This is one of those medical revolutions that I am waiting dearly for.<p>Facilities that are not hospitals(to avoid the risk of occupying medical devices that sick people need) built to _regulary_ check up otherwise healthy people for preventive care.<p>Heck, I have so many alerts defined on my monitoring setup for servers to watch for signals of failure before they get too big. But, my own body is not observed until something bad needs treatment. Why can’t we observe ourselves medically and analyze that record for early signs of trouble before it becomes serious?!<p>All the advancement in technology in recent years, this ought to happen sooner than later.
0.5% false positive rate is actually <i>quite</i> high. I don't have the numbers, but I'd be surprised if more than 1% of the people that would receive such a screening <i>actually already have cancer</i>, which would mean that it produces about as many false positives as false negatives.<p>Picture a population of a million people, all receiving the test. 1% of them <i>have</i> cancer (unknown to them) - half of <i>those</i> people get a negative screening result and half get a positive one (51% successful identification), and 0.5% of the <i>full population</i> get a negative screening result. Those numbers are roughly equal.<p>That doesn't make it useless by any means, but it's not nearly as 'specific' as it sounds on paper.
For those interested in reading the actual paper, this appears to be it (open access too):<p><a href="https://www.annalsofoncology.org/article/S0923-7534(21)02046-9/fulltext" rel="nofollow">https://www.annalsofoncology.org/article/S0923-7534(21)02046...</a><p>I’ve only skimmed it, but seen quite a few limitations, notably those with “non-malignant conditions at enrolment” (it would be nice to know more what that means), previous cancer or recent corticosteroids use were excluded. Additionally, it’s a case controlled trial which don’t always translate to screening tools (as mentioned in their own discussion).<p>The problem with screening is that you are doing something on healthy patients, so particularly for rare cancers even a small risk of a false positive is significant. In this case it’s 99.5%. So if you test 1,000,000 people, 5000 (*correction from 500) people will come up positive. This is great for common diseases, but if it’s rare and only 1 in a million has it then you have got 5000 false positive for every 1 positive.<p>I think this will likely be a useful test (if it translates well to a wider screening population), but it’s not as good as it first seems.
If you’re tested positive, what’s the probably that you actually have cancer?<p>(The article states 0.5% false positive rate and about 50% true positive rate, but I would need to know the the prevalence of cancer in the population to compute what I am asking for).
"It correctly identified when cancer was present in 51.5% of cases, across all stages of the disease, and wrongly detected cancer in only 0.5% of cases."<p>Doesn't this seem kind of low? Just a bit better than a coin flip. Of course, it rises to 65% and 87% for certain circumstances and the false positive rate is low, but it seems like a lot of cancers could fail to be detected and give a false assurance patients are cancer free. When symptoms emerge later they may be less concerned with getting it checked out. Is this in line with standard performance of tests?
More technologies like this are coming...<p>Is it worth taking blood samples now, keeping them in a freezer, and then waiting for the test tech to come so I can get it tested to see disease progression of whatever disease I get when I get older?
I hope this test becomes mainstream in the next few years. One of those things that the world desperately needs.<p>There is another company in India, Tzar Labs, that has been working on a similar test and is almost ready to launch as well.<p><a href="https://epaper.livemint.com/Home/ShareArticle?OrgId=75ef980ce4" rel="nofollow">https://epaper.livemint.com/Home/ShareArticle?OrgId=75ef980c...</a>
There is a big discussion in this post about iatrogenics being introduced by these tests. It sounds like the fear here is that it'll tax an already burdened system with more patients (some of which just got a false positive).<p>I'm failing to understand how that's any different from other blood tests: they signal a problem, then more has to be established to validate the signal. Nobody is saying "this test proves someone has cancer" and I'm pretty sure doctors already have to discuss how tests can be inaccurate with patients. I believe it's the case that symptoms of cancer (just like other illnesses) may be ignored exactly because there aren't any other signals to indicate cancer. Plus zebras and horses and all of that too. Perhaps someone does have symptoms but nobody connects the dots, and these tests might connect those dots. Is there some reason that this argument is invalid in medical science?
