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

譯者: Esther Lam 審譯者: Shelley Tsang 曾雯海

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

我很幸運，過去十七年 能在斯克里普斯研究所工作。

for the last 17 years.

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

它是全國最大的非營利 生物醫學機構。

And I've watched some of my colleagues,

我看過我的一些同事

who have spent two to three years

花兩到三年的時間

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

去定義一種蛋白質的三維立體結構。

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

現在只要兩到三分鐘就能完成。

And that's because of the work of AlphaFold,

那要歸功於 AlphaFold，

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

它是由 DeepMind 衍生出來的， 戴米斯‧哈薩比斯和約翰‧瓊波

recognized by the American Nobel Prize in September.

他們是被美國人認可的專家學者 在九月得到諾貝爾獎。

What's interesting, this work,

有趣的是，這項產品

which is taking the amino acid sequence in one dimension

是用一個維度上的氨基酸序列

and predicting the three-dimensional protein at atomic level,

去預測原子層級的三維蛋白質，

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

現在已經促成了許多 其他的蛋白質結構預測模型，

as well as RNA and antibodies,

還有核糖核酸和抗體，

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

甚至可以挑出基因組中的錯義突變，

and even being able to come up wit proteins

甚至還能提供

以前從未發明出來, 在大自然中也不存在的蛋白質。

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

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

從而我現在唯一的想法是， 它是個轉換模型，

we'll talk about that in a moment,

我們稍後會討論這個，

in this award, since Demis and John

在這個獎項中，因為戴米斯和約翰

以及他們的三十名科學家團隊

and their team of 30 scientists

don't understand how the transformer model works,

不了解轉換模型怎麼運作，

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

難道人工智慧不應該也得到一個詮釋星號， 作為該獎項的一部分？

I'm going to switch from life science,

我接著要從生命科學，

which has been the singular biggest contribution just reviewed,

就是我們剛才提到的，迄今為止最大的貢獻，

to medicine.

換到醫學。

And in the medical community,

在醫學界，

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

我們不太會去談診斷醫療錯誤。

And according to the National Academy of Medicine,

根據國家醫學院，

我們所有人一生中都至少 會遇到一次這種錯誤。

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

And we know from a recent Johns Hopkins study

根據約翰‧霍普金斯大學 最近的一項研究，

that these errors have led to 800,000 Americans dead

這些錯誤，每年會導致八十萬名美國人死亡

或嚴重殘疾。

or seriously disabled each year.

So this is a big problem.

所以這是個大問題。

And the question is, can AI help us?

問題是，人工智慧能幫我們嗎？

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

我們不斷聽到「精準醫療」這個詞。

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

如果你持續不斷犯同樣的錯誤， 那的確很精準。

(Laughter)

（笑聲）

We don't need that,

那不是我們需要的， 我們需要正確精準醫療。

we need accuracy and precision medicine.

So can we get there?

我們能做到嗎？

Well, this is a picture of the retina.

好，這是一張視網膜圖。

And this was the first major hint,

這是第一個重要提示，

training 100,000 images with supervised learning.

背後是監督式學習方法搭配 十萬張影像的訓練資料。

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

機器能看見人類看不見的東西嗎？

And so the question was, to the retinal experts,

給視網膜專家的問題是：

is this from a man or a woman?

這是男性或女性的視網膜？

And the chance of getting it accurate was 50 percent.

而答對的機率是 50%。

(Laughter)

（笑聲）

But the AI got it right, 97 percent.

但人工智慧答對了，97% 的正確率。

So that training,

所以，那種訓練，

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

這些特徵甚至還沒有完全定義出來，究竟是如何可能的。

Well that gets then to all of medical images.

嗯，這就涉及到所有的醫學圖像。

This is just representative, the chest X-ray.

這只是個代表性的例子，胸部 X 光。

And in fact with the chest X-ray,

事實上，就胸部 X 光來說，

the ability here for the AI to pick up,

人工智慧的挑選能力，

the radiologists, expert radiologists missing the nodule,

放射科醫生，專業的放射科醫生 沒發現的小瘤，

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

結果卻被人工智慧挑出來， 且判斷是惡性的，

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

這個例子可以代表所有的醫學掃描，

whether it’s CT scans, MRI, ultrasound.

