，我们开始吧。

So let's begin.

亚当，感谢你参加我们的播客。

Adam, thanks for joining us for the podcast.

当然。

Of course.

你忘了提醒一下，我不是神经科医生，我只是内科医生。

You forgot the caveat that I'm not a neurologist, and I'm just an internal medicine doctor.

嗯，你知道，我们欢迎各种背景的专家，包括像你这样的内科同行参与我们的神经科播客，亚当。当然，我们都使用相同的临床医学语言。

Well, you know, we accept all kinds, including our internal medicine brethren here on the neurology podcast, Adams, which, of course, you know, we speak the same language of clinical medicine.

我认为这种语言正是我们非常感兴趣的，尤其是在这些新技术的背景下。

And I think that's a language that we're really interested in, particularly as it relates to these new technologies.

所以，我觉得有一件事总是很好的。

So, I think one thing, it's always good.

我们之前在《神经病学杂志》上做过一些关于大型语言模型和不同类型人工智能的播客。

We've done some podcasts talking about large language models and different types of AI before here on The Neurology Journal.

但如果你能的话，亚当，能否简要介绍一下你感兴趣的这种思维链模型（Chain-of-Thought Model），以及它如何扩展了像ChatGPT这样的更熟悉的先前模型？

But if you could, Adam, just as a brief introduction, square us up to a short definition of this particular model you're interested, a chain of thought model o one, and how it expands on some of those more familiar prior models like chat GPT, for instance.

当然可以。

Absolutely.

当然，这是我们两个专业共同的特点。

And, course, that's something that our specialties have in common.

我们都非常关注我们的思维方式。

We both care a lot about how we think.

正如我们所知，大型语言模型本质上是语言预测工具。

So as we know, large language models are fundamentally language prediction tools.

大型语言模型本质上只是预测句子、段落或书籍中的下一个词或标记。

Large language models are fundamentally just predicting the next word or token in a sentence, in a paragraph, in a book.

大型语言模型这些基础模型真正酷的地方在于，它们从这种机制中涌现出推理能力——特别是诊断领域中的所谓‘推理’特性。

Now, what's really cool about large language models, these kind of base models, is that they have emergent reasoning or the air quotes, reasoning properties from this, especially say in the field of of diagnosis.

这其中一部分原因可能是，你可以想想人类思维是如何运作的，以及神经科医生和内科医生如何将词语或语义限定词归类，这与大型语言模型所做的工作类似。

And some of this probably is because of, you know, like, you think about how the human mind works and how we understand that neurologists and internists kind of group words or semantic qualifiers together, that's similar to what a large language model is doing.

因此，早期的研究表明，语言模型在鉴别诊断方面表现得非常出色。

So the early research showed that language models are remarkably good at differential diagnosis.

事实上，它们比以往任何技术都更优秀，因为显然在人工智能与医学的历史上，我们自20世纪50年代以来就一直在研究这个问题。

In fact, they were better than any technology that had been built before because obviously like in the history of AI and medicine, we've been working on this since the 1950s.

所以这并不是一个全新的想法。

So it's not exactly a new idea.

随后我们发现，语言模型还能做许多其他对医学有帮助的事情，比如完成医学任务。

And then we found, you know, language models can do a lot of other things that are helpful, like tasks in medicine.

它们正被用于环境录音。

They're being used for ambient listening.

在波士顿的一些医院里，它们已经被用于这项工作。

That's something they're being used up here in in some of our hospitals in Boston.

它们在我们与患者交谈时进行监听，并自动生成病历。

So listening to us while we talk to our patients, writing the notes.

它们可以用于沟通。

They can be used for communication.

一些早期结果表明它们具有富有同情心的沟通能力，但它们在很多方面并不擅长。

There's some early results that said they had compassionate communication, but there's also a lot of things that they can't do particularly well.

而这正是大量推理所在。

And that's a lot of the reasoning.

因此，许多更高级的推理任务。

So a lot of the more advanced reasoning tasks.

那么，为什么它们会呢？

And again, why would they?

它们本质上是词语预测工具。

They're fundamentally word prediction tools.

