13 ms·

In our eyes, Google’s software sees heart attack risk

52 points by helloworld 9y ago

carbocation 9y ago
As my PI notes, a lot of the success here can be explained by the model's excellent ability to determine age from the images of the retina. https://twitter.com/skathire/status/965661596503638016 https://twitter.com/skathire/status/965661596503638016 This is quite good age detection - other efforts that I've seen (from face images, rather than retina images) are +/- 5 years, whereas this paper gets age to within +/- 3 years or so. The fact that the model automatically detects anatomical features that humans recognize as being important is promising for a variety of applications, I think.
logronoide 9y ago
Any optician can do exactly the same thing during a simple check of your eyes. In this case I can’t see the benefit of this type of screening. The tool used by the optician is just a cheap lantern, but only the tool needed by the computer is way more expensive.
- ridgeguy 9y ago
  Perhaps a cheap lens/illuminator attachment for smartphones could be engineered for retinal imaging. If this method could be adapted to phones, many more people could be checked at low cost.
  - logronoide 9y ago
    It already exists. Most opticians take a image of your retina and store it in the cloud for further analysis and tracking the evolution. In Spain (where I live) most opticians do not charge anything for this test, because it’s a minimal part of a typical prescription.
    - froindt 9y ago
      That's like a $40 upcharge at my optometrist (USA). But she is a good family friend, walks me through everything she's seeing on there, and gives me glowing compliments on how beautiful my optical nerve is. It sounds stupid, but I'm going to miss that when I go to my new optometrist.
    - cortesoft 9y ago
      Yeah, but I have great eyesight and haven't been to an optician in 15 years. However, I might still be at risk of a heart attack.
      - logronoide 9y ago
        You don’t need to visit an optician. A family doctor with a lantern can do it (and normally they do). It’s a very simple test. It’s not a ML or AI challenge at all.
        mkagenius 9y ago
        But if you have AI, you do not need the optician/doctor. For example people in Poor countries have no access to such a person.
        JustSomeNobody 9y ago
        I don't think anyone would say don't build the AI/ML. Just that the technology already exists. Until then, people should continue to see their eye Dr when one is available to them.
        txcwpalpha 9y ago
        There are a few startups/companies that specialize in going to poor countries and using lens attachments on cell phones to do eye exams. They send in people who aren't doctors, take a picture of the eye, and email/upload the pictures to doctors who can analyze it for issues. IIRC, in those situations, the problem isn't the accessibility to a doctor, but rather the accessibility to someone who has a device capable of imaging the eye. I imagine this would be the same. Replacing the doctor with an AI doesn't really solve the problem.
- ceejayoz 9y ago
  I'd suspect the tools will eventually be dramatically cheaper than a trained optician. Put it in a pharmacy, grocery store, or other central location and people could screen weekly instead of once every couple years (assuming they even see an optician).
  - logronoide 9y ago
    As I have explained before, in Spain an optician does not charge anything for it because it’s ridiculously simple and it’s a small percentage of the effort that an optician needs to do for a prescription. This test is different of a blood test for example. You don’t need to control the evolution weekly, or even monthly. This is one of the most simple screening process. Because it’s easy to train the model with tons of images already made by opticians and optometrists. It does not have any value, it’s too simple. My brother is an optician and has a huge database with retina images. If even me with limited knowledge of ML and my brother as field expert could train a model to detect blood pressure problems. Too simple, no value.
    - ceejayoz 9y ago
      You're missing a very key point. Many people with normal eyesight never see an optician. They might use a kiosk at the grocery store. It's also a first step proof-of-concept towards harder diagnoses.
      - logronoide 9y ago