不論是電腦斷層、 磁振造影、超音波。

透過監督式學習方法搭配大型、 加上標籤、加上註解的資料集，

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

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

我們可以看到人工智慧 表現至少會和專業醫生一樣好，

甚至更好。

than expert physicians.

在二十一次的隨機試驗中 去找取息肉，

And 21 randomized trials of picking up polyps --

machine vision during colonoscopy -- have all shown

全部試驗都顯示, 在大腸鏡檢查中,

借助機器視覺

that polyps are picked up better

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

比單靠胃腸科醫師的表現更好

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

有趣的是，在一天的 越後段時期越是如此。

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

我們不知道另外多找出 這些息肉是否會

changes the natural history of cancers,

改變癌症的自然史，

but it tells you about machine eyes,

但它能告訴你機器眼睛，

the power of machine eyes.

機器眼睛的力量。

Now that was interesting.

那是很有趣的。

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

但是现在仍然使用深度學習模型， 而不是變壓器模型，

we've seen and learned that the ability

我們已經看到並了解到，

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

電腦視覺 拾取人眼看不見的東西的能力

is quite remarkable.

非常出色。

Here's the retina.

這是視網膜。

Picking up the control of diabetes and blood pressure.

控制糖尿病和血壓.。

Kidney disease.

腎臟疾病。

Liver and gallbladder disease.

肝胆疾病。

The heart calcium score,

心臟钙分数,

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

這通常是通過心臟掃描得出的。

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

早於阿茲海默症之前 就已明顯的臨床症狀。

Predicting heart attacks and strokes.

預測心臟病發作和中風。

Hyperlipidemia.

高脂血症。

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

甚至在出现帕金森氏病症状的七年前，

to pick that up.

都可以被機器查出來。

Now this is interesting because in the future,

這很有趣 因為在未來，

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

我們在檢查的時候 會拍照視網膜。

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

這是通往身體幾乎所有系統的門戶。

It's really striking.

這真的很驚人。

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

我們會回到這個主題， 因為這些研究中的每一項

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

都是透過監督學習 使用數十或數百,數千張圖像

with supervised learning,

完成的，

and they’re all separate studies by different investigators.

而且它们都是 由不同的研究人员單獨的研究。

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

现在，作為一名心臟科專家， 我喜歡閱讀心電圖。

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.

但我看不到這些東西。

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,

或者心臟的射血分數，

making difficult diagnoses that are frequently missed.

做出經常被遗漏並且困難的診斷。

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

患者的貧血情況， 即血红蛋白精確到小數點。

Predicting whether a person,

預測一個從未出現

who's never had atrial fibrillation or stroke

房顫或中風的人

from the ECG,

是否從心電圖上

whether that's going to likely occur.

能看出這種情況是否可能發生。

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

糖尿病，從心電圖檢查中診斷糖尿病和糖尿病前期。

The filling pressure of the heart.

心臟充盈壓。

Hypothyroidism

甲狀腺功能减退

and kidney disease.

和腎臟疾病。

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

想像一下，透過心電圖檢查告訴您 有關所有這些其他事物的資訊，

not really so much about the heart.

而不僅僅是關於心臟的資訊。

Then there's the chest X-ray.

然後是胸部 X 光片。

Who would have guessed that we could accurately determine

誰能猜到 我們竟然可以透過機器視覺從胸腔X光中準確地判斷，

the race of the patient,

患者的種族，

no less the ethical implications of that,

更不用說這其中的倫理影響呢。

from a chest X-ray through machine eyes?

誰會猜到我们可以透過機器眼睛進行胸部X光檢查？

And interestingly, picking up the diagnosis of diabetes,

而有趣的是，

as well as how well the diabetes is controlled,

透過胸部 X光片 了解糖尿病的診斷

以及糖尿病的控制情况。 當然，心臟有這麼多不同的参数，

through the chest X-ray.

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

無論是

which we could never,

radiologists or cardiologists, could never be able to come up

放射科醫生或心臟科醫生，

永遠無法想像機器視覺能做什麼。

with what machine vision can do.

