Can AI Catch What Doctors Miss? | Eric Topol | TED

翻译人员: Yip Yan Yeung 校对人员: suya f.

I've had the real fortune of working at Scripps Research

在过去的 17 年里，我很幸运 能在斯克利普斯研究所工作。

for the last 17 years.

It's the largest nonprofit biomedical institution in the country.

它是美国最大的 非营利生物医学机构。

And I've watched some of my colleagues,

我看到过一些同事

who have spent two to three years

花了两到三年时间

to define the crystal 3-D structure of a protein.

定义一种蛋白质的晶体三维结构。

Well, now that can be done in two or three minutes.

现在两三分钟就搞定了。

And that's because of the work of AlphaFold,

由于 AlphaFold 的成果，

which is a derivative of DeepMind, Demis Hassabis and John Jumper,

AlphaFold 是 DeepMind 开发的技术，

戴密斯·哈萨比斯（Demis Hassabis） 和约翰·乔普（John Jumper）

recognized by the American Nobel Prize in September.

于 9 月获得了美国诺贝尔奖。

What's interesting, this work,

有趣的是，这项成果，

which is taking the amino acid sequence in one dimension

从一维层面提取氨基酸序列

and predicting the three-dimensional protein at atomic level,

并以原子级别预测三维蛋白质，

[has] now inspired many other of these protein structure prediction models,

（已经）启发了许多其他 蛋白质结构预测模型

as well as RNA and antibodies,

以及 RNA 和抗体，

and even being able to pick up all the missense mutations in the genome,

甚至能发现基因组中 所有的错义突变，

and even being able to come up wit proteins

甚至能提出以前从未被创造、

that have never been invented before, that don't exist in nature.

自然界中不存在的蛋白质。

Now, the only thing I think about this is it was a transformer model,

我唯一想到的一点是 它是一个 Transformer 模型，

we'll talk about that in a moment,

我们之后会谈到，

in this award, since Demis and John

在这个奖项中，由于戴密斯和约翰

and their team of 30 scientists

以及他们由 30 名科学家组成的团队

don't understand how the transformer model works,

不了解 Transformer 模型的工作原理，

shouldn't the AI get an asterisk as part of that award?

难道 AI 不应该 从这个奖项中分一杯羹吗？

I'm going to switch from life science,

我将从生命科学，

which has been the singular biggest contribution just reviewed,

也就是我们刚刚说到最重大的贡献，

to medicine.

谈到医学。

And in the medical community,

在医学界，

the thing that we don't talk much about are diagnostic medical errors.

我们不太谈论的是医疗诊断错误。

And according to the National Academy of Medicine,

根据美国国家医学院的说法，

all of us will experience at least one in our lifetime.

我们所有人一生中 都会经历至少一次。

And we know from a recent Johns Hopkins study

我们从约翰·霍普金斯大学 最近的一项研究中得知，

that these errors have led to 800,000 Americans dead

这些错误每年会导致 80 万美国人死亡

or seriously disabled each year.

或严重残疾。

So this is a big problem.

这可是个大问题。

And the question is, can AI help us?

问题是， AI 能帮助我们吗？

And you keep hearing about the term “precision medicine.”

你总是会听到“精准医疗”这个词。

Well, if you keep making the same mistake over and over again, that's very precise.

如果你一遍又一遍地 犯同样的错误，那确实非常精确。

(Laughter)

（笑声）

We don't need that,

我们才不要这样，

we need accuracy and precision medicine.

我们需要准确和精准的医疗。

So can we get there?

我们能达成这个目标吗？

Well, this is a picture of the retina.

这是一张视网膜的照片。

And this was the first major hint,

这是第一个重要迹象，

training 100,000 images with supervised learning.

通过监督学习训练十万张图像。

Could the machine see things that people couldn't see?

机器能看见人们看不见的东西吗？

And so the question was, to the retinal experts,

对视网膜专家问这么一个问题：

is this from a man or a woman?

它来自男性还是女性？

And the chance of getting it accurate was 50 percent.

答对的概率是 50％。

(Laughter)

（笑声）

But the AI got it right, 97 percent.

但 AI 做对了 97％。

So that training,

在这种训练中，

the features are not even fully defined of how that was possible.

能达成这样效果的特征 甚至没有被完整地定义。

Well that gets then to all of medical images.

来说说各种医疗图像。

This is just representative, the chest X-ray.

举个代表性的例子， 胸部 X 光检查。

And in fact with the chest X-ray,

其实在胸部 X 光片中，

the ability here for the AI to pick up,

AI 能够发挥的能力是识别，

the radiologists, expert radiologists missing the nodule,

在放射科医生、 放射科专家医生遗漏了结节时，

which turned out to be picked up by the AI as cancerous,

发现结节是癌性的，

and this is, of course, representative of all of medical scans,

当然这也代表了所有医学扫描，

whether it’s CT scans, MRI, ultrasound.

