The doctor determines that you have high cholesterol

你去看醫生，接受了一些檢查。

and you would benefit from medication to treat it.

醫生診斷出你的膽固醇過高，

So you get a pillbox.

建議你服藥治療可能有幫助。

You have some confidence,

所以，你拿到了藥罐子。

your physician has some confidence that this is going to work.

你有點信心，

The company that invented it did a lot of studies, submitted it to the FDA.

你的醫師也有信心，認為這藥會有效。

They studied it very carefully, skeptically, they approved it.

發明這個藥的公司做了很多的研究， 然後呈送給食品藥物管理局。

They have a rough idea of how it works,

他們很仔細、審慎地研究， 並核准了這藥物上市。

they have a rough idea of what the side effects are.

他們大概知道這藥物如何運作，

It should be OK.

也大略知道會有什麼副作用，

You have a little more of a conversation with your physician

應該沒問題。

and the physician is a little worried because you've been blue,

你跟醫師又多聊了一會，

haven't felt like yourself,

而醫師有點擔心，因為你很憂鬱，

you haven't been able to enjoy things in life quite as much as you usually do.

精神欠佳。

Your physician says, "You know, I think you have some depression.

無法像平常一樣盡情享受生活點滴。

I'm going to have to give you another pill."

你的醫師說： 「我認為你有一點精神憂鬱，

So now we're talking about two medications.

我再開個藥給你。」

This pill also -- millions of people have taken it,

所以，我們現在有兩種藥了。

the company did studies, the FDA looked at it -- all good.

這個藥也有好幾百萬人服用過，

Think things should go OK.

公司做了研究，食品藥物管理局 也檢查過，全部都沒問題。

Think things should go OK.

想一下，這東西沒問題，OK的。

Well, wait a minute.

想一下，這東西沒問題，OK的。

How much have we studied these two together?

但，請等一下。

Well, it's very hard to do that.

我們對這兩種藥混在一起吃 做了多少研究？

In fact, it's not traditionally done.

其實，這很難評估。

We totally depend on what we call "post-marketing surveillance,"

事實上，傳統上都不會做。

after the drugs hit the market.

在藥物上市後，我們完全倚賴一種

How can we figure out if bad things are happening

叫做「上市後監察系統」的機制，

between two medications?

我們要如何確認，兩種藥之間

Three? Five? Seven?

是否有什麼不好的事會發生？

Ask your favorite person who has several diagnoses

三種？五種？七種呢？

how many medications they're on.

問你身邊有各種疾病在身的人，

Why do I care about this problem?

他們正在吃多少藥。

I care about it deeply.

為什麼我在乎這個問題？

I'm an informatics and data science guy and really, in my opinion,

我非常在乎。

the only hope -- only hope -- to understand these interactions

我是念資訊和數據科學的人， 真的，在我看來，

is to leverage lots of different sources of data

了解藥物彼此間的交互影響 唯一的希望只有

in order to figure out when drugs can be used together safely

運用不同來源的龐大資料，

and when it's not so safe.

才能找出這些藥 何時可以安全地一起服用，

So let me tell you a data science story.

以及何時不行。

And it begins with my student Nick.

所以，讓我來告訴各位 一個數據科學的故事。

Let's call him "Nick," because that's his name.

這要從我的學生尼克開始講起。

(Laughter)

我們就稱呼他為尼克吧， 因為那就是他的本名。

Nick was a young student.

（笑聲）

I said, "You know, Nick, we have to understand how drugs work

尼克很年輕，

and how they work together and how they work separately,

我說：「尼克， 我們必須了解藥物如何運作，

and we don't have a great understanding.

以及藥物在一起會如何運作、 分開會如何運作，

But the FDA has made available an amazing database.

而我們並沒有了解很深。」

It's a database of adverse events.

