4 ms·

This is absurd. Cancer is diagnosed incredibly late far too often. If we would dedicate capacity to screen much more often and much earlier we would be far bett

This is absurd. Cancer is diagnosed incredibly late far too often. If we would dedicate capacity to screen much more often and much earlier we would be far better off and many more people would survive. The main issue with universal screening is cost. Medicine is far too expensive and now we need to divide up a scarce resource. The 2nd issue is subclinical cancers that would not kill you before you die of something else. But we don't spend much time on the 2nd issue because we don't screen as much as we could if we made healthcare cheaper.

brianpan 5y ago
It's not absurd, it's statistics. If you're trying to detect something that's fairly rare, it's very easy to have a test produce more false positives than real positives. That means if you randomly take a test and it has a positive result, it's more likely you do NOT have the cancer than do. More detection of cancer and earlier is great. But testing has it's limitations and the general public doesn't always understand the statistics involved to apply the tests correctly. https://www.youtube.com/watch?v=M8xlOm2wPAA https://www.youtube.com/watch?v=M8xlOm2wPAA
KennyBlanken 5y ago
No, the issue with universal screening is that many screenings return positive results but the patient never develops cancer. In the case of some cancer treatments, like prostate cancer, the treatment is invasive and life-alerting on its own. Further, some screening methods involve radiation exposure, and in large population groups, the effects of that additional radiation starts to have a significant effect and may overtake the delta between "screen who we estimate needs it" and "screen everyone", which means you've given cancer to more people, looking for cancer. Wrist injuries are a pain in the ass to x-ray looking for broken bones, but CT scans are very reliable. But...CT scans involve a _fuckton_ more x-ray exposure. They're also more time consuming, use much more expensive equipment, generate a lot more data, and require more work to analyze. So generally, ERs continue to use x-rays to check for broken wrist bones. Also, since you replied to a comment about breast cancer: breast cancer is among the most diagnosed, most successfully treated, and most survivable cancers. There is no scientific basis for easing or promoting more breast cancer screenings.
- vimy 5y ago
  With enough data we can filter out false positives. Thus we need mass scanning to gather this data and we will have a transition period where there will be a lot of false positives. I think the benefits outweigh the negatives in the longterm.
  - Groxx 5y ago
    In many cases, more data does not mean higher accuracy / fewer false positives. It just means your estimates on how many false positives the system produces get more accurate.
  - foerbert 5y ago
    This is not actually necessarily true at all, at least not in a useful way. Even if it happens to be true, "with enough data" is a hell of a qualifier for many things, never mind one that involves practicing medicine on real patients. I can't help but suspect that the brilliant success of machine learning in some tasks with a ludicrous amount of data available has blinded us a bit. "Just use more data" is a seductively easy answer to all questions if we ignore any and all factors beyond what is theoretically possible in the 'true Turing machine on a perfectly flat friction-less plane...' sense. Mammograms are used for preliminary screening. They are cheap, relatively easy to perform, and close to as (physically) noninvasive as you can get for something reasonably called a test. So immediately most data is simply unavailable without obviating the entire point of doing better preliminary screening. Any data we can consider without making the entire endeavor pointless will be extraordinarily challenging to work with. Even if you use the magic hand-wave to move few 'little' fields like probably most all of NLP into the realm of the trivial, this is an extremely challenging problem. Even if you can magically encode the entire medical history, you'd have what, a massively sparse set of massively sparse time series? And lots of information is going to be in the relationships of different time series, but also you have to deal with doctors not being a single immutable entity... and doctors can only see so many patients in a day, and they only work so many years, and and andandand... So many difficulties, so many confounding problems, and so little data to brute force the solution. And if we somehow rally the world to our cause and somehow throw enough resources at it to amazingly solve all the problems... what do we do for the world? Let's use the best imaginable case. The new test is perfectly accurate, specific, and sensitive. It is instant, risk-free, and otherwise just free in all ways you can imagine, to the point that all breast cancer is now detected at the literal possible earliest moment. Not only that, we can do way better. We can even determine if it should be treated or not. Almost 13% of women in the US will get invasive breast cancer in their lifetime, but only 2.6% die of it [0]. We can spare that 10% or so of dreadful side effects from chemo, radiation, major surgery, and all that. Just to put the icing on the impossibly-large cake, we can even suggest optimal treatments, so that none of those 2.6% die from treatment. We have effectively eradicated death by breast cancer in any way. Of course, this would (very seriously) be extraordinarily great. I don't even have the words, so I'll just leave it. This is best case scenario. Everything is impossibly perfect. We're probably at least a couple dozen plausible-actual-miracles deep. So how does this stack up? Well... 2.6% probably clued you in. Breast