无论是 CT 扫描、 核磁共振成像还是超声波。

That through supervised learning of large, labeled, annotated data sets,

通过监督学习 带标签和注释的大型数据集，

we can see AI do at least as well, if not better,

我们可以看到 AI 的表现 即使不胜过，

than expert physicians.

也与专家医生相当。

And 21 randomized trials of picking up polyps --

还有 21 场检测息肉的随机试验，

machine vision during colonoscopy -- have all shown

在结肠镜检查过程中使用机器视觉，

that polyps are picked up better

全都表明，在机器视觉的帮助下，

with the aid of machine vision than by the gastroenterologist alone,

比单靠胃肠病学家 发现息肉的效果更好，

especially as the day goes on, later in the day, interestingly.

尤其是到了一天晚些时候。

We don't know whether picking up all these additional polyps

我们不知道 额外识别出这些息肉

changes the natural history of cancers,

是否会改变癌症的自然病史，

but it tells you about machine eyes,

但是它展现了机器眼，

the power of machine eyes.

机器眼的力量。

Now that was interesting.

很有意思。

But now still with deep learning models, not transformer models,

但它用的还是深度学习模型， 不是 Transformer 模型，

we've seen and learned that the ability

我们已经见证并了解到

for computer vision to pick up things that human eyes can't see

计算机视觉识别 人眼看不见的东西的能力

is quite remarkable.

非常出色。

Here's the retina.

这是视网膜。

Picking up the control of diabetes and blood pressure.

检测到糖尿病和血压控制。

Kidney disease.

肾脏疾病。

Liver and gallbladder disease.

肝胆疾病。

The heart calcium score,

心脏钙化积分

which you would normally get through a scan of the heart.

通常是通过心脏扫描得出的。

Alzheimer's disease before any clinical symptoms have been manifest.

在出现临床症状之前 就诊断出阿尔茨海默病。

Predicting heart attacks and strokes.

预测心脏病发作和中风。

Hyperlipidemia.

高脂血症。

And seven years before any symptoms of Parkinson's disease,

出现帕金森氏病症状的 七年以前发现患病。

to pick that up.

Now this is interesting because in the future,

这很有趣，因为将来，

we'll be taking pictures of our retina at checkups.

我们将在检查的时候 拍下视网膜的照片。

This is the gateway to almost every system in the body.

这是通往人体几乎所有系统的门户。

It's really striking.

真的很惊人。

And we'll come back to this because each one of these studies

我们还会回来讨论这个问题， 因为每项研究

was done with tens or hundreds [of] thousands of images

都是通过监督学习 使用成千上万张图像完成的，

with supervised learning,

and they’re all separate studies by different investigators.

是由不同的研究人员 分别进行的研究。

Now, as a cardiologist, I love to read cardiograms.

作为一名心脏病专家， 我喜欢阅读心电图。

I've been doing it for over 30 years.

我已经做了 30 多年了。

But I couldn't see these things.

但我看不到这些东西。

Like, the age and the sex of the patient,

比如，患者的年龄和性别，

or the ejection fraction of the heart,

或者心脏射血分数，

making difficult diagnoses that are frequently missed.

做出常常被忽略的困难诊断。

The anemia of the patient, that is, the hemoglobin to the decimal point.

患者的贫血， 即血红蛋白容量极低。

Predicting whether a person,

通过心电图预测一个 从未发生过房颤或中风的人

who's never had atrial fibrillation or stroke

from the ECG,

whether that's going to likely occur.

是否会出现症状。

Diabetes, a diagnosis of diabetes and prediabetes, from the cardiogram.

糖尿病，根据心电图 诊断糖尿病和糖尿病前期。

The filling pressure of the heart.

心脏充盈压。

Hypothyroidism

甲状腺功能减退和肾脏疾病。

and kidney disease.

Imagine getting an electrocardiogram to tell you about all these other things,

想象一下，让心电图 告诉你这些额外的事，

not really so much about the heart.

不仅仅关乎心脏。

Then there's the chest X-ray.

然后是胸部 X 光片。

Who would have guessed that we could accurately determine

谁能猜到我们可以准确地判断

the race of the patient,

患者的种族

no less the ethical implications of that,

以及与之相关的伦理意蕴，

from a chest X-ray through machine eyes?

而这都是通过由机器眼 看到的胸部 X 光片得出的？

And interestingly, picking up the diagnosis of diabetes,

有趣的是，还能通过胸片 诊断糖尿病

as well as how well the diabetes is controlled,

和糖尿病的控制情况，

through the chest X-ray.