但食品藥物管理局已經 有一個很驚人的資料庫，

They literally put on the web --

是一個藥物不良反應通報資料庫。

publicly available, you could all download it right now --

資料真的直接放在網路上

hundreds of thousands of adverse event reports

供大眾查詢，你現在就可以全部下載，

from patients, doctors, companies, pharmacists.

從病人、醫生、公司、藥劑師通報上來

And these reports are pretty simple:

好幾百萬個的藥物不良反應通報。

it has all the diseases that the patient has,

這些報告都相當簡單：

all the drugs that they're on,

上面有病人所有疾病

and all the adverse events, or side effects, that they experience.

及所有藥物的使用狀況，

It is not all of the adverse events that are occurring in America today,

還有他們經歷過的 所有不良反應事件或副作用。

but it's hundreds and hundreds of thousands of drugs.

雖然沒有現今在美國 發生的所有不良反應事件，

So I said to Nick,

但卻有上百萬種藥物資科。

"Let's think about glucose.

所以，我跟尼克說：

Glucose is very important, and we know it's involved with diabetes.

「我們來想一想葡萄糖。

Let's see if we can understand glucose response.

葡萄糖非常重要，而且 大家都知道它與糖尿病有關。

I sent Nick off. Nick came back.

讓我們來看看是否可以 了解葡萄糖的反應。」

"Russ," he said,

我請尼克去找資料，

"I've created a classifier that can look at the side effects of a drug

他回來後說：「洛斯，

based on looking at this database,

我已經建造了一個分辨器， 可以透過這個資料庫

and can tell you whether that drug is likely to change glucose or not."

來檢視一種藥物的副作用，

He did it. It was very simple, in a way.

而且還可以告訴你，這個藥 會否改變病人血糖狀況。」

He took all the drugs that were known to change glucose

他用一個方法做到了，很簡單。

and a bunch of drugs that don't change glucose,

他把所有已知會改變葡萄糖的藥物

and said, "What's the difference in their side effects?

及所有不會改變的藥物拿出來做比較，

Differences in fatigue? In appetite? In urination habits?"

「它們之間的副作用有什麼分別？

All those things conspired to give him a really good predictor.

疲勞狀況上的差異？食慾上的差異？ 排尿習慣上的差異？」

He said, "Russ, I can predict with 93 percent accuracy

所有這些事情都可以協助他 做出一個很棒的預測器。

when a drug will change glucose."

他說：「洛斯，我能預測 哪種藥可改變血糖，

I said, "Nick, that's great."

準確率可以高達93%。」

He's a young student, you have to build his confidence.

我說：「尼克，這太棒了！」 他是個年輕的學生，

"But Nick, there's a problem.

你必須建立他的信心。

It's that every physician in the world knows all the drugs that change glucose,

「但，尼克，有一個問題。

because it's core to our practice.

就是全世界的醫師都知道 這些藥會改變葡萄糖，

So it's great, good job, but not really that interesting,

因為這是我們實務上的核心。

definitely not publishable."

所以，你很棒，幹得好， 但並沒有人對這有興趣，

(Laughter)

絕對還不適合公布你的研究結果。」

He said, "I know, Russ. I thought you might say that."

（笑聲 ）

Nick is smart.

他說：「我知道，洛斯。 我知道你可能會這麼說。」

"I thought you might say that, so I did one other experiment.

尼克很聰明。

I looked at people in this database who were on two drugs,

「我知道你會這麼說， 所以我多做了另一項實驗。

and I looked for signals similar, glucose-changing signals,

我仔細觀察資料庫裡 同時服用兩種藥的人，

for people taking two drugs,

然後尋找他們之間

where each drug alone did not change glucose,

葡萄糖改變的相似訊號，

but together I saw a strong signal."

但前提是，這些藥單獨服用 不會改變葡萄糖，

And I said, "Oh! You're clever. Good idea. Show me the list."

一起服用時，會有強烈訊號的藥物。」

And there's a bunch of drugs, not very exciting.