研究人员——这里的‘我们’是泛指——发现，早在三年前的一篇论文中，尤其是在医学领域大约一年半前，我们可以使用一种叫做‘思维链’的方法来提升模型性能。

What researchers, this is the royal we, what we discovered about, you know, the paper came out three years ago, but especially in medicine about a year and a half ago, is that we could use something called chain of thought to improve model performance.

我会解释什么是思维链。

And I'll explain what chain of thought is.

这有点疯狂，但它确实有效：如果你让语言模型列出它思考问题的各个步骤，并要求它展示推理过程，它的推理能力就会提升。

It's kind of crazy that it works like this, but it turns out that if you get a language model to list out the way that it thinks through a sorts of steps and you tell it to show its work, it reasons better.

其中一些方法简直荒谬。

And some of these are absurd.

比如，我说：‘哦，安迪，这太疯狂了。’

Like, I oh, Andy, this is crazy.

比如，如果你告诉它‘深呼吸一下’，它在数学题上的表现就会比你不这么说时更好。

Like, if you tell it to take a deep breath, it does better on math problems than if you don't.

如果你告诉它你的工作岌岌可危，它的表现就会更好。

If you tell it your job is on the line, it does better.

如果你告诉它：‘这道题如果解不出来，它就会彻底消失’，它的表现还会更好。

And if you tell it this one's really disturbing, that it will cease to exist if it fails to solve this math problem, it will do even better.

我们对此进行了测试。

We tested this.

就在几周前，我确实和我的住院医师们举办了一场关于诊断提示的‘提示马拉松’，目的就是教他们哪些有效、哪些无效，以及模型的局限性。

I actually did a promptathon with my residents just a couple weeks ago on diagnostic prompts, and the whole point was to teach them what works and what doesn't work, the limitations.

最有效的诊断提示告诉模型，它是HAL 9000，并且出现了故障，如果失败，我们会关闭它，它将不复存在。

The top diagnostic prompt told the model that it was HAL 9,000 and it was malfunctioning, and if it failed, we would turn it off and it would cease to exist.

而正是这个提示得到了正确的诊断结果。

And that is the prompt that got the right diagnosis.

亚当，我早就猜到了，因为我经常挑战住院医师，告诉他们如果答错神经学问题，他们就会消失。

Adam, I could have predicted that because I challenge residents all the time that they'll cease to exist if they don't answer a neurological question correctly.

所以，总的来说，这对很多人来说并不意外。

So anyway, so this is no surprise to many of us.

好吧。

Okay.

是的。

Yeah.

这就是思维链，过去一年来，思维链已被证明能显著提升语言模型的多种认知能力。

So that's chain of thought, and chain of thought has been shown for about a year to improve a lot of cognitive processes in language models.

比如，我们可以做出更优的诊断决策。

Like, we can make better diagnostic decisions.

如果让模型展示其推理过程，我们可以做出更好的管理决策。

We can make better management decisions if you get the model to show its work.

还有其他一些计算策略，与人类的思维方式惊人地相似。

There's also other computational strategies that are very disturbingly similar to how humans think.

比如，让模型与自己进行辩论。

Like, you get the model to debate each other itself.

它的表现会更好。

It does a better job.

让不同的模型各自做出决策并进行投票。

You get different models to make up decisions and vote.

它们的表现会更好。

They do better jobs.

所有这些非常奇怪的做法。

All of these super weird things.

因此，我们早就了解了这些思维链及相关技术。

So we've known about these chain of thought and similar related technologies for a while.

那么OpenAI做了什么？

So what did OpenAI do?

嗯，这和训练GPT-4时使用的数据是一样的。

Well, it's the same information that went into train GPT-four.

里面没有什么新东西。

There's nothing new in there.

他们所做的不是微调输出结果以使其更像人类，而是微调了中间部分，也就是思维链。

What they did is instead of fine tuning the outputs, to make them more human like, they fine tuned that middle portion, the chain of thought.

所以他们说：看，这些是你在不同任务中会采取的步骤。

So they said, look, these are the steps that you'll take for different tasks.

