Deep Nose may one day be the best diagnostician in medicine
nautil.us
nautil.us
Digital chemical sensing is hard, and yet biological chemical sensing is amazing. Understandably, there's always been immense interest in digitizing the biological process of smell. People have been pitching various technologies, from the early days of the WaspHound[0] and DARPA's largely failed RealNose project[1], to current efforts from startups[2] and big players in perfumery[3].
It is my opinion that all of these technologies are solving the wrong problem. Olfactory receptors are just plain horrible proteins. They're extremely difficult to produce, and even when you can express them, they have exactly the wrong type of properties. Each olfactory receptor responds to many different chemicals with really bad sensitivity. While that works really well for animals, these properties are a nightmare for digitization.
If we want to digitize chemical sensing, I believe we will need better proteins than naturally occurring olfactory receptors. And that's what I am trying to do currently in my Ph.D. research!
[0] https://en.wikipedia.org/wiki/Hymenoptera_training [1] https://www.defensedaily.com/darpa-awards-contracts-for-sens... [2] https://yesse.tech/ [3] https://www.firmenich.com/company/research
I totally agree, it's a feature not a bug. A wide spectrum is absolutely an essential feature for an animal to sense thousands (to millions) of different chemicals without thousands (to millions) of different receptors. So, calling olfactory receptors "just plain horrible proteins" was a bit rude to them, haha!
However, in my opinion, digital sensors based on olfactory receptors have and will continue to suffer from such wide spectra and poor sensitivities.
If Spectre can use a timing side channel to peer into regions of memory that were not supposed to be accessed, how many more side channels are there in a human body that mostly wasn't designed to conceal anything? (That's not exactly true, we've evolved to hide quietly, many things hunt by smell, women conceal their state of fertility, etc. But technology can access a much lower threshold of signal than anything evolution has had to worry about.)
If you listen closely enough, I'd expect we're all open books. Time the exact duration that nerve impulses travel, that blood coagulates, that sweat is produced, that hormones are released, etc. Look at the ratios of chemicals produced by digestion, respiration, locomotion, etc.
Technologists are already mining some of this wealth of signal from just filtered speech recordings. There's so much more that could be harnessed (and abused).
Instead of creating a single modality i.e. artificial noses or ears, medicine would benefit from a multimodal sensory system that uses vision, smell, auditory and perhaps even taste, _in conjunction_, as the human body is rich with signals through all these modalities.
Thats a great thought right there. Adding an over-time perspective would make this even more powerful, as a lot of information is encapsulated in change, not just absolute observation.
I wonder how much of the required sensing tech is already available, but most likely not rated for medical usage yet.
Given the decades of research into an electronic sense of smell and the billions of $ it will be worth if we can get a electronic nose, why are we jumping the gun to the other things?
> a sensor to detect molecules for which the mammalian nose has evolved to detect.
This would be a modern marvel and so far impossible. It also wouldn't have to be mammalian, there's no reason to limit it.
> If we have an excitatory pattern, then most AI techniques should be able to detect it easily.
Exactly. I'm not sure why you'd need AI at all. It should just be a look up table.
An electronic nose is a big deal.
We are talking the ability to know what people or vehicles are carrying without a search, tracking people and animals over land, the ability to find every truffle in a field, it should help us create smells like 4D cinema, find pollution in water and air and land, track leaks, cooking, find animal and mineral resources, animal husbandry and medicine.
Perhaps all those kinds of sensors could be cleverly combined in ways that make them more feasible/reliable, not less.
If we limit ourselves to designing it with the same specific conceptions of smell that we have, we might be unnecessarily making the job harder for ourselves, or missing easy optimizations that would make it more effective.
Yes. The sensors are the problem. Processing the data probably isn't a huge job.
There was a startup trying to do this in 1999.[1] Cyrano Sciences, Inc. They got some Small Business Innovation grants.[2] They were trying to build a chemical weapons detector badge, an oil/fuel leak detector, and related devices. They developed a handheld unit, the Cyranose 320.[3] They were acquired by Smiths Detection (UK). They sell various devices for detecting explosives, drugs, disease carriers, and such. "Currently trialing rapid detection airborne COVID-19."[4]
[1] https://www.scientificamerican.com/article/how-close-are-art...
[2] https://www.sbir.gov/sbc/cyrano-sciences-inc
[3] https://spinoff.nasa.gov/spinoff