我說：「喔！你真聰明， 好主意，讓我看一下清單。」

But what caught my eye was, on the list there were two drugs:

有一大堆藥，並沒有令人非常興奮。

paroxetine, or Paxil, an antidepressant;

但引起我注意的是，清單上有兩種藥：

and pravastatin, or Pravachol, a cholesterol medication.

帕羅西汀或稱克憂果， 這是一種治療憂鬱症的藥，

And I said, "Huh. There are millions of Americans on those two drugs."

還有普伐他汀或稱美百樂， 一種治療心臟疾病的藥。

In fact, we learned later,

然後我說：「哈！有上百萬 美國人正在服用這兩種藥」。

15 million Americans on paroxetine at the time, 15 million on pravastatin,

事實上，我們之後才知道，

and a million, we estimated, on both.

當時有1500萬美國人正在服用帕羅西汀， 1500萬人正在服用普伐他汀，

So that's a million people

而我們預估有100萬人， 同時服用這兩個藥。

who might be having some problems with their glucose

所以，有100萬人

if this machine-learning mumbo jumbo that he did in the FDA database

可能有葡萄糖上的問題，

actually holds up.

如果他用食品藥物管理局的資料庫

But I said, "It's still not publishable,

做的機械學習判讀器真的有用的話。

because I love what you did with the mumbo jumbo,

但我說：「還是不能發表，

with the machine learning,

因為我雖然喜歡你做的

but it's not really standard-of-proof evidence that we have."

機械學習判讀器，

So we have to do something else.

但我們沒有真正的證明標準 來證明我們是正確的。」

Let's go into the Stanford electronic medical record.

所以，我們來必須做些其他事來驗證。

We have a copy of it that's OK for research,

我們去找史丹佛的電子病例紀錄。

we removed identifying information.

我們有一個副本，可以用來研究，

And I said, "Let's see if people on these two drugs

我們移除了病人個資。

have problems with their glucose."

我說：「讓我們來看看， 服用這兩種藥的人

Now there are thousands and thousands of people

是否有葡萄糖上的疾病。」

in the Stanford medical records that take paroxetine and pravastatin.

在史丹佛病例紀錄中有成千上萬的人

But we needed special patients.

同時服用這兩種藥。

We needed patients who were on one of them and had a glucose measurement,

但我們需要特定病患。

then got the second one and had another glucose measurement,

我們需要已經做葡萄糖檢測 且服用其中一種藥的病人，

all within a reasonable period of time -- something like two months.

另外再找到另一個已經做過 另一個葡萄糖檢測的病人，

And when we did that, we found 10 patients.

全部都在合理期間做的， 例如兩個月內。

However, eight out of the 10 had a bump in their glucose

當我們開始著手進行時， 我們找到十個病人。

when they got the second P -- we call this P and P --

然而，十個人裡面 有八個葡萄糖異常增加現象，

when they got the second P.

在他們服用第二個P時 ─我們稱呼這個叫 P&P─

Either one could be first, the second one comes up,

當他們服用了第二個 P。

glucose went up 20 milligrams per deciliter.

哪一個先服用都行， 當第二個藥服用後，

Just as a reminder,

葡萄糖濃度每公升會增加20毫克。

you walk around normally, if you're not diabetic,

提醒各位一下，

with a glucose of around 90.

如果你能正常走動，沒有糖尿病，

And if it gets up to 120, 125,

你的葡萄糖濃度約90毫克／公升。

your doctor begins to think about a potential diagnosis of diabetes.

如果上升到120、125，

So a 20 bump -- pretty significant.

你的醫生會開始認為 你有潛在的糖尿病症狀。

I said, "Nick, this is very cool.

所以，一下子增加20是相當明顯的。

But, I'm sorry, we still don't have a paper,

我說：「尼克，這很酷。

because this is 10 patients and -- give me a break --

但，很抱歉，我們仍然沒辦法寫報告，

it's not enough patients."

因為只有十個病人，饒了我吧，

So we said, what can we do?