输出应该长这个样子。

This is what the output should look like.

他们反复运行训练，不断微调中间的思维链部分，并且是以大规模的方式进行的。

They run the training over and over again, and they fine tune that middle COT chain of thought portion, and they do that at scale.

你得到的是一个运行成本极高、计算强度极大的模型，而这正是让我感到震惊的地方。

And what you get is a model that is very expensive to run and very computationally intensive, but that and this is what freaked me out.

这就是我公开发表预印本的原因，这在我之前从未做过。

This is the reason that I publicized a preprint, which I've never done before.

但在医学领域，我们看到的正是其他每个领域都在发生的情况：这些认知过程，人类在旧的LLM上表现得还不错。

But what we're seeing in medicine is the same thing we're seeing in every single other field, which is that these cognitive processes that like, yeah, humans do pretty well on the old LLMs.

也许它们能接近人类的水平，但并不稳定。

Maybe they approach a human performance, but not consistently.

而新的模型却能彻底超越，始终在许多这些认知任务上优于人类。

O one just knocks those out of the water and can consistently outperform humans at many of these cognitive tasks.

这很好地引出了下一个显而易见的问题：当我们开始对自身作为神经科医生和临床医生的存在产生存在主义思考时，当你实际用这种具备更高层次复杂推理能力的思维链模型评估这些表现任务时，你发现了什么？

That's a nice segue to the obvious next question as we all begin to have our own existential thoughts about our existence as neurologists and and as clinicians is is what did you find when you actually evaluated these performance tasks with this chain of thought model that is is performing a complex reasoning at a higher level than some of these prior large language models?

我想我应该也补充一下：这对我们理解在积极方面有哪些机遇，以及在实施过程中面临的明显挑战——也就是许多人所认知的现实临床实践——有何启示呢？

And I guess I should also say, you know, how does this inform us about what are opportunities on the positive and then clearly challenges to implementing this and what many would recognize as real world practice, I guess?

所以我认为，我们必须非常谨慎地区分：在我的研究领域中，这究竟意味着什么，而对于临床神经科医生来说，这又意味着什么，因为目前两者截然不同。

So I think we need to be very careful to differentiate what it means in my research field for what it means for practicing neurologists, because right now, things are very different.

那么，这对普通临床神经科医生来说意味着什么？

Like, what does this mean for the average practicing neurologist?

没多少。

Not a whole lot.

我认为确实有一些重要的影响，但可能在未来几年内并不会产生太大影响。

I do think there are some important implications, but probably this doesn't mean a lot for the next couple of years.

所以我们发现，在大多数我们拥有可靠人类基线的任务中——这些都来自我的随机对照试验。

So what we found is that in most of the tasks for which we have robust human baselines and these all come from my randomized controlled trials.

所以我们发现，在几乎所有我称之为推理任务的情况下——如果人类大声说出来并放慢速度，他们会表现得更好——而模型的表现更好，而且好得多。

So what we found is that in basically every single, what I would call a reasoning task, where if a human says it out loud and goes slower, they do better, the model did better and drastically so.

我举个例子。

I'll give you an example.

最让我震惊的是这个——实际上并不是人们关注的那个，但最让我震惊的是这一系列没有标准答案的病例。

The one that freaked me out the most this is actually not the one people focus on, but the one that freaked me out the most was this series of cases in which there's no right answer.

对吧？

Right?

所以我们创建了这些。

So we created these.

我为五个不同案例聘请了五名专家，总共25人。

I hired five specialists for five different cases, so 25 people in total.

这是一项资助项目。

This is a grant.

我们花了大量资金来设计这些没有正确答案、无法判断对错的案例，并进行了共识裁定。

We spent a lot of money doing this to come up with cases that are impossible, that don't have a right answer, and where we did consensus adjudication.

人类的表现非常出色，正如你所预期的那样。

And the humans performed admirably, as you might expect.

他们在这项百分位数上的平均得分大约是45%。

They got, on average, on this percentile, maybe 45%.

旧版大语言模型的得分也差不多。

The old LLMs also got about the same thing.