        As I said, not only opticians can do it. A family doctor with a lantern can do it (and they do). It’s not an ML or AI challenge at all.
        ceejayoz 9y ago
        There are plenty of Americans who don't have access to a "family doctor". Not everyone can afford or bothers to get an annual checkup. Same reason grocery stores often have a "check your blood pressure" kiosk, or offer flu shots. It's quick, it's convenient, and you reach people who would otherwise miss out. Sure, a doctor or nurse can check your BP or give you the vaccine, but there's a significant benefit to expanded access.
- elbasti 9y ago
  Well, not really. The tool used by the optician is a $5 dollar lantern connected to a brain that had to go through a $300,000 dollar education.
matte_black 9y ago
What do you do after you discover you are at risk? Get your affairs in order?
- ceejayoz 9y ago
  Adjust diet, exercise, statins, get a cardiology consult?
  - matte_black 9y ago
    But if you’re already at the point where you are at risk and your eyes show it, isn’t it already too late to really do anything about it?
    - maxerickson 9y ago
      Weight loss and exercise are recommended to people that survive heart attacks.
    - arkades 9y ago
      Morbidity and mortality can be reduced with the prescription of beta blockers, aspirin, statins, and ace inhibitors. +\- clopidogrel. Being at risk isn’t fate: you have modifiable risk factors that can be mitigated (eg, weight, cholesterol, blood pressure.) Additionally, CPR training of your immediate family can be - literally - lifesaving in the event you do have that heart attack.
- logronoide 9y ago
  Visit your family doctor ASAP.
- tinymollusk 9y ago
  It gives a high-risk group the ability to notify their friends and family and practice what to do in event of a heart attack. My mom frequently stays with me, but outside of "call 911" I don't really know what to do in teh event of a heart attack. Time lost is lives lost, so better preparation could quite literally save lives.
amelius 9y ago
I suspect that it's mostly the amount of data that allows this, and that the software is just using run-off-the-mill statistics/ML. The title should reflect that. It's not the software, it's the data.
davidw 9y ago
Pretty cool to see those photos. I used to work at a company that makes fundus cameras that take those kinds of pictures.
bromuro 9y ago
> opening a new opportunity for artificial intelligence in the vast and lucrative global health industry. well done! google will join the lucrative industry of the people’s health. the rich will survive, the other will build the machine to make them survive.
la6470 9y ago
This whole medical market is ripe for disruption but the tight regulations , data privacy laws etc is preventing this to reach escape velocity.
leeoniya 9y ago
i think it'll be quite some time until ML and CV can do the breadth of diagnostics for all possible things. my wife is an optometrist. just from looking at retinas, she finds brain tumors among other scary systemic problems. even if you have perfect vision, go in for an exam, it can literally save your life.
sjg007 9y ago
Anybody have the link or how can I test myself?
hemapani 9y ago
Remember what happened they tried to predict the flu. https://www.wired.com/2015/10/can-learn-epic-failure-google-flu-trends/ https://www.wired.com/2015/10/can-learn-epic-failure-google-... In https://www.kaggle.com/c/diabetic-retinopathy-detection https://www.kaggle.com/c/diabetic-retinopathy-detection actually ML algo did better than humans. Yet it will take time for apply these widely.
- candiodari 9y ago
  Those are two entirely unrelated methods. One worked, one failed. One is a sort of "wisdom of the crowd" algorithm (as in human crowds) and it failed after initially showing promise (essentially it worked until someone told the crowd), one is machine learning from images.
okigan 9y ago
>Google’s algorithms were able to predict whether someone had high blood pressure or was at risk of a heart attack or stroke How is that different from just measuring the blood pressure (with really inexpensive device)?
mping 9y ago
Honest question: why is that an alternative medicine eg. Iridology is seen as quackery but a computer predicting heart attack risk is not? I know the science behind AI but from a patient point of view, they both analyze the eye.

