Slashdot Mirror
Scientists Develop a Breathalyzer That Detects 17 Diseases With One Breath From a Patient (qz.com)
randomErr quotes a report from Quartz: In the last 10 years, researchers have developed specific sniff tests for diagnosing tuberculosis, hypertension, cystic fibrosis, and even certain types of cancer. A group of global researchers led by Hossam Haick at the Israel Institute of Technology have taken the idea a step further. They've built a device -- a kind of breathalyzer -- that is compact and can diagnose up to 17 diseases from a single breath of a patient. The breathalyzer has an array of specially created gold nanoparticles, which are sized at billionths of a meter, and mixed with similar-sized tubes of carbon. These together create a network that is able to interact differently with each of the nearly 100 volatile compounds that each person breaths out (apart from gases like nitrogen, oxygen, and carbon dioxide). Haick's team collected 2,800 breaths from more than 1,400 patients who were each suffering from at least one of 17 diseases (in three classes: cancer, inflammation, and neurological disorders). Each sample of the disease was then passed through the special breathalyzer, which then produced a dataset of the types of chemicals it could detect and in roughly what quantities. The team then applied artificial intelligence to the dataset to search for patterns in the types of compounds detected and the concentrations they were detected at. As they report in the journal ACS Nano, the data from the breathalyzer could be used to accurately detect that a person is suffering from a unique disease nearly nine out of ten times.
their promises didn't turn out so well for the investors
Pain is merely failure leaving the body
Here is the list of detected diseases from the source report. ;
lung cancer, colorectal cancer, head and neck cancer, ovarian cancer, bladder cancer, prostate cancer, kidney cancer, gastric cancer, Crohn’s disease, ulcerative colitis, irritable bowel syndrome, idiopathic Parkinson’s, atypical Parkinsonism, multiple sclerosis, pulmonary arterial hypertension, pre-eclampsia, and chronic kidney disease.
Halitosis.
Have gnu, will travel.
But can it tell if someone is drunk or under the influence of drugs? If it can this could be a nice side business for the police.
Minimum threshold fixed. Thanks!
We seem to be coming closer by the day to a Star Trek tricorder.
A country that does so much good, I knew that's where that good news was coming from. It's a real shame that they just got sucker punched by the lame ducker.
is "nanometer" outside the vocabulary range of their normal readers? o_O
Anons need not reply. Questions end with a question mark.
And you have cancer but if you punch me in the face I can get you the jail / prison will the really good doctors at no cost to you other then being in lock up for a some time.
And in which order ?
Votez ecolo : Chiez dans l'urne !
We've known dogs can accurately detect several cancers by smelling a person's breath for almost a century. India is beginning to use rats (they're easier to train and have a more sensitive nose) as an auxiliary screening system for things like tuberculosis, with results generally more accurate than screenings by human experts. And he sensitivity of electronic "noses" has been advancing rapidly, so it seems perfectly reasonable that they could achieve similarly impressive results, with the added advantage that they offer objective, quantitative results that lend themselves to easy analysis and lookup, without any individual training needed.
--- Most topics have many sides worth arguing, allow me to take one opposite you.
Wow. You really have to feel sorry for that patient. Its a wonder they can generate one breath.
The summary only gives the rate for people that have the disease, if you want to know the false positive rate you'll have to read the actual report.
Also, the article says they consider it a good proof of concept but still far away from being used in actual diagnosis.
I don't think it is a 10% false positive rate, but if so that would be great. From the description, it sounds like the cost per test would be very low after economies of scale are realized. Therefore, the doctor could use this as a routine part of the annual checkup. If the machine says "Parkinson's is likely", then the doctor would know to investigate the possibility of Parkinson's. Many (most?) of the routine screening tests aren't definitive - they provide evidence that the doctor will then follow up on.
Have you ever had a throat culture? The doctor did a culture because there was some evidence of an infection that could be definitely diagnosed by a culture. First there's the screening which tells the doctor which more reliable (and expensive) tests should be run, THEN you run the more reliable test.
I've been working in this field for a long time. If you look around the literature, you'll see my name on several papers on nanoelectronic detection of disease via breath. This is a great demo, and Haik is a very good guy in this field, but he's done only the easiest part. I've learned the hard way that publishing an academic paper and making something that doctors actually would buy to make treatment decisions are completely different things. This is the first step in the development process, not the last.
