Full text of Stanford big data techno-optimist Russ Altman’s talk titled “What Really Happens When You Mix Medications?” at TED conference. In this fascinating talk, Russ Altman shows how doctors are studying unexpected drug interactions using a surprising resource: search engine queries.
Listen to the MP3 Audio here:
Russ Altman – Professor of Bioengineering, Genetics, Medicine and Computer Science at Stanford University
So you go to the doctor and get some tests. The doctor determines that you have high cholesterol and you would benefit from medication to treat it. So you get a pillbox.
You have some confidence, your physician has some confidence that this is going to work. The company that invented it did a lot of studies, submitted it to the FDA. They studied it very carefully, skeptically, they approved it. They have a rough idea of how it works, they have a rough idea of what the side effects are. It should be okay.
You have a little more of a conversation with your physician, and the physician is a little worried because you’ve been blue, you haven’t felt like yourself, you haven’t been able to enjoy things in life quite as much as you usually do.
Your physician says, “You know, I think you have some depression. I’m going to have to give you another pill.”
So now we’re talking about two medications. This pill also — millions of people have taken it, the company did studies, the FDA looked at it — all good. Think things should go okay.
Think things should go okay.
Well, wait a minute.
How much have we studied these two together?
Well, it’s very hard to do that. In fact, it’s not traditionally done. We totally depend on what we call post-marketing surveillance, after the drugs hit the market.
How can we figure out if bad things are happening between two medications? Three? Five? Seven? Ask your favorite person who has several diagnoses how many medications they’re on.
So why do I care about this problem?
I care about it deeply. I’m an informatics and data science guy and really, in my opinion, the only hope — only hope — to understand these interactions is to leverage lots of different sources of data in order to figure out when drugs can be used together safely and when it’s not so safe.
So let me tell you a data science story. And it begins with my student Nick. Let’s call him Nick, because that’s his name.
Nick was a young student. And I said, “You know, Nick, we have to understand how drugs work and how they work together and how they work separately, and we don’t have a great understanding.
But the FDA has made available an amazing database. It’s a database of adverse events. They literally put on the web — publicly available, you could all download it right now — hundreds of thousands of adverse event reports from patients, doctors, companies, pharmacists.
And these reports are pretty simple: it has all the diseases that the patient has, all the drugs that they’re on, and all the adverse events, or side effects, that they experience. It is not all of the adverse events that are occurring in America today, but it’s hundreds and hundreds of thousands of drugs.
So I said to Nick, “Let’s think about glucose. Glucose is very important, and we know it’s involved with diabetes. Let’s see if we can understand glucose response.”
I sent Nick off. Nick came back. “Russ,” he said, “I’ve created a classifier that can look at the side effects of a drug based on looking at this database, and can tell you whether that drug is likely to change glucose or not.” He did it. It was very simple, in a way. He took all the drugs that were known to change glucose and he took a bunch of drugs that don’t change glucose, and he said, “What’s the difference in their side effects? Differences in fatigue? In appetite? In urination habits?”
All those things conspired to give him a really good predictor. He said, “Russ, I can predict with 93% accuracy when a drug will change glucose.”
I said, “Nick, that’s great.” He’s a young student, you have to build his confidence. “But Nick, there’s a problem. It’s that every physician in the world knows all the drugs that change glucose, because it’s core to our practice. So it’s great, good job, but not really that interesting, definitely not publishable.”
He said, “I know, Russ. I thought you might say that.” Nick is smart. “I thought you might say that, so I did one other experiment. I looked at people in this database who were on two drugs, and I looked for signals similar, glucose-changing signals, for people taking two drugs, where each drug alone did not change glucose, but together I saw a strong signal.”
And I said, “Oh! You’re clever. Good idea. Show me the list.”
And there’s a bunch of drugs, not very exciting. But what caught my eye was, on the list there were two drugs: paroxetine, or Paxil, an antidepressant; and pravastatin, or Pravachol, a cholesterol medication.
And I said, “Huh. There are millions of Americans on those two drugs.” In fact, we learned later, 15 million Americans on paroxetine at the time, 15 million on pravastatin, and a million, we estimated, on both.