And of course, so many different parameters about the heart,

当然，心脏有这么多不同的指标，

which we could never,

无论是放射科医生还是心脏病专家， 我们永远都无法做到

radiologists or cardiologists, could never be able to come up

with what machine vision can do.

机器视觉能做到的这么多诊断。

Pathologists often argue about a slide,

病理学家总是会争论一张片子 到底展现了什么。

about what does it really show?

But with this ability of machine eyes,

但是，凭借这种机器眼的能力，

the driver genomic mutations of the cancer can be defined,

可以定义癌症的 驱动基因组突变，

no less the structural copy number variants

还可以看出导致或出现在 这个肿瘤内的结构拷贝数变异。

that are accounting or present in that tumor.

Also, where is that tumor coming from?

还有，肿瘤从何而起？

For many patients, we don’t know.

对于许多患者来说，我们不知道。

But it can be determined through AI.

但可以通过 AI 确定。

And also the prognosis of the patient,

还有患者的预后，

just from the slide,

只需要通过各种训练 分析片子得出。

by all of the training.

Again, this is all just convolutional neural networks,

同样，这只是卷积神经网络，

not transformer models.

不是 Transformer 模型。

So when we go from the deep neural networks to transformer models,

当我们从深度神经网络 转向 Transformer 模型时，

this classic pre-print,

这份经典的预印本，

one of the most cited pre-prints ever,

有史以来被引用次数最多的预印本之一，

《Attention is All You Need》 （意为“注意力足矣”），

"Attention is All You Need,"

the ability to now be able to look at many more items,

可以处理更多对象的能力，

whether it be language or images,

无论是语言还是图像，

and be able to put this in context,

并能够将其置于上下文中，

setting up a transformational progress in many fields.

在许多领域取得了变革性进展。

The prototype is, the outgrowth of this is GPT-4.

这个模型的原型 或成果就是 GPT-4。

With over a trillion connections.

其拥有超过一万亿个连接。

Our human brain has 100 trillion connections or parameters.

我们的人脑有 100 万亿个 神经连接或参数。

But one trillion,

但是，一万亿，

just think of all the information, knowledge,

想一想这一万亿连接中 包含的所有信息、知识。

that's packed into those one trillion.

And interestingly, this is now multimodal with language, with images,

有趣的是，现在已经支持多模态， 包括语言、图像、语音。

with speech.

And it involves a massive amount of graphic processing units.

还包含大量的图形处理单元。

And it's with self-supervised learning,

还有自监督学习，

which is a big bottleneck in medicine

这是医学界的一大瓶颈，

because we can't get experts to label images.

因为我们不能让专家给图像打标签。

This can be done with self-supervised learning.

这可以通过自监督学习来完成。

So what does this set up in medicine?

这在医学中起到了什么作用呢？

It sets up, for example, keyboard liberation.

比如，它带来了键盘解放。

The one thing that both doctors, clinicians

这是医生、临床医师

and patients would like to see.

和患者都希望看到的一件事。

Everyone hates being data clerks as clinicians,

每个临床医师都不想当数据员，

and patients would like to see their doctor

等了好久终于可以看病的时候， 患者希望可以见到医生。

when they finally have the visit they've waited for a long time.

So the ability to change the face-to-face contact

因此，改变面对面接触的能力

is just one step along the way.

只是前进道路中的一步。

By having the liberation from keyboards with synthetic notes

借助从对话中得到、生成的合成笔记

将人们从键盘解放出来，

that are driven, derived from the conversation,

and then all the downstream normal data clerk functions that are done,

在非工作时间完成 各种数据员的常规后续工作。

often off-hours.

Now we're seeing in health systems across the United States

我们能在美国各地的卫生系统中看到，

where people, physicians are saving many hours of time

人们、医生节省了大量的时间，

and heading towards ultimately keyboard liberation.

最终走向键盘解放。

We recently published, with the group at Moorfields Eye Institute,

最近，我们与皮尔斯·基恩 （Pearse Keane）领导的

led by Pearse Keane,

莫菲尔德眼科研究所的 研究小组一起发布了

the first foundation model in medicine from the retina.

医学界第一个基于视网膜的基础模型。

And remember those eight different things that were all done by separate studies?

还记得那八件由不同研究完成的事吗？

This was all done with one model.

都是用一个模型完成的。

This is with 1.6 million retinal images

用了 160 万张视网膜图像

predicting all these different outcome likelihoods.

预测了各种不同结果的可能性。

And this is all open-source,

这都是开源的，

which is of course really important that others can build on these models.