In this case, there are already medical breath tests, and entire clinics devoted to this kind of medical test (without the nanotech part). The tools are already cleared for use, and medical doctors have protocols and billing methods for using them. If the key part of this is really those 13 compounds, there's no need for nano wizardry; use the mass spec or whatever that the clinic already has. That's really the key here, why would anyone use his device, and not just his results? Often in sensor research, we don't understand the distinction there when the results get us such great publications and press. The grant manager paid for the nanotechnology (and the citations that come with it), but everyone else is interested in the medicine.
As a hypochondriac I'd be scared to take this test. Best case, I have one or all of these diseases. Worst case, I have no clue what I'm dying from.
Although to detect diseases 2, 9, 10, and 11 - you have to "blow" in a somewhat different manner.
#DeleteChrome
Really, wow would consider anything accurate if it diagnosed over 10% of patients incorrectly. For instance mamography for women under 50 about 80%, which is why it is not always recommended for y0unger women. Here we are talking about 17 diagnosis with a 10% error. Regular use would likely mean that you would be guaranteed to be diagnosed with something eventually, something you did not have, while missing the thing you do.
"She's a scientist and a lesbian. She's not going to let it slide." Orphan Black
That way i would feel safer on one night stands and fucking hookers.
A fart would probably tell them a lot more.
If my buddies are any indication, they died years ago from various loathsome diseases, and are sending back evidence to their still-ambulatory bodies about what the air is like in hell.
Or maybe China.
I've calculated my velocity with such exquisite precision that I have no idea where I am.
Reminds me of this old joke.
I'm sorry, but a 10% false positive rate is terrible, especially for screening purposes.
For example:
Lets assume that a true diagnosis for disease X is at a rate of 10 in 100,000 people per year (fairly reasonable incidence rate for a lot of diseases).
Assuming there are no false negatives, the number of diagnoses that are made using the test will be:
Number of true positives + 10% of the true negatives
Number of true positives = 10
Number of true negatives = 99,990
# of diagnoses made = 10 + 0.10*99,990
= 10,009
The 9999 false positives would have to have further follow up tests and possibly erroneous treatment. This then of course puts a significant financial burden on the heathcare system.
For instance mamography for women under 50 about 80%, which is why it is not always recommended for y0unger women.
I was under the impression that the added risk of getting cancer from repeated X-rays over decades was why it isn't done on younger people.
Some countries don't recommend mammograms at all except for high risk individuals, and instead teach young women how to feel for irregularities.
and to add to my own post:
There are 17 diseases that the breath analyser detects, if each has a false positive rate of 10%, we have the following:
9999 * 17 = 169,983 false positives per 100,000 people per year.
It'd be like one of those kids' athletic competitions, everyone gets a prize.
If you want to eliminate neural nets as a form of artificial intelligence, you are going to need to conclude that most human "reasoning" is similarly not really intelligent behavior. Plenty of research shows that humans make most decisions in a manner highly analogous to those in neural nets (and you can predict the result of this "reasoning" by brain monitoring before the subject knows which decision he is going to make). It is true that humans, if challenged, will attempt to justify their decisions, but their justifications are often pretty nonsensical. Meanwhile, while complex neural nets, trained by large volumes of empirical data, indeed cannot simply explain how decisions are reached, we at least can control the data used for that training. Human "decision making" is based on data that is often highly dubious, and (although subject to attempted justification) similarly shaped by complex training from large data sets that cannot be explained simply.
Fair enough. Thank you for that. I guess I just prefer they call this machine learning rather than artificial intelligence, but that's my own idiosyncrasy.
Error: NSE - No Signature Error
Some diseases have an incubation period of thirty and more years. But they can be diagnosed very early by modern medical tools and be a good source of an income for doctors and clinics.
Thanks for that; this is what most people don't get about the term "false positive rate". In your example, if you are diagnosed by the machine as having one of the diseases, the odds that the machine is wrong (i.e. that you don't have the disease after all) is .999 Most people figure it's 0.1 (the false positive rate)
If construction was anything like programming, an incorrectly fitted lock would bring down the entire building...
