An AI smartwatch that detects seizures

This is Henry, a cute boy, and when Henry was three, his mom found him having some febrile seizures(发热痉挛). Febrile seizures are seizures that occur when you also have a fever, and the doctor said, “Don’t worry too much. Kids usually outgrow (when they grow up, then can get rid of some sickness) these.” When he was four, he had a convulsive(抽搐,震动) seizure, the kind that you lose consciousness and shake — a generalized tonic-clonic(强直阵挛) seizure — and while the diagnosis(诊断) of epilepsy was in the mail, Henry’s mom went to get him out of bed one morning, and as she went in his room, she found his cold, lifeless body.

Henry died of SUDEP, sudden unexpected death in epilepsy. I’m curious how many of you have heard of SUDEP. This is a very well-educated audience, and I see only a few hands. SUDEP is when an otherwise(原本) healthy person with epilepsy dies and they can’t attribute it to(归咎于) anything they can find in an autopsy(验尸). There is a SUDEP every seven to nine minutes. That’s on average two per TED Talk. Now, a normal brain has electrical activity. You can see some of the electrical waves coming out of this picture of a brain here. And these should look like typical electrical activity that an EEG(脑电图分析仪) could read on the surface. When you have a seizure, it’s a bit of unusual electrical activity, and it can be focal(焦点,病灶). It can take place in just a small part of your brain. When that happens, you might have a strange sensation(feeling). Several could be happening here in the audience right now, and the person next to you might not even know. However, if you have a seizure where that little brush fire spreads like a forest fire over the brain, then it generalizes(form the concept of something), and that generalized seizure takes your consciousness away and causes you to convulse.(这种常规意义上的癫痫把你的意识并且引起你的抽搐)

There are more SUDEPs in the United States every year than sudden infant death syndrome(综合征). Now, how many of you have heard of sudden infant death syndrome? Right? Pretty much every hand goes up. So what’s going on here? Why is this so much more common and yet people haven’t heard of it? And what can you do to prevent it? Well, there are two things, scientifically shown, that prevent or reduce the risk of SUDEP. The first is: “Follow your doctor’s instructions, take your medications.” Two-thirds of people who have epilepsy get it under control with their medications. The second thing that reduces the risk of SUDEP is companionship(陪伴). It’s having somebody there at the time that you have a seizure. Now, SUDEP, even though most of you have never heard of it, is actually the number two cause of years of potential life lost of all neurological(神经学) disorders. The vertical axis is the number of deaths times the remaining life span, so higher is much worse impact. SUDEP, however, unlike these others, is something that people right here could do something to push that down.

Now, what is Roz Picard, an AI researcher, doing here telling you about SUDEP, right? I’m not a neurologist. When I was working at the Media Lab on measurement of emotion, trying to make our machines more intelligent about our emotions, we started doing a lot of work measuring stress. We built lots of sensors that measured it in lots of different ways. But one of them in particular grew out of some of this very old work with measuring sweaty palms(手掌心出汗) with an electrical signal. This is a signal of skin conductance(导率) that’s known to go up when you get nervous, but it turns out it also goes up with a lot of other interesting conditions. But measuring it with wires on your hand is really inconvenient. So we invented a bunch of other ways of doing this at the MIT Media Lab. And with these wearables, we started to collect the first-ever clinical(临床的) quality data 24-7. Here’s a picture of what that looked like the first time an MIT student collected skin conductance on the wrist 24-7. Let’s zoom in a little bit here. What you see is 24 hours from left to right, and here is two days of data. And first, what surprised us was sleep was the biggest peak of the day. Now, that sounds broken, right? You’re calm when you’re asleep, so what’s going on here?

Well, it turns out that our physiology during sleep is very different than our physiology during wake, and while there’s still a bit of a mystery why these peaks are usually the biggest of the day during sleep, we now believe they’re related to memory consolidation(巩固) and memory formation(记忆合成) during sleep.