cancer is not a particularly huge cause of death. From [1] (if you forgive the year gap), we're looking at something on the scale of diabetes (2.7%). We're well below death by unintentional injury (4.4%). Heart attack (21.8%) and all cancer (20.7%) are the two major ones. But breast cancer, well, isn't the only cancer. But the other side of the coin is that we expanded preliminary screening to young folks. So maybe if we save a lot of them, we've added more human-years-lived than it seems from total death rates. Well... no. Mammograms are used later in life for a reason. Younger folks aren't very likely at all to have invasive breast cancer. At 20, 0.1% of women get it, and 0.055% die of it. At 40 (the age suggested in [0] to start offering mammograms) it's 1.5%/0.2%. You get the idea. Maybe for some age it's at least a significant cause of death? Nope. [1] has a breakdown by age. So frankly... absolute impossible best-case we mostly prolong the lives of a few percent of the oldest women. Large numbers obviously means there's more than a just a few younger women too, but the distribution of possible beneficiaries is rather skewed towards the oldest. Still. It'd be unfathomably awesome to see for many reasons - both technical and humanitarian. But the degree of difficulty is extreme and the far, far upper bound isn't quite proportional. There's easier targets with better rewards. My random speculation from thinking about this for way too long to write this is that solving many of those huge challenges along the way would probably be the bigger boon for humanity in the end. TL;DR (yes it's still long, but relative, okay?) * Preliminary screens need to be simple and non-invasive means way little data, means just throwing machine learning at it very hard * Many very hard problems means maybe tons of data is needed to push through with sheer weight of data, but human interactions can only generate so much data so fast * Not that many people of any age get breast cancer, and far fewer actually die of it * And the age distribution of those who get it skews substantially older... hence the entire reason for mammograms not being used on young folk in the first place... and hence this entirely way too long response to a rather short and probably not-much-thought-of comment [0] https://www.cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/breast-cancer-facts-and-figures/breast-cancer-facts-and-figures-2019-2020.pdf https://www.cancer.org/content/dam/cancer-org/research/cance... [1] https://www.cdc.gov/women/lcod/2017/all-races-origins/index.htm https://www.cdc.gov/women/lcod/2017/all-races-origins/index....
eru 5y ago
Compare and contrast this video from 'Medlife Crisis': https://www.youtube.com/watch?v=yNzQ_sLGIuA https://www.youtube.com/watch?v=yNzQ_sLGIuA > Trying to detect disease early seems like an obviously good thing to do. But reality isn't always quite so straightforward, especially in medicine. If It Ducks Like A Quack is a series where I try to go beyond the hype and explore how complicated the apparently simple phenomena in medicine are. One point from the video: five year survival rates of prostate cancer in the US are massively better than in the UK. However, roughly the same proportion of people die of prostate cancer in both countries. The US just screens a lot more.
bsder 5y ago
> This is absurd. It is not, and the issues are central to both statistics and medicine. 1) Medical tests have four quadrants--not 2. The accurate quadrants are "positive test--positive diagnosis" and "negative test--negative diagnosis". The problem quadrants are "positive test--negative diagnosis" and "negative test--positive diagnosis". A positive test with a negative diagnosis can cause issues like they do for mammograms. The problem is that the followup procedures have a non-zero probability of killing the patient. Biopsies can result in infections. Anxiety from a positive test can cause heart attacks and strokes. If your test is only 99% accurate, then 1% of people being tested are incurring those risks when they don't have to. If you are testing a lot of people and the disease is relatively rare, you may kill more people with false positives than the disease itself. A negative test with a positive diagnosis is problematic because people will tend to blow off other signs since "the test was negative". This can delay lifesaving treatment. Finally, optimizing one of those quadrants generally occurs at the expense of another. So, which should a test prefer? 2) The body often clears a lot of diseases. Lumps that could be cancerous often disappear. Swellings come and go. Viruses do more and varied damage than we originally thought. The body repairs itself and reversion to the mean is a thing. 3) Diseases may not even be that problematic. Sure, aggressive prostate cancer is an issue. However, practically every single man will eventually get prostate cancer. Most men die of something else before prostate cancer is an issue. Is it worth risking becoming incontinent or impotent from a prostate cancer treatment? 4) Will the positive test even lead to a correct request for treatment? People demand antibiotics for viral infections all the time. My body never gets rid of strep throat, so I will force my doctor to give me a quick strep test when I have a persistent sore throat. The first time I did this my doctor was literally stunned when I simply thanked him and got ready to walk out since the test came back negative. Everyone else will demand antibiotics even if the test is negative.
- tomohawk 5y ago
  I wonder if some of this could be mitigated with a sort of double blind take home test. You take the test, but you do not know what it is screening for, or what the results mean. It tests for N things which can be economically packaged. The results can then be used to produce a recommendation for additional consultation with a doctor.