Thanks. What I was thinking of, I have now learned, is called the "false discovery rate". FDR is "10% of the samples flagged positive were actually negative". If a test is cheap, a 10% FDR os okay, a 10% FPR is not, (unless perhaps a large percentage of samples actually are positive).
I just studied the two for a few minutes to get an idea of which rate is most useful to consider for the tests I create. It seems false DISCOVERY rate is often useful when there are many tests done on a relatively small number of samples. That somewhat describes my testing - I test for about 90,000 hypotheses (90,000 conditions) on approximately 90,000 samples. I normally think about "what percentage of our positives are false? (FDR)" and it seems that's appropriate for the testing we do.
Indeed I was thinking of the false discovery rate - what percentage of positive results are false. After doing some reading, I just learned that false discovery rate is most useful when testing a small number of samples for many conditions. False positive rate is most useful when testing a large number of samples for a small number of conditions.
That's interesting to me because I develop a testing system that tests for about 90,000 conditions and tests about 90,000 "patients". My patients are computers, and I test for 90,000 different security weaknesses.
Don't kiss that guy.
I'm on whichever side doesn't practice eugenics.
They just have a hard time communicating the results. Somebody should work on that part.
A neural network does not implement an algorithm, so no.
Guns don't kill people; Physics kills people! - John Lithgow as Dick Solomon on Third Rock From The Sun
That's nice, but I'll settle for a breathalyzer that can be scrutinized by anyone but the manufacturer in court. I'm not sure I trust a DUI conviction to a black box nobody can look inside including the courts.
You mis-spelled $50,000.
Well, technically it does. But the algorithms are the relatively trivial ones that do the neuron implementations, as opposed to the more traditional algorithmic direct solution.
I can't necessarily comment on other diseases in the list, but it's absurd to claim that a person's blood-pressure is detectable by breath.
The smart money says this is fake.
---- I'll take you in a Hunt deathmatch any day.
There are algorithms at its core, but it does not implement an algorithm. My point stands.
Guns don't kill people; Physics kills people! - John Lithgow as Dick Solomon on Third Rock From The Sun
Single breath analysis sounds hype-py. I would go for a series of tests at specified intervals for each disease. It also seems to be biased toward the presence of fecal matter in breath, but such can come from bad food, organics or specialty cafeteria coffee chain coffee... Ahem, anyway, assumptions also seem detached from actual hypothesis -mechanisms on the relation between diseases and breath presence of chemicals. Like they run an AI associative network and called it results? Surely several other molecules can be detected with organometallic devices... So overall sounds a good idea with bad scientists. Though what seems alarming is that if I use that device after and before a meal AND IT DETECTS COLON CANCER, chances are I can sue the place for mixing excrement in food! Still to achieve epidemiological levels they should use a bigger sample, five thousand at least over a number of periods. But they would be infringing Afroarab interests if we can show a place gave me bad breath out of bad hygiene or worse practices.
The term used for the popular recent solutions is "deep learning". To be more specific, the most effective solutions are "guided deep learning". The term "guided" means that the important inputs and outputs are partly chosen by humans trying to tune the learning process. Progress has been rapid in just the last two or three years. Image recognition, for instance, an extremely tough area until very recently, is now pretty much solved. In this area, the next frontier to be cracked is totally independent learning without any need for humans to be involved. Such a breakthrough may or may not be achieved quickly.
Another very interesting area of research is how to deal with imperfect information. Where large amounts of data is available, and that data can unambiguously be used to determine a correct solution (such as moves in chess, or analysis of MRI scans for tumor analysis) artificial intelligence can already surpass the performance of any human if the AI system is given sufficient training. With AIs that must deal with imperfect information (especially prediction of what humans or other AIs might do) progress is being made, but the best humans are generally still superior to the best AIs. Examples are playing poker, and stock market decisions (though the latter is still heavily AI assisted).
Still a major problem for AIs is where limited clearly relevant data to guide decision making is available. Clearly, humans rely on a lot of peripheral experience to suggest a plan of action. The actions taken may be imperfect, but at least there is a basis for the decision. Before AIs can be made equally (or hopefully more) adept, the process needs to be better understood.