So that’s a million people who might be having some problems with their glucose if this machine-learning mumbo jumbo that he did in the FDA database actually holds up.
But I said, “Nick, it’s still not publishable, because I love what you did with the mumbo jumbo, with the machine learning, but it’s not really standard-of-proof evidence that we have.”
So we have to do something else. Let’s go into the Stanford electronic medical record. We have a copy of it that’s OK for research, we removed identifying information. And I said, “Let’s see if people on these two drugs have problems with their glucose.”
Now there are thousands and thousands of people in the Stanford medical records that take paroxetine and pravastatin. But we needed special patients. We needed patients who were on one of them and had a glucose measurement, then got the second one and had another glucose measurement, all within a reasonable period of time — something like two months.
And when we did that, we found 10 patients. However, eight out of the 10 had a bump in their glucose when they got the second P — we call this P and P — when they got the second P. Either one could be first, the second one comes up, glucose went up 20 milligrams per deciliter.
Just as a reminder, you walk around normally, if you’re not diabetic, with a glucose of around 90. And if it gets up to 120, 125, your doctor begins to think about a potential diagnosis of diabetes. So a 20 bump — pretty significant.
I said, “Nick, this is very cool. But, I’m sorry, we still don’t have a paper, because this is 10 patients and — give me a break — it’s not enough patients.”
So we said, what can we do? And we said, let’s call our friends at Harvard and Vanderbilt, who also — Harvard in Boston, Vanderbilt in Nashville, who also have electronic medical records similar to ours. Let’s see if they can find similar patients with the one P, the other P, the glucose measurements in that range that we need.
God bless them, Vanderbilt in one week found 40 such patients, same trend. Harvard found 100 patients, same trend. So at the end, we had 150 patients from three diverse medical centers that were telling us that patients getting these two drugs were having their glucose bump somewhat significantly.
More interestingly, we had left out diabetics, because diabetics already have messed up glucose. When we looked at the glucose of diabetics, it was going up 60 milligrams per deciliter, not just 20. This was a big deal, and we said, “We’ve got to publish this.”
We submitted the paper. It was all data evidence, data from the FDA, data from Stanford, data from Vanderbilt, data from Harvard. We had not done a single real experiment. But we were nervous.
So Nick, while the paper was in review, went to the lab. We found somebody who knew about lab stuff. I don’t do that. I take care of patients, but I don’t do pipettes. They taught us how to feed mice drugs. We took mice and we gave them one P, paroxetine. We gave some other mice pravastatin. And we gave a third group of mice both of them.
And lo and behold, glucose went up 20 to 60 milligrams per deciliter in the mice. So the paper was accepted based on the informatics evidence alone, but we added a little note at the end, saying, oh by the way, if you give these to mice, it goes up.
That was great, and the story could have ended there. But I still have six and a half minutes.
So we were sitting around thinking about all of this, and I don’t remember who thought of it, but somebody said, “I wonder if patients who are taking these two drugs are noticing side effects of hyperglycemia. They could and they should. How would we ever determine that?”
We said, well, what do you do? You’re taking a medication, one new medication or two, and you get a funny feeling. What are you going to do? You’re going to go to Google and you’re going to type in the two drugs you’re taking or the one drug you’re taking, and you type in “side effects.” What are you experiencing?
So we said OK, let’s ask Google if they will share their search logs with us, so that we can look at the search logs and see if patients are doing these kinds of searches. Google, I am sorry to say, denied our request. So I was bummed.
And I was at a dinner with a colleague who works at Microsoft Research and I said, “Hey, we wanted to do this study, Google said no, it’s kind of a bummer.”
He said, “Well, we have the Bing searches.”
Yeah. That’s great.
Now I felt like I was — I felt like I was talking to Nick again. He works for one of the largest companies in the world, and I’m already trying to make him feel better.
But he said, “No, Russ – he said, you might not understand. We not only have Bing searches, but if you use Internet Explorer to do searches at Google, Yahoo, Bing, any … Then, for 18 months, we keep that data for research purposes only.”
I said, “Now you’re talking!”