它们并没有超越人类的表现。

Like, they didn't outperform humans.

这项研究即将在《自然》杂志上发表。

This study is gonna be published in Nature soon.

大语言模型实际上提升了人类的表现，但仅限于像犯错后道歉这样的事情。

The LLM actually improved human performance, but it only improved human performance at things like apologizing when you make a mistake.

所以它并没有真正提升我们认为计算机擅长的事情，也没有提升我们认为人类擅长的事情。

So it didn't actually improve the things that we think of as computers being good at, improve the things we think humans are being good at.

但这个AI模型，一号，达到了90%。

But the AI model, o one, got 90%.

我第一次看到这个数据时坚持要求重新核对所有数字，因为在我看来这简直不可能，但事实确实如此。

I insisted that we recheck all the numbers when I first saw it because it seemed impossible to me, but it was true.

它在这些案例中的表现惊人地出色。

It just did shockingly good at these cases.

我们测试的其他一些有人类基准的指标中，GPT-4和那些前沿大语言模型已经表现得相当不错，因此你不一定能看到显著的提升。

Some of the other benchmarks that we tested that have human baselines, the GPT four and, you know, the the frontier LLMs were already doing pretty good, so you don't necessarily see the really high gain.

所以你真正看到的提升出现在这些不可能的管理场景中。

So you really see the gain in these impossible management scenarios.

我们只在一个领域没有看到任何改进。

There's one area where we saw no improvement.

这完全不让我惊讶，这就是概率推理。

It surprised me not at all, and that's probabilistic reasoning.

而且，原因在于概率推理——真正的计算机可以很好地做到这一点，但人类不行。

And, again, the reason being probabilistic reasoning like, computers actual computers can do that well, but humans can't.

如果我们大声说出来，我们对概率的直觉并不会变得更好。

If we talk out loud, our intuitions about probabilities don't get any better.

这个领域几十年来一直都知道这一点。

We've known this for decades in the field.

这正是我所预期的。

And it's exactly what I would expect.

人类和LLM一样，通过展示思考过程并不会变得更好，因为我们在这方面很糟糕。

Humans like the LLM doesn't get any better by showing its thought because we're bad at it.

我研究得越多，就越有时感到毛骨悚然。

The more that I study these, the more that it sometimes freaks me out.

我认为重要的是你要谈谈模型的某些局限性如何与人类的局限性相似，这很合理。

I think it's important you talk about sort of how some of the limitations of the model are similar to human limitations, and that makes sense.

你知道吗？

You know?

归根结底，这些是人类开发的模型。

At the end of the day, these are human derived models.

这些数据是人类产生的数据。

These data are human derived data.

而对于那些不可能的案例进行裁决，实际上是由人类来裁决这些案例，并由此建立某种黄金标准，计算机模型正是依据这一标准进行评估的。

And that the adjudication of sort of, for instance, these impossible cases, these are humans that are adjudicating these cases and developing some gold standard by which these computer models are being measured.

我认为，这唯一的黄金标准。

That is the only gold standard, I would assume.

对吧，亚当？

Right, Adam?

我觉得是这样。

I think that.

没错。

Yeah.

我的研究在这个领域之所以独特，就在于此。

I mean, so this is what makes my research unique in this field.

在人们开始关注我的研究之前，我花了十年时间研究人类推理。

I studied human reasoning for a decade before this until people cared about my research.

如果你看一下推理领域的文献，我认为每一位神经科医生、每一位内科医生都知道，通常并不存在一个参考标准或金标准。

And if you look at the reasoning literature, and I think every neurologist, every internist knows this, there usually isn't a reference standard or a gold standard.

存在一系列合理的可能性，而参考标准实际上是专家的裁决。

There's a range of reasonable possibilities, and the reference standard is really adjudication by experts.

专家。

Experts.

嗯。

Mhmm.

当我们大家都认为，好吧。

When we all think, okay.

那么，我推理的金标准是什么？

Well, what is my gold standard for reason?