brianpan 5y ago

It's not absurd, it's statistics. If you're trying to detect something that's fairly rare, it's very easy to have a test produce more false positives than real positives. That means if you randomly take a test and it has a positive result, it's more likely you do NOT have the cancer than do. More detection of cancer and earlier is great. But testing has it's limitations and the general public doesn't always understand the statistics involved to apply the tests correctly. https://www.youtube.com/watch?v=M8xlOm2wPAA https://www.youtube.com/watch?v=M8xlOm2wPAA

KennyBlanken 5y ago

No, the issue with universal screening is that many screenings return positive results but the patient never develops cancer. In the case of some cancer treatments, like prostate cancer, the treatment is invasive and life-alerting on its own. Further, some screening methods involve radiation exposure, and in large population groups, the effects of that additional radiation starts to have a significant effect and may overtake the delta between "screen who we estimate needs it" and "screen everyone", which means you've given cancer to more people, looking for cancer. Wrist injuries are a pain in the ass to x-ray looking for broken bones, but CT scans are very reliable. But...CT scans involve a _fuckton_ more x-ray exposure. They're also more time consuming, use much more expensive equipment, generate a lot more data, and require more work to analyze. So generally, ERs continue to use x-rays to check for broken wrist bones. Also, since you replied to a comment about breast cancer: breast cancer is among the most diagnosed, most successfully treated, and most survivable cancers. There is no scientific basis for easing or promoting more breast cancer screenings.

vimy 5y ago
With enough data we can filter out false positives. Thus we need mass scanning to gather this data and we will have a transition period where there will be a lot of false positives. I think the benefits outweigh the negatives in the longterm.
- Groxx 5y ago
  In many cases, more data does not mean higher accuracy / fewer false positives. It just means your estimates on how many false positives the system produces get more accurate.
- foerbert 5y ago
  This is not actually necessarily true at all, at least not in a useful way. Even if it happens to be true, "with enough data" is a hell of a qualifier for many things, never mind one that involves practicing medicine on real patients. I can't help but suspect that the brilliant success of machine learning in some tasks with a ludicrous amount of data available has blinded us a bit. "Just use more data" is a seductively easy answer to all questions if we ignore any and all factors beyond what is theoretically possible in the 'true Turing machine on a perfectly flat friction-less plane...' sense. Mammograms are used for preliminary screening. They are cheap, relatively easy to perform, and close to as (physically) noninvasive as you can get for something reasonably called a test. So immediately most data is simply unavailable without obviating the entire point of doing better preliminary screening. Any data we can consider without making the entire endeavor pointless will be extraordinarily challenging to work with. Even if you use the magic hand-wave to move few 'little' fields like probably most all of NLP into the realm of the trivial, this is an extremely challenging problem. Even if you can magically encode the entire medical history, you'd have what, a massively sparse set of massively sparse time series? And lots of information is going to be in the relationships of different time series, but also you have to deal with doctors not being a single immutable entity... and doctors can only see so many patients in a day, and they only work so many years, and and andandand... So many difficulties, so many confounding problems, and so little data to brute force the solution. And if we somehow rally the world to our cause and somehow throw enough resources at it to amazingly solve all the problems... what do we do for the world? Let's use the best imaginable case. The new test is perfectly accurate, specific, and sensitive. It is instant, risk-free, and otherwise just free in all ways you can imagine, to the point that all breast cancer is now detected at the literal possible earliest moment. Not only that, we can do way better. We can even determine if it should be treated or not. Almost 13% of women in the US will get invasive breast cancer in their lifetime, but only 2.6% die of it [0]. We can spare that 10% or so of dreadful side effects from chemo, radiation, major surgery, and all that. Just to put the icing on the impossibly-large cake, we can even suggest optimal treatments, so that none of those 2.6% die from treatment. We have effectively eradicated death by breast cancer in any way. Of course, this would (very seriously) be extraordinarily great. I don't even have the words, so I'll just leave it. This is best case scenario. Everything is impossibly perfect. We're probably at least a couple dozen plausible-actual-miracles deep. So how does this stack up? Well... 2.6% probably clued you in. Breast cancer is not a particularly huge cause of death. From [1] (if you forgive the year gap), we're looking at something on the scale of diabetes (2.7%). We're well below death by unintentional injury (4.4%). Heart attack (21.8%) and all cancer (20.7%) are the two major ones. But breast cancer, well, isn't the only cancer. But the other side of the coin is that we expanded preliminary screening to young folks. So maybe if we save a lot of them, we've added more human-years-lived than it seems from total death rates. Well... no. Mammograms are used later in life for a reason. Younger folks aren't very likely at all to have invasive breast cancer. At 20, 0.1% of women get it, and 0.055% die of it. At 40 (the age suggested in [0] to start offering mammograms) it's 1.5%/0.2%. You get the idea. Maybe for some age it's at least a significant cause of death? Nope. [1] has a breakdown by age. So frankly... absolute impossible best-case we mostly prolong the lives of a few percent of the oldest women. Large numbers obviously means there's more than a just a few younger women too, but the distribution of possible beneficiaries is rather skewed towards the oldest. Still. It'd be unfathomably awesome to see for many reasons - both technical and humanitarian. But the degree of difficulty is extreme and the far, far upper bound isn't quite proportional. There's easier targets with better rewards. My random speculation from thinking about this for way too long to write this is that solving many of those huge challenges along the way would probably be the bigger boon for humanity in the end. TL;DR (yes it's still long, but relative, okay?) * Preliminary screens need to be simple and non-invasive means way little data, means just throwing machine learning at it very hard * Many very hard problems means maybe tons of data is needed to push through with sheer weight of data, but human interactions can only generate so much data so fast * Not that many people of any age get breast cancer, and far fewer actually die of it * And the age distribution of those who get it skews substantially older... hence the entire reason for mammograms not being used on young folk in the first place... and hence this entirely way too long response to a rather short and probably not-much-thought-of comment [0] https://www.cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/breast-cancer-facts-and-figures/breast-cancer-facts-and-figures-2019-2020.pdf https://www.cancer.org/content/dam/cancer-org/research/cance... [1] https://www.cdc.gov/women/lcod/2017/all-races-origins/index.htm https://www.cdc.gov/women/lcod/2017/all-races-origins/index....