This was Eric Horvitz, my friend at Microsoft.
So we did a study where we defined 50 words that a regular person might type in if they’re having hyperglycemia, like “fatigue,” “loss of appetite,” “urinating a lot,” “peeing a lot” — forgive me, but that’s one of the things you might type in.
So we had 50 phrases that we called the “diabetes words.” And we did first a baseline. And it turns out that about 0.5% to 1% of all searches on the Internet involve one of those words. So that’s our baseline rate.
If people type in “paroxetine” or “Paxil” — those are synonyms — and one of those words, the rate goes up to about 2% of diabetes-type words, if you already know that there’s that “paroxetine” word.
If it’s “pravastatin,” the rate goes up to about 3% from the baseline. If both “paroxetine” and “pravastatin” are present in the query, it goes up to 10%, a huge three- to four-fold increase in those searches with the two drugs that we were interested in, and kind of diabetes-type words or hyperglycemia-type words.
We published this, and it got some attention. The reason it deserves attention is that patients are telling us their side effects indirectly through their searches. We brought this to the attention of the FDA. They were interested. They have set up social media surveillance programs to collaborate with Microsoft, which had a nice infrastructure for doing this, and others, to look at Twitter feeds, to look at Facebook feeds, to look at search logs, to try to see early signs that drugs, either individually or together, are causing problems.
So what do I take from this? Why do I tell this story?
Well, first of all, we have now the promise of big data and medium-sized data to help us understand drug interactions and really, fundamentally, drug actions. How do drugs work? This will create and has created a new ecosystem for understanding how drugs work and to optimize their use.
Nick went on; he’s a professor at Columbia now. He did this in his PhD for hundreds of pairs of drugs. He found several very important interactions, and so we replicated this and we showed that this is a way that really works for finding drug-drug interactions.
However, there’s a couple of things.
We don’t just use pairs of drugs at a time. As I said before, there are patients on three, five, seven, nine drugs.
Have they been studied with respect to their nine-way interaction? Yes, we can do pair-wise, A and B, A and C, A and D. But what about A, B, C, D, E, F, G all together, being taken by the same patient, perhaps interacting with each other in ways that either makes them more effective or less effective or causes side effects that are unexpected?
We really have no idea. It’s a blue sky, open field for us to use data to try to understand the interaction of drugs.
Two more lessons:
I want you to think about the power that we were able to generate with the data from people who had volunteered their adverse reactions through their pharmacists, through themselves, through their doctors, the people who allowed the databases at Stanford, Harvard, Vanderbilt, to be used for research.
People are worried about data. They’re worried about their privacy and security — they should be. We need secure systems. But we can’t have a system that closes that data off, because it is too rich of a source of inspiration, innovation and discovery for new things in medicine.
And then the final thing I want to say is, in this case we found two drugs and it was a little bit of a sad story. The two drugs actually caused problems. They increased glucose. They could throw somebody into diabetes who would otherwise not be in diabetes, and so you would want to use the two drugs very carefully together, perhaps not together, make different choices when you’re prescribing.
But there was another possibility. We could have found two drugs or three drugs that were interacting in a beneficial way. We could have found new effects of drugs that neither of them has alone, but together, instead of causing a side effect, they could be a new and novel treatment for diseases that don’t have treatments or where the treatments are not effective.
If we think about drug treatment today, all the major breakthroughs — for HIV, for tuberculosis, for depression, for diabetes — it’s always a cocktail of drugs.
And so the upside here, and the subject for a different TED Talk on a different day, is how can we use the same data sources to find good effects of drugs in combination that will provide us new treatments, new insights into how drugs work and enable us to take care of our patients even better?
Thank you very much.
Resources for Further Reading:
- How Literature Can Help Us Develop Empathy: Beth Ann Fennelly (Transcript)
- The Human Skills We Need In An Unpredictable World: Margaret Heffernan (Transcript)
- How AI Could Empower Any Business: Andrew Ng (Transcript)
- Agoraphobia: The Fear of Fear – Linda Bussey (Transcript)
- Why Politicians Shouldn’t Appoint Judges: Aziz Huq (Transcript)