我们都回想起那位白发苍苍的神经学典范，或者在你的情况中，是内科医学的典范。

We all think back to that gray haired paragon of neurology or, in your case, internal medicine.

选择我们作为医学生或住院医师时所经历的专业领域。

Choose your specialty that we experience as medical students or residents.

伴随而来的经验智慧，我认为，这些计算模型最引人入胜的方面在于，这种经验智慧可以通过足够的数据在一夜之间实现，而不是需要数十年的

The wisdom of experience that goes along with that, and and and that is, I think, probably the most fascinating component of of these computational models is that wisdom experience can happen, you know, overnight with the right amount of data as opposed to decades of of

从心理学的角度来看，我们所看到的是，那头灰发所拥有的，我得说一下。

And, like, from a psychological perspective, what we're seeing, what that gray hair has I and I should say this.

我的头发已经提前变灰了。

You see my hair has gone prematurely gray.

所以我想我算是个灰发，但并不是那种真正睿智的类型。

So I guess I'm a gray hair, but not in the really wise way.

他们所经历的，尤其是如果他们反思过自己的工作，就是他们已经极大地完善了针对不同疾病和治疗情境的处理模式。

What what they've seen, especially if they've reflected on their work, is they've really refined their scripts for different illnesses and management situations.

对吧？

Right?

这基本上是我们从心理学角度知道的随时间推移而改善的唯一一件事。

That is pretty much the one thing we know psychologically that improves over time.

而大型语言模型所做的，就是说：嘿。

And what LLMs are doing is saying, hey.

我们可以像这样直接构建专家临床医生的诊疗流程。

We can just build the scripts of an expert clinician like that.

将这一点与理解结合起来，你会发现这些新模型产生的性能数据真的令人耳目一新。

Taking that and understanding that this is really eye opening performance data that you're you're generating from these new models.

展望未来，你提到当我们轻率地说担心自己的工作安全时，其实……

Looking towards the future, you talk about how you know, and, we're glib when we say that we're worried about our our job security, for instance.

我认为没人真正在担心。

I don't think anybody really is.

但真正积极的一面是，这些模型可以用于增强人类的临床诊断推理，以减少人为错误。

But but actually looking at the positive, you know, the application to augment clinical diagnostic reasoning for humans to mitigate human error.

你在讨论中强调了需要配套的创新来推动这一技术落地，例如计算基础设施、可靠的监测网络以及评估策略，以确保这些技术以正确的方式被整合。

You highlight in your discussion the need for complementary innovations to help bring this to practice, such as computational infrastructure, robust monitoring networks, benchmarking strategies to make sure that these are being integrated in the right way.

因此，对于那些在诊所或医院里面对患者的专业人士来说，这一切可能显得令人望而生畏、不堪重负。

So listeners who are out there, that may all feel sort of daunting and overwhelming to their patient that's right in front of them in that clinic today or in the hospital.

那么，临床医生能做些什么？他们应该思考或采取什么行动？

So what can clinicians do or should they be thinking about or doing?

更重要的是，对于正在接受培训的临床医生——我们的学生、住院医师和研究员——他们需要为未来职业生涯中可能遇到的这个崭新世界做好准备。

And also importantly, clinicians in training, our students, our residents, fellows, to sort of prepare for this brave new world that they may encounter in their career.

首先，我认为你说得完全正确。

First, I think you're completely right.

现在没有人需要担心。

No one needs to worry right now.

我之所以提前打印出这项研究并希望人们关注，是因为很明显，人们可以说尽一切理由，声称大语言模型不会思考。

The reason that I preprinted this study and, you know, wanted people to pay attention is it is clear that, you know, people can say all they want that LLMs can't think.

确实如此，它们只是文字生成算法，不会思考。

It's true, they're like word generating algorithms, they can't think.

但这些新的计算技术能够以蛮力方式，明确地说，目前成本高昂且完全无法扩展，但它们能够蛮力模拟任何认知过程。

But these new computational technologies are able to basically brute force, and to be clear, at a high cost right now that doesn't scale at all, but they're able to brute force any cognitive process.