病人樣本數根本不夠。」

And we said, let's call our friends at Harvard and Vanderbilt,

所以，那怎麼辦？

who also -- Harvard in Boston, Vanderbilt in Nashville,

我們來打電話給哈佛 及范德堡大學的朋友，

who also have electronic medical records similar to ours.

就是波士頓的哈佛 及納許維爾的范德堡，

Let's see if they can find similar patients

他們都有跟我們很像的 電子病歷紀錄。

with the one P, the other P, the glucose measurements

讓我們看看，他們是否 也可以找到相同的病人，

in that range that we need.

也有我們需要的已經服用這兩種藥，

God bless them, Vanderbilt in one week found 40 such patients,

並做過葡萄糖檢測的病人。

same trend.

上天保佑，范德堡一個星期內找到40個

Harvard found 100 patients, same trend.

有同樣趨勢的病人。

So at the end, we had 150 patients from three diverse medical centers

哈佛找到100個有同樣趨勢的病人。

that were telling us that patients getting these two drugs

所以，最後，我們從三個不同的 醫學中心找到150個病人

were having their glucose bump somewhat significantly.

服用過這兩種藥，

More interestingly, we had left out diabetics,

然後有葡萄糖異常增加現象。

because diabetics already have messed up glucose.

有趣的是，我們沒有考慮糖尿病患者，

When we looked at the glucose of diabetics,

因為糖尿病患者本身的 血糖濃度就已經很混亂。

it was going up 60 milligrams per deciliter, not just 20.

當我們觀察糖尿病患者的血糖濃度時，

This was a big deal, and we said, "We've got to publish this."

會上升到每公升60毫克， 不只20毫克。

We submitted the paper.

這事情很重要，我們說： 「我們必須發佈這件事。」

It was all data evidence,

我們遞交報告，

data from the FDA, data from Stanford,

裡面全部都是資料證明，

data from Vanderbilt, data from Harvard.

有來自食品藥物管理局、史丹佛的資料、

We had not done a single real experiment.

有來自范德堡、哈佛醫學院的資料，

But we were nervous.

我們完全沒有做任何實驗。

So Nick, while the paper was in review, went to the lab.

但我們很緊張。

We found somebody who knew about lab stuff.

所以，當報告送去審核時， 尼克就去了實驗室。

I don't do that.

我們找到會做實驗的人。

I take care of patients, but I don't do pipettes.

我不做實驗的。

They taught us how to feed mice drugs.

我會看病人，但我不會做分量管。

We took mice and we gave them one P, paroxetine.

他們教我們如何餵老鼠吃藥。

We gave some other mice pravastatin.

我們給第一組老鼠餵食帕羅西汀，

And we gave a third group of mice both of them.

給第二組老鼠餵食普伐他汀。

And lo and behold, glucose went up 20 to 60 milligrams per deciliter

第三組的老鼠兩種藥都餵食。

in the mice.

驚奇的是，葡萄糖每公升上升20到60毫克，

So the paper was accepted based on the informatics evidence alone,

老鼠也有相同的反應。

but we added a little note at the end,

所以，只有資料證據的報告被接受了，

saying, oh by the way, if you give these to mice, it goes up.

但我們在最後加了註記說，

That was great, and the story could have ended there.

如果把藥物給老鼠，葡萄糖也會上升。

But I still have six and a half minutes.

太棒了，故事其實就到這裡結束。

(Laughter)

但，我還有六分半鐘。

So we were sitting around thinking about all of this,

（笑聲）

and I don't remember who thought of it, but somebody said,

所以，我們坐下來想一下所有的事，

"I wonder if patients who are taking these two drugs

我忘記誰曾經說過，但有人說：

are noticing side effects of hyperglycemia.

「不曉得同時服用這兩種藥的病人，

They could and they should.

是否有注意到高血糖症的副作用。

How would we ever determine that?"

他們可能知道，也必須知道。

We said, well, what do you do?