52 points by helloworld 9y ago

As my PI notes, a lot of the success here can be explained by the model's excellent ability to determine age from the images of the retina. https://twitter.com/skathire/status/965661596503638016 https://twitter.com/skathire/status/965661596503638016 This is quite good age detection - other efforts that I've seen (from face images, rather than retina images) are +/- 5 years, whereas this paper gets age to within +/- 3 years or so. The fact that the model automatically detects anatomical features that humans recognize as being important is promising for a variety of applications, I think.

logronoide 9y ago

Any optician can do exactly the same thing during a simple check of your eyes. In this case I can’t see the benefit of this type of screening. The tool used by the optician is just a cheap lantern, but only the tool needed by the computer is way more expensive.

ridgeguy 9y ago
Perhaps a cheap lens/illuminator attachment for smartphones could be engineered for retinal imaging. If this method could be adapted to phones, many more people could be checked at low cost.
- logronoide 9y ago
  It already exists. Most opticians take a image of your retina and store it in the cloud for further analysis and tracking the evolution. In Spain (where I live) most opticians do not charge anything for this test, because it’s a minimal part of a typical prescription.
  - froindt 9y ago
    That's like a $40 upcharge at my optometrist (USA). But she is a good family friend, walks me through everything she's seeing on there, and gives me glowing compliments on how beautiful my optical nerve is. It sounds stupid, but I'm going to miss that when I go to my new optometrist.
  - cortesoft 9y ago
    Yeah, but I have great eyesight and haven't been to an optician in 15 years. However, I might still be at risk of a heart attack.
    - logronoide 9y ago
      You don’t need to visit an optician. A family doctor with a lantern can do it (and normally they do). It’s a very simple test. It’s not a ML or AI challenge at all.
      - mkagenius 9y ago
        But if you have AI, you do not need the optician/doctor. For example people in Poor countries have no access to such a person.
        JustSomeNobody 9y ago
        I don't think anyone would say don't build the AI/ML. Just that the technology already exists. Until then, people should continue to see their eye Dr when one is available to them.
        txcwpalpha 9y ago
        There are a few startups/companies that specialize in going to poor countries and using lens attachments on cell phones to do eye exams. They send in people who aren't doctors, take a picture of the eye, and email/upload the pictures to doctors who can analyze it for issues. IIRC, in those situations, the problem isn't the accessibility to a doctor, but rather the accessibility to someone who has a device capable of imaging the eye. I imagine this would be the same. Replacing the doctor with an AI doesn't really solve the problem.
ceejayoz 9y ago
I'd suspect the tools will eventually be dramatically cheaper than a trained optician. Put it in a pharmacy, grocery store, or other central location and people could screen weekly instead of once every couple years (assuming they even see an optician).
- logronoide 9y ago
  As I have explained before, in Spain an optician does not charge anything for it because it’s ridiculously simple and it’s a small percentage of the effort that an optician needs to do for a prescription. This test is different of a blood test for example. You don’t need to control the evolution weekly, or even monthly. This is one of the most simple screening process. Because it’s easy to train the model with tons of images already made by opticians and optometrists. It does not have any value, it’s too simple. My brother is an optician and has a huge database with retina images. If even me with limited knowledge of ML and my brother as field expert could train a model to detect blood pressure problems. Too simple, no value.
  - ceejayoz 9y ago
    You're missing a very key point. Many people with normal eyesight never see an optician. They might use a kiosk at the grocery store. It's also a first step proof-of-concept towards harder diagnoses.
    - logronoide 9y ago
      As I said, not only opticians can do it. A family doctor with a lantern can do it (and they do). It’s not an ML or AI challenge at all.
      - ceejayoz 9y ago
        There are plenty of Americans who don't have access to a "family doctor". Not everyone can afford or bothers to get an annual checkup. Same reason grocery stores often have a "check your blood pressure" kiosk, or offer flu shots. It's quick, it's convenient, and you reach people who would otherwise miss out. Sure, a doctor or nurse can check your BP or give you the vaccine, but there's a significant benefit to expanded access.
elbasti 9y ago
Well, not really. The tool used by the optician is a $5 dollar lantern connected to a brain that had to go through a $300,000 dollar education.

matte_black 9y ago

What do you do after you discover you are at risk? Get your affairs in order?

ceejayoz 9y ago
Adjust diet, exercise, statins, get a cardiology consult?
- matte_black 9y ago
  But if you’re already at the point where you are at risk and your eyes show it, isn’t it already too late to really do anything about it?
  - maxerickson 9y ago
    Weight loss and exercise are recommended to people that survive heart attacks.
  - arkades 9y ago
    Morbidity and mortality can be reduced with the prescription of beta blockers, aspirin, statins, and ace inhibitors. +\- clopidogrel. Being at risk isn’t fate: you have modifiable risk factors that can be mitigated (eg, weight, cholesterol, blood pressure.) Additionally, CPR training of your immediate family can be - literally - lifesaving in the event you do have that heart attack.
logronoide 9y ago
Visit your family doctor ASAP.
tinymollusk 9y ago
It gives a high-risk group the ability to notify their friends and family and practice what to do in event of a heart attack. My mom frequently stays with me, but outside of "call 911" I don't really know what to do in teh event of a heart attack. Time lost is lives lost, so better preparation could quite literally save lives.

amelius 9y ago

I suspect that it's mostly the amount of data that allows this, and that the software is just using run-off-the-mill statistics/ML. The title should reflect that. It's not the software, it's the data.

davidw 9y ago

Pretty cool to see those photos. I used to work at a company that makes fundus cameras that take those kinds of pictures.

bromuro 9y ago

> opening a new opportunity for artificial intelligence in the vast and lucrative global health industry. well done! google will join the lucrative industry of the people’s health. the rich will survive, the other will build the machine to make them survive.

la6470 9y ago

This whole medical market is ripe for disruption but the tight regulations , data privacy laws etc is preventing this to reach escape velocity.

leeoniya 9y ago

i think it'll be quite some time until ML and CV can do the breadth of diagnostics for all possible things. my wife is an optometrist. just from looking at retinas, she finds brain tumors among other scary systemic problems. even if you have perfect vision, go in for an exam, it can literally save your life.

sjg007 9y ago

Anybody have the link or how can I test myself?

hemapani 9y ago

Remember what happened they tried to predict the flu. https://www.wired.com/2015/10/can-learn-epic-failure-google-flu-trends/ https://www.wired.com/2015/10/can-learn-epic-failure-google-... In https://www.kaggle.com/c/diabetic-retinopathy-detection https://www.kaggle.com/c/diabetic-retinopathy-detection actually ML algo did better than humans. Yet it will take time for apply these widely.

candiodari 9y ago
Those are two entirely unrelated methods. One worked, one failed. One is a sort of "wisdom of the crowd" algorithm (as in human crowds) and it failed after initially showing promise (essentially it worked until someone told the crowd), one is machine learning from images.

okigan 9y ago

>Google’s algorithms were able to predict whether someone had high blood pressure or was at risk of a heart attack or stroke How is that different from just measuring the blood pressure (with really inexpensive device)?

mping 9y ago

Honest question: why is that an alternative medicine eg. Iridology is seen as quackery but a computer predicting heart attack risk is not? I know the science behind AI but from a patient point of view, they both analyze the eye.