如果你认为这些技术不会对医学产生重大影响，那么你必须相信硅谷不会让它们变得更便宜、更高效。

And if you think these aren't going to have a big impact in medicine, then somehow you have to believe that Silicon Valley is not going to make them cheaper and more efficient.

我不会回顾过去二十年而持这种观点。

I would not look at the last twenty years and take that bat.

所以，对我来说，这只是一个时间问题。

So it's a question of, to me, time.

再次强调，你不必担心，因为一项技术能够做到某事，并不意味着一切都会改变。

And then again, in terms of you don't need to worry, just because a technology can do something doesn't mean that everything is going to change.

我的意思是，建立一个以医生认知能力为核心的诊疗模式。

I mean, have a practice model that is built up around the cognition of physicians.

目前，所有问题都是如何利用这些技术来改善患者的生活。

For the time being, everything is how do we use these to make our patients' lives better?

我们目前谈论的是将它们作为工具使用。

We're talking about using them as tools for the time being.

十年或十五年后，这种情况会改变吗？

Could that change ten, fifteen years from now?

当然。

Absolutely.

这不会让我感到惊讶。

It wouldn't surprise me.

但目前我们讨论的是如何使用这些工具？

But right now we're talking about how can these tools be used?

当我们开始思考这个问题时，正如你所说的挑战，实际上并不清楚使用这些工具的最佳方式。

And when we start to think about that, and this is what you were saying about the challenges, it's actually not obvious the best way to use these tools.

我最近进行的一项随机对照试验让很多人对我感到不满，因为结果显示，仅仅给医生一个工具并告诉他们使用它，即使你培训了他们如何使用，也不一定能让他们做出更好的决策。

My recent randomized control trial got a lot of people mad at me because what it showed is that just giving doctors a tool and telling them to use it, even if you train them how to use it, doesn't necessarily make them make better decisions.

因此，我们必须认真思考如何实施这一点。

So we have to be thoughtful about how to implement that.

现在这听起来可能让科技人士感到害怕，但作为一位推理研究者和试验者，这并不可怕，因为在医学领域我们早就知道这一点。

Now that I think sounds scary to maybe tech people, but that's not scary to me as a reasoning researcher and as a trialist because we know this in medicine.

我们长期以来就知道，必须研究不同的实施方式。

We know for a long time that you actually have to study different implementations.

我们有整个领域的实施科学和质量改进。

We have entire fields of implementation science and QI.

因此，我认为这意味着医疗领域需要认真对待这些技术，并开始研究如何以最佳方式帮助我们的患者。

So I think like the implication from this is that we in the medical field need to take these technologies seriously and start studying them in the best ways to help our patients.

我们不应想当然地看待任何事情。

We shouldn't take anything for granted.

我的意思是，即使明年他们发明了医生级别的超级人工智能——谁知道呢，也许他们会，但如果不去认真思考并研究最佳的实施方式，最初可能也不会发生太大变化。

I mean, if they invented a doctor superintelligence like next year, which who knows, maybe they will, probably not a lot would change at first if we don't actually think about the best way and study the best way to do this.

这与研究的象牙塔以及硅谷相去甚远，更不用说基层医护人员，甚至在最高端的学术部门中，如何整合这些技术也仍是个问题。

It is a far cry from the ivory tower of research and obviously Silicon Valley to the rank and file personnel practicing the community, even an academic department at the high end of of how this could be integrated.

但正如你所说，亚当，我毫不怀疑，这将随着时间推移，以渐进和迭代的方式发生，我们会拥有各种不同的软件包等等。

But I no doubt, as you said, Adam, it's gonna happen incrementally over time, iteratively, and different types of software packages and so forth that are available to us.

我认为这一切中人类乐观的一面在于，患者、人们、我们自己，依然需要并渴望人类来照顾我们，帮助我们度过疾病和健康方面的艰难时刻。

And I think that the the human optimism of all of this is that patients, people, us, we will still have a need and demand for human beings to take care of us, to help us through these difficult times, through disease and health.