我們要如何確定？」

You're taking a medication, one new medication or two,

我們說，好吧，你會怎麼做？

and you get a funny feeling.

你服用了一種藥，一個或兩個新藥，

What do you do?

然後你感覺怪怪的。

You go to Google

你會怎麼做？

and type in the two drugs you're taking or the one drug you're taking,

你會去問 Google，

and you type in "side effects."

然後搜尋你在服用的一或兩個藥名，

What are you experiencing?

然後加上「副作用」。

So we said OK,

你會找到什麼？

let's ask Google if they will share their search logs with us,

所以，我們說，好，

so that we can look at the search logs

我們來問 Google 能否 跟我們分享搜尋紀錄，

and see if patients are doing these kinds of searches.

讓我們可以觀察搜尋紀錄，

Google, I am sorry to say, denied our request.

看是否有病人也在做同樣的搜尋。

So I was bummed.

很抱歉我得這麼說， 但 Google 拒絕了我們的請求。

I was at a dinner with a colleague who works at Microsoft Research

所以，我很煩惱。

and I said, "We wanted to do this study,

我跟一個在微軟研究室的同事吃晚餐時，

Google said no, it's kind of a bummer."

我跟他說：「我們想做這個研究，

He said, "Well, we have the Bing searches."

Google 說不行，我有點煩惱。」

(Laughter)

他說：「我們有 Bing 搜尋引擎啊。」

Yeah.

（笑聲）

That's great.

是啊！

Now I felt like I was --

太棒了。

(Laughter)

現在，我感覺...

I felt like I was talking to Nick again.

（笑聲）

He works for one of the largest companies in the world,

我好像又在鼓勵尼克一樣。

and I'm already trying to make him feel better.

他在全世界數一數二的公司上班，

But he said, "No, Russ -- you might not understand.

我已經開始要安慰他了。

We not only have Bing searches,

但他說：「不，洛斯，你可能沒搞懂。

but if you use Internet Explorer to do searches at Google,

我們不只有 Bing 啊，

Yahoo, Bing, any ...

如果你用 IE 在 Google、

Then, for 18 months, we keep that data for research purposes only."

雅虎、Bing 等任何搜尋引擎，

I said, "Now you're talking!"

之後18個月，我們保留這些數據 僅做研究目的使用。

This was Eric Horvitz, my friend at Microsoft.

我說：「這才像話嘛！」

So we did a study

這就是我的微軟朋友艾瑞克．霍維茲。

where we defined 50 words that a regular person might type in

我們做了一項研究，

if they're having hyperglycemia,

我們定義出了50個

like "fatigue," "loss of appetite," "urinating a lot," "peeing a lot" --

如果一般人有高血糖症時 會鍵入的關鍵字，

forgive me, but that's one of the things you might type in.

像是疲勞、沒食慾、頻尿等。

So we had 50 phrases that we called the "diabetes words."

請原諒我，但這些就是 你可能會鍵入的關鍵字。

And we did first a baseline.

所以，我們有了50個短語， 我們稱之為「糖尿病關鍵字」。

And it turns out that about .5 to one percent

我們先設定了一條基準線。

of all searches on the Internet involve one of those words.

原來，網路上有包含這些關鍵字的搜尋

So that's our baseline rate.

占了0.5~1%的比例。

If people type in "paroxetine" or "Paxil" -- those are synonyms --

所以，這就是我們的基準線率，

and one of those words,

如果大家鍵入「帕羅西汀」或「克憂果」 ──這些是同義字──

the rate goes up to about two percent of diabetes-type words,

以及剛剛其中一個關鍵字，

if you already know that there's that "paroxetine" word.

那糖尿病類型的基準線率會上升到2%，

If it's "pravastatin," the rate goes up to about three percent from the baseline.

如果你已經知道 「帕羅西汀」這個字的話。

If both "paroxetine" and "pravastatin" are present in the query,