Pathologists often argue about a slide,

病理學家經常爭論投影片，

about what does it really show?

爭論它到底顯示了什麼？

But with this ability of machine eyes,

但是，憑藉著機器眼睛的這種能力，

the driver genomic mutations of the cancer can be defined,

可以定義癌症的驅動基因組突變，

no less the structural copy number variants

更不用說

that are accounting or present in that tumor.

構成 或存在於該腫瘤中的結構 或者拷貝數變異了。

Also, where is that tumor coming from?

另外，那腫瘤來自哪裡？

For many patients, we don’t know.

對於許多患者來說，我們不知道。

But it can be determined through AI.

但這可以通過人工智能來確定。

And also the prognosis of the patient,

還有患者的預後，

just from the slide,

光從幻燈片來看，

by all of the training.

透過所有的訓練。

Again, this is all just convolutional neural networks,

同樣的， 這只是卷積神經網絡，

not transformer models.

而不是變壓器模型。

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

所以當我們從深層神經網路到變壓器模型，

this classic pre-print,

這本經典的預印本，

one of the most cited pre-prints ever,

有史以來被引用最多的預印本之一，

"Attention is All You Need,"

“注意力就是你所需要的”

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

现在 能够查看更多项目，

whether it be language or images,

無論是語言還是圖像，

and be able to put this in context,

並能夠將其放在上下文中，

setting up a transformational progress in many fields.

從而在許多領域取得變革性進展。

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

直接地說，這的延伸就是 GPT-4

With over a trillion connections.

擁有超過一兆的連接。

Our human brain has 100 trillion connections or parameters.

我們人的大腦有100兆 連接或參數。

But one trillion,

但是，一兆，

just think of all the information, knowledge,

想一想

that's packed into those one trillion.

就一兆中所包含的信息、知識。

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

有趣的是，现在這是一種多模式， 包括語言、圖像、

with speech.

語音。

And it involves a massive amount of graphic processing units.

而且它涉及大量 的圖形處理單元。

And it's with self-supervised learning,

還有透過自我監督學習，

which is a big bottleneck in medicine

這是醫學上的大瓶頸，

because we can't get experts to label images.

因為我們找不到專家 來標記圖像。

This can be done with self-supervised learning.

這可以透過自我監督學習 來完成。

So what does this set up in medicine?

那麼這在醫學上起到了什麼作用？

It sets up, for example, keyboard liberation.

例如，它设置了 鍵盤解放。

The one thing that both doctors, clinicians

這是 醫師、臨床醫師

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,

每個人都討厭把床醫師的身份當數據員，

and patients would like to see their doctor

患者希望

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

在終於等到他們期待已久的就診時 見到他們的醫生。

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

因此，改變面對面接觸的型式

is just one step along the way.

只是這條路上的一步。

By having the liberation from keyboards with synthetic notes

通過從鍵盤中解放出來，使用合成音符

that are driven, derived from the conversation,

這些音符是從對話中生成的，

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

然後是通常在非工作時間完成的所有下游

often off-hours.

常規數據員功能。

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

現在，我們在美國各地的衛生系統中看到，

where people, physicians are saving many hours of time

人們和醫生都在節省大量時間

and heading towards ultimately keyboard liberation.

最終走向鍵盤解放。

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

最近，我們與皮爾斯·基恩

led by Pearse Keane,

領導的莫菲爾德眼科研究所的研究小組一起發布了

the first foundation model in medicine from the retina.

第一個視網膜醫學基礎模型。

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

還記得我剛才提到那八件不同的事情 都是通過單獨的研究完成的嗎？

This was all done with one model.

這一切都是用一個模型完成的。

This is with 1.6 million retinal images

也就是说，有160萬張視網膜圖像

predicting all these different outcome likelihoods.

預測了所有這些不同的結果的可能性。

And this is all open-source,

這些都是公開的資源，

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

當然非常重要的，其他人可以在這些模型上進行構建。

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

現在我只想回顧幾個非常有趣的患者。

Andrew, who is now six years old.