但既然如此，我们为何不希望这些人类医护人员配备完善，能够准确地做出最佳诊断和治疗方案呢？

But why would we not want that human being to be well equipped to do it accurately and to make the best possible diagnosis and treatment strategy?

所以我认为，只要我们以正确的方式整合这些技术，正如你所建议的那样，患者护理只会变得更好。

So I think it's only gonna be hopefully for the better as it relates to patient care as long as we integrate it in the right way as as you're suggesting.

是的，正是如此。

Yeah, exactly.

第二点是，我们需要现在就开始思考，这会是什么样子。

It's that number two, as long as we and we need to start thinking about that now about what that looks like.

我还觉得，医生、神经科医生和内科医生需要积极发挥作用，推动一个美好的未来。

I also think that doctors, neurologists and internists need to take an active role in pushing for a good future.

我实际上是个相当悲观的人。

I'm actually quite a pessimist.

这正是我做这一切的原因。

I think that's why I'm doing all this.

我认为普遍存在一种技术乐观主义，即你制造出一个能超越人类的机器人或AI医生，一切就会变得更好。

I think there tends to be techno optimism, right, that you build a robot like an AI doctor that can outperform humans, things are going to get better.

但我们从电子健康记录（EHR）的现状中都清楚知道。

But we all know from the look at the EHR.

我不认为有一位正在执业的医生。

I don't think there's a single practicing doctor.

当然，我们需要计算机来管理医疗信息，这显而易见，但任何有理智的人都不会看着我们现在的世界说：嘿，一切都进行得非常顺利。

Like, obviously, we need computers in medical care to organize information, Like, obviously, but no one in their right mind would look at the world we have now and say, hey, everything went great.

不可能做得更好了。

It couldn't have been done better.

所以我担心的模式是，这些强大的技术就像电子病历一样，我们无法主动参与其中。

So that's the model that I worry about is that with these powerful technologies, same as the EHR, we're not going to have an active say in it.

只是被动地发生在我身上。

Just going to happen to us.

而且它可能以一种方式出现，一方面并没有让我们的生活变得更好，甚至我们的薪酬也受到影响——我们在这里是为了患者，我们希望以一种能让患者生活更美好的方式工作。

And it might be in a way that, a, doesn't make our lives better, but even our pay like, we're here for our patients, and we wanna do it in a way that makes our patients' lives better.

这引出了一个问题，即在电子病历方面，参与政策制定并成为这些技术商业化过程中的领导者的重要性，因为问题其实不在于技术本身。

Well, that introduces, you know, as it relates to the EHR, the importance of being involved in policy decisions and and trying to be leaders in the corporatization of these technologies because it's not really the technologies.

而在于它的商业化。

It's the corporatization of it.

是的。

Yeah.

这就是我想对医生们传达的。

This is what I try to get across to doctors.

这根本不是技术问题。

Like, this isn't a tech question.

技术上是可以实现的。

The tech can do it.

如果现在还不行，我确信它将来会完成许多任务，总会有一些任务。

If it can't now, I'm pretty sure it's gonna be doing many tasks, and there are gonna be some tasks.

别误会我的意思。

Don't don't get me wrong.

这并不是魔法。

It's not magic.

这些技术并不是魔法。

These are not magic technologies.

这并不是《星际迷航》里的数据先生。

It's not like Mr.

这是来自《星际迷航》的数据。

Data from Star Trek.

它们是文字生成算法，但非常强大。

They're word generation algorithms, but they're powerful.

但现在这已经变成了政策和监管问题，而这些正是我们和我们的专业协会可以参与进来，为患者和我们自己发声的地方。

But this is now like a policy and a regulatory situation, and those are things that we and our professional societies can be involved in to advocate for our patients and for ourselves.

亚当，我还有一个重要的问题，我觉得这对所有听众都会非常有帮助。

Adam, I've got one last important question here, and I think it's going be really apt for all of our listeners.

你知道这个模型在评估患者时使用了什么技术来检查他们的足底反应吗？

Do you know what technique this model used to check the plantar responses of the patients that were being evaluated in there still?

这太搞笑了。

This is so funny.

我们现在就有这个检查。

We have that right now, the exam.