eru 5y ago

Compare and contrast this video from 'Medlife Crisis': https://www.youtube.com/watch?v=yNzQ_sLGIuA https://www.youtube.com/watch?v=yNzQ_sLGIuA > Trying to detect disease early seems like an obviously good thing to do. But reality isn't always quite so straightforward, especially in medicine. If It Ducks Like A Quack is a series where I try to go beyond the hype and explore how complicated the apparently simple phenomena in medicine are. One point from the video: five year survival rates of prostate cancer in the US are massively better than in the UK. However, roughly the same proportion of people die of prostate cancer in both countries. The US just screens a lot more.

bsder 5y ago

> This is absurd. It is not, and the issues are central to both statistics and medicine. 1) Medical tests have four quadrants--not 2. The accurate quadrants are "positive test--positive diagnosis" and "negative test--negative diagnosis". The problem quadrants are "positive test--negative diagnosis" and "negative test--positive diagnosis". A positive test with a negative diagnosis can cause issues like they do for mammograms. The problem is that the followup procedures have a non-zero probability of killing the patient. Biopsies can result in infections. Anxiety from a positive test can cause heart attacks and strokes. If your test is only 99% accurate, then 1% of people being tested are incurring those risks when they don't have to. If you are testing a lot of people and the disease is relatively rare, you may kill more people with false positives than the disease itself. A negative test with a positive diagnosis is problematic because people will tend to blow off other signs since "the test was negative". This can delay lifesaving treatment. Finally, optimizing one of those quadrants generally occurs at the expense of another. So, which should a test prefer? 2) The body often clears a lot of diseases. Lumps that could be cancerous often disappear. Swellings come and go. Viruses do more and varied damage than we originally thought. The body repairs itself and reversion to the mean is a thing. 3) Diseases may not even be that problematic. Sure, aggressive prostate cancer is an issue. However, practically every single man will eventually get prostate cancer. Most men die of something else before prostate cancer is an issue. Is it worth risking becoming incontinent or impotent from a prostate cancer treatment? 4) Will the positive test even lead to a correct request for treatment? People demand antibiotics for viral infections all the time. My body never gets rid of strep throat, so I will force my doctor to give me a quick strep test when I have a persistent sore throat. The first time I did this my doctor was literally stunned when I simply thanked him and got ready to walk out since the test came back negative. Everyone else will demand antibiotics even if the test is negative.

tomohawk 5y ago
I wonder if some of this could be mitigated with a sort of double blind take home test. You take the test, but you do not know what it is screening for, or what the results mean. It tests for N things which can be economically packaged. The results can then be used to produce a recommendation for additional consultation with a doctor.