安德魯，現年六歲。

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

他三年來一直遭受不斷增加的疼痛，停滯的生長。

His gait suffered with a dragging of his left foot,

他的步態因左腳拖曳而受到影響，

he had severe headaches.

頭痛嚴重。

He went to 17 doctors over three years.

他在三年内去看了17位醫生。

His mother then entered all his symptoms into ChatGPT.

然後，他的母親將所有他的症狀 都輸入到了ChatGPT中。

It made the diagnosis of occulta spina bifida,

它診斷為隱性脊柱裂，

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

這意味著他患有脊髓系繩， 三年多來

over three years.

所有17位醫生都漏掉了這條脊髓。

He had surgery to release the cord.

他接受了鬆開索帶的手術。

He's now perfectly healthy.

他现在非常健康。

(Applause)

（掌聲）

This is a patient that was sent to me,

這是一位被送到我這裡的病人，

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

她被告知患有長期新冠病毒

She saw many different physicians, neurologists,

她見過許多不同的醫生、神經科醫生，

and her sister entered all her symptoms after getting nowhere,

她的姐姐在別無選擇之後 輸入了所有症狀，

no treatment for long COVID,

沒有治療長期新冠病毒，

there is no treatment validated,

沒有經過驗證的治療方法，

and her sister put all her symptoms into ChatGPT.

姐姐將所有症狀都輸入了ChatGPT。

It found out it actually was not long COVID,

它發現實際上不是長期新冠病毒，

she had limbic encephalitis, which is treatable.

她患有邊緣性腦炎，這是可以治療的。

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

她接受了治療，現在情況非常好。

But these are not just anecdotes anymore.

但這些不再只是軼事了。

70 very difficult cases

將70例非常困難的病例

that are the clinical pathologic conferences

在

at the New England Journal of Medicine

《新英格蘭醫學雜誌》的臨床病理學會議上，

were compared to GPT-4,

與 GPT-4 進行了比較，

and the chatbot did as well

聊天機器人在做出診斷方面的表現

or better than the expert master clinicians

與大師級的臨床醫師一樣好

in making the diagnosis.

或更好

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

因此，我只想以最近與我的同伴的談話作為結束。

Medicine is still an apprenticeship,

醫學仍然是學徒，

and Andrew Cho is 30 years old,

安德魯·曹 今年30歲

in his second year of cardiology fellowship.

是他以獎學金攻讀心臟病學的第二年。

We see all patients together in the clinic.

我們一起在診所為所有患者看病。

And at the end of clinic the other day,

前幾天在診所結束時，

I sat down and said to him,

我坐下來對他說：

"Andrew, you are so lucky.

安德鲁，你真幸運。

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

在鍵盤解放的時代，你將從事醫學工作。

You're going to be connecting with patients

你將以幾十年來從未有過的方式與患者

the way we haven't done for decades."

建立聯繫。

That is the ability to have the note

這就是能夠從對話中獲得醫療筆記

and the work from the conversation

和工作成果，

to derive things like pre-authorization,

從而得出諸如預授權、

billing, prescriptions, future appointments --

帳單、處方、未來預約

all the things that we do,

等所有我們所做的事情，

including nudges to the patient.

包括向患者推動。

For example, did you get your blood pressure checks

例如，你有沒有收到血壓檢查，

and what did they show

檢查結果顯示了什麼，

and all that coming back to you.

所有這些都回覆給你。

But much more than that,

但不僅如此，

to help with making diagnoses.

還有助於做出診斷。

And the gift of time

還有時間的禮物，

that having all the data of a patient

那就是在看病人之前，

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

就已經準備好了患者的所有數據。

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

所有這些支持都改變了未來醫生和病人的關係的，

bringing in the gift of time.

引入時間的恩賜。

So this is really exciting.

所以這真的很令人興奮。

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

我對安德魯說，當然，一切都必須得到驗證，

that the benefit greatly outweighs any risk.

確保收益遠大於任何風險。

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

但是對於醫療保健的未來來说， 這確實是一個非凡的時刻，

it's so damn exciting.

真的令人興奮。

Thank you.

謝謝。

(Applause)

（掌聲）