> The test, which is also being piloted by NHS England in the autumn, is aimed at people at higher risk of the disease including patients aged 50 or older.<p>Does anyone know if the study was performed with a population that matches this description? Curious if the rates are in a general population or this higher risk group.
Whole area of having tests that give definitive answers in the field of health is the one area in which it can't get enough of. Biggest issue health wise many have is time from issue to getting a correct diagnoses, whatever the ailment.
If it's detecting 50 types of cancer and the patient doesn't have symptoms, do they just do a full body MRI to find the source? If so, why not just do full body scans which can find other issues, like aneurysms or the other almost 50% of positives that it missed?<p>I get that cost is a big issue, but it seems like the test is missing a lot and you might get more bang for the buck with a periodic MRI from the perspective of the number of potential issues it can find. Either way can result in false results.
Took me quite some time to find the paper that seemed to describe the ML algo[1]. Pretty disappointingly domain-specific and barebones ("logistic regression").<p>And indeed, model form was so unimportant that it was relegated to the supplement. [2]<p>Yet again evidence that access to data and domain knowledge trump fancy ML algos 10/10 times.<p>[1] <a href="https://www.sciencedirect.com/science/article/pii/S0923753420360580" rel="nofollow">https://www.sciencedirect.com/science/article/pii/S092375342...</a><p>[2]
> Custom software was built to classify samples using source models that recognized methylation patterns per region as similar to those derived from a particular cancer type, fol- lowed by a pair of ensemble logistic regressions: one to determine cancer/non-cancer status and the other to resolve the TOO to one of the listed sites (supplementary information)
Highly encourage reading the paper referred to (but not linked!) in the article:<p><a href="https://www.annalsofoncology.org/article/S0923-7534(21)02046-9/fulltext" rel="nofollow">https://www.annalsofoncology.org/article/S0923-7534(21)02046...</a><p>Some good breakdown of test performance by stage and cancer types.
I don’t have much to add to the discussion here. Sensitivity is poor, specificity is pretty good and could be better if used as a confirmatory test(?). What I found interesting in the comments is how many people involved in healthcare lurk on a forum very much not dedicated to healthcare. Kind of cool. You all should chime in more, I learned a lot.
> It correctly identified when cancer was present in 51.5% of cases, across all stages of the disease, and wrongly detected cancer in only 0.5% of cases.<p>While I think this is a great step forward, How can this be described as highly accurate when it missed identifying cancer in 48.5% of the people?
One day you'll put your finger on a sensor on your phone that pricks you to get a drop of blood and analyze it right there on the spot. Kinda like square does for mobile credit card payments.<p>Maybe that's how you'll unlock it too, which might help with phone addiction /s
What makes this even more revolutionary to me (retired anesthesiologist) is that the methodology allows for use of Theranos-style finger-prick-size blood samples rather than IV blood draw, since fragment identity is the indicator rather than level/concentration.
Why wait until 2023 for a larger study? This is a game changing technology that should be pursued like the COVID vaccine. We should be dumping billions into this across the globe to improve its accuracy and drive costs down.
So many tech people here thinking they know how the entire medical industry works... medical problems and the research that arises from them are staggeringly complex and messy. Please... if you've never had experience with this kind of thing, don't pretend that you do. Some uninformed person might see your comment and think you actually know something.
One of my great fears in life is that employers/insurers will impose this en masse on people, causing immeasurable harm to the mental health of people who can’t handle the anxiety that this type of approach to medicine would cause. I’m definitely a big proponent of evidence-based medicine and using it to help living quality life as long as possible, but this sort of thing would just wreck my day to day existence and destroy my quality of life. I am not a server in a data center. I am a human with consciousness and emotions, the desire to live, and a fear of suffering and death.