We also saw things that were exactly what we expected. When an MIT student is working hard in the lab or on homeworks, there is not only emotional stress, but there’s cognitive(认知) load, and it turns out that cognitive load, cognitive effort, mental engagement, excitement about learning something — those things also make the signal go up.

Unfortunately, to the embarrassment of we MIT professors,


the low point every day is classroom activity. Now, I am just showing you one person’s data here, but this, unfortunately, is true in general.

This sweatband(防汗带) has inside it a homebuilt skin-conductance sensor, and one day, one of our undergrads knocked on my door right at the end of the December semester, and he said, “Professor Picard, can I please borrow one of your wristband sensors? My little brother has autism(自闭症), he can’t talk, and I want to see what’s stressing him out.” And I said, “Sure, in fact, don’t just take one, take two,” because they broke easily back then. So he took them home, he put them on his little brother. Now, I was back in MIT, looking at the data on my laptop, and the first day, I thought, “Hmm, that’s odd, he put them on both wrists instead of waiting for one to break. OK, fine, don’t follow my instructions.” I’m glad he didn’t. Second day — chill. Looked like classroom activity.


A few more days ahead. The next day, one wrist signal was flat and the other had the biggest peak I’ve ever seen, and I thought, “What’s going on? We’ve stressed people out at MIT every way imaginable. I’ve never seen a peak this big.” And it was only on one side. How can you be stressed on one side of your body and not the other? So I thought one or both sensors must be broken. Now, I’m an electroengineer by training, so I started a whole bunch of stuff to try to debug this, and long story short, I could not reproduce this.

So I resorted to old-fashioned debugging. I called the student at home on vacation. “Hi, how’s your little brother? How’s your Christmas? Hey, do you have any idea what happened to him?” And I gave this particular date and time, and the data. And he said, “I don’t know, I’ll check the diary.”

Diary? An MIT student keeps a diary? So I waited and he came back. He had the exact date and time, and he says, “That was right before he had a grand mal seizure.” (在大的抽搐之前发生的)

Now, at the time, I didn’t know anything about epilepsy, and did a bunch of research, realized that another student’s dad is chief of neurosurgery at Children’s Hospital Boston, screwed up my courage and called Dr. Joe Madsen.

“Hi, Dr. Madsen, my name’s Rosalind Picard. Is it possible somebody could have a huge sympathetic(交感神经) nervous system surge” — that’s what drives the skin conductance — “20 minutes before a seizure?” And he says, “Probably not.” He says, “It’s interesting. We’ve had people whose hair stands on end on one arm 20 minutes before a seizure.” And I’m like, “On one arm?” I didn’t want to tell him that, initially, because I thought this was too ridiculous(可笑).

He explained how this could happen in the brain, and he got interested. I showed him the data. We made a whole bunch more devices, got them safety certified. 90 families were being enrolled in a study, all with children who were going to be monitored 24-7 with gold-standard EEG on their scalp(头皮) for reading the brain activity, video to watch the behavior, electrocardiogram(心电图) — ECG — and now EDA, electrodermal activity, to see if there was something in this periphery(边缘) that we could easily pick up, related to a seizure.

We found, in 100 percent of the first batch of grand mal seizures, this whopper of responses in the skin conductance. The blue in the middle, the boy’s sleep, is usually the biggest peak of the day. These three seizures you see here are popping out of the forest like redwood trees. Furthermore, when you couple the skin conductance at the top with the movement from the wrist and you get lots of data and train machine learning and AI on it, you can build an automated AI that detects these patterns much better than just a shake detector can do. So we realized that we needed to get this out, and with the PhD work of Ming-Zher Poh and later great improvements by Empatica, this has made progress and the seizure detection is much more accurate.