当然非常重要，这样其他人 可以基于这些模型开发。

Now I just want to review a couple of really interesting patients.

我想回顾几个非常有趣的患者。

Andrew, who is now six years old.

安德鲁，现年六岁。

He had three years of relentlessly increasing pain, arrested growth.

三年来，他痛苦持续加剧，成长受阻。

His gait suffered with a dragging of his left foot,

他的步态因左脚 牵扯性疼痛而受到影响，

he had severe headaches.

头痛严重。

He went to 17 doctors over three years.

他在三年内去看了 17 位医生。

His mother then entered all his symptoms into ChatGPT.

然后，他的母亲将他所有的症状 输入进了 ChatGPT。

It made the diagnosis of occulta spina bifida,

它诊断为隐性脊柱裂，

which meant he had a tethered spinal cord that was missed by all 17 doctors

这意味着他患有脊髓栓系， 三年内的所有 17 位医生

over three years.

都没有注意到。

He had surgery to release the cord.

他接受了脊髓栓系松解手术。

He's now perfectly healthy.

现在非常健康。

(Applause)

（掌声）

This is a patient that was sent to me,

这是一位被送到我这里的病人，

who was suffering with, she was told, long COVID.

她被告知患有“长新冠”。

She saw many different physicians, neurologists,

她看了许多不同的医生、 神经科医生，

and her sister entered all her symptoms after getting nowhere,

她的姐妹把她所有的症状， 在经历了一路碰壁、

no treatment for long COVID,

长新冠无药可救、

there is no treatment validated,

没有经过验证的治疗方法后，

and her sister put all her symptoms into ChatGPT.

将所有症状都输入了 ChatGPT。

It found out it actually was not long COVID,

它发现其实并不是长新冠，

she had limbic encephalitis, which is treatable.

而是边缘系统脑炎， 是可以治疗的。

She was treated, and now she's doing extremely well.

她接受了治疗，现在情况非常好。

But these are not just anecdotes anymore.

但这些不再只是个例了。

70 very difficult cases

70 例非常困难的病例

that are the clinical pathologic conferences

登上《新英格兰医学杂志》的 临床病理学会议，

at the New England Journal of Medicine

were compared to GPT-4,

与 GPT-4 进行了比较，

and the chatbot did as well

聊天机器人在做出诊断方面的表现

or better than the expert master clinicians

与临床专家相当或更好。

in making the diagnosis.

So I just want to close with a recent conversation with my fellow.

我想以最近与我的同事的对话收尾。

Medicine is still an apprenticeship,

医学仍然采用的是“师徒制”，

and Andrew Cho is 30 years old,

安德鲁·赵（Andrew Cho） 今年 30 岁，

in his second year of cardiology fellowship.

是他攻读心脏病学培训的第二年。

We see all patients together in the clinic.

我们一起在诊所为所有患者看病。

And at the end of clinic the other day,

有一天在看诊结束时，

I sat down and said to him,

我坐下来对他说：

"Andrew, you are so lucky.

“安德鲁，你真幸运。

You're going to be practicing medicine in an era of keyboard liberation.

你能在键盘解放的时代 从事医学工作。

You're going to be connecting with patients

你会以我们几十年来前所未有的方式

the way we haven't done for decades."

与患者接触。”

That is the ability to have the note

这就是能够从对话中获取笔记

and the work from the conversation

和工作成果的能力，

to derive things like pre-authorization,

从而得出诸如预授权、

billing, prescriptions, future appointments --

账单、处方、未来预约之类 我们要做的事，

all the things that we do,

including nudges to the patient.

包括提示患者。

For example, did you get your blood pressure checks

比如，你有没有量血压，

and what did they show

得到的结果是什么意思，

and all that coming back to you.

这些都能回到你的手中。

But much more than that,

但不仅如此，

to help with making diagnoses.

还有助于做出诊断。

And the gift of time

还有时间上的优势，

that having all the data of a patient

在见到病人之前 就已经准备好了患者的所有数据。

that's all teed up before even seeing the patient.

And all this support changes the future of the patient-doctor relationship,

这些帮助都改变了医患关系的未来，

bringing in the gift of time.

带来了时间的恩赐。

So this is really exciting.

这真的很令人兴奋。

I said to Andrew, everything has to be validated, of course,

我对安德鲁说， 当然，这一切都必须经过验证，

that the benefit greatly outweighs any risk.

证明好处远大于任何风险。

But it is really a remarkable time for the future of health care,

但是对于医疗保健的未来来说， 这确实是一个重大的时刻，

it's so damn exciting.

真是令人兴奋。

Thank you.

谢谢。

(Applause)

（掌声）