But we also learned some other things about SUDEP during this. One thing we learned is that SUDEP, while it’s rare after a generalized tonic-clonic seizure, that’s when it’s most likely to happen — after that type. And when it happens, it doesn’t happen during the seizure, and it doesn’t usually happen immediately afterwards, but immediately afterwards, when the person just seems very still and quiet, they may go into another phase, where the breathing stops, and then after the breathing stops, later the heart stops. So there’s some time to get somebody there. We also learned that there is a region deep in the brain called the amygdala(杏仁核), which we had been studying in our emotion research a lot. We have two amygdalas, and if you stimulate(刺激) the right one, you get a big right skin conductance response. Now, you have to sign up right now for a craniotomy(穿颅术) to get this done, not exactly something we’re going to volunteer to do, but it causes a big right skin conductance response. Stimulate the left one, big left skin conductance response on the palm. And furthermore, when somebody stimulates your amygdala while you’re sitting there and you might just be working, you don’t show any signs of distress, but you stop breathing, and you don’t start again until somebody stimulates you. “Hey, Roz, are you there?” And you open your mouth to talk. As you take that breath to speak, you start breathing again.

So we had started with work on stress, which had enabled us to build lots of sensors that were gathering high quality enough data that we could leave the lab and start to get this in the wild; accidentally found a whopper(谎言) of a response with the seizure, neurological activation that can cause a much bigger response than traditional stressors; lots of partnership with hospitals and an epilepsy monitoring unit, especially Children’s Hospital Boston and the Brigham; and machine learning and AI on top of this to take and collect lots more data in service of trying to understand these events and if we could prevent SUDEP.

This is now commercialized by Empatica, a start-up that I had the privilege to cofound, and the team there has done an amazing job improving the technology to make a very beautiful sensor that not only tells time and does steps and sleep and all that good stuff, but this is running real-time AI and machine learning to detect generalized tonic-clonic seizures and send an alert for help if I were to have a seizure and lose consciousness. This just got FDA-approved as the first smartwatch to get approved in neurology.


Now, the next slide is what made my skin conductance(电导率) go up. One morning, I’m checking my email and I see a story from a mom who said she was in the shower, and her phone was on the counter by the shower, and it said her daughter might need her help. So she interrupts her shower and goes running to her daughter’s bedroom, and she finds her daughter facedown in bed, blue and not breathing. She flips her over(把她翻过来) — human stimulation(刺激) — and her daughter takes a breath, and another breath, and her daughter turns pink and is fine.

I think I turned white reading this email. My first response is, “Oh no, it’s not perfect. The Bluetooth could break, the battery could die. All these things could go wrong. Don’t rely on this.” And she said, “It’s OK. I know no technology is perfect. None of us can always be there all the time. But this, this device plus AI enabled me to get there in time to save my daughter’s life.”

Now, I’ve been mentioning children, but SUDEP peaks, actually, among people in their 20s, 30s and 40s, and the next line I’m going to put up is probably going to make some people uncomfortable, but it’s less uncomfortable than we’ll all be if this list is extended to somebody you know. Could this happen to somebody you know? And the reason I bring up this uncomfortable question is because one in 26 of you will have epilepsy at some point, and from what I’ve been learning, people with epilepsy often don’t tell their friends and their neighbors that they have it. So if you’re willing to let them use an AI or whatever to summon(recover/help sb to recover) you in a moment of possible need, if you would let them know that, you could make a difference in their life.

Why do all this hard work to build AIs? A couple of reasons here: one is Natasha, the girl who lived, and her family wanted me to tell you her name. Another is her family and the wonderful people out there who want to be there to support people who have conditions that they’ve felt uncomfortable in the past mentioning to others. And the other reason is all of you, because we have the opportunity to shape the future of AI. We can actually change it, because we are the ones building it.

So let’s build AI that makes everybody’s lives better.

Thank you.


This entry was posted in TED 演讲. Bookmark the permalink. Follow any comments here with the RSS feed for this post. Both comments and trackbacks are currently closed.

1 Comment

  1. cme
    Posted 2019年5月16日 at 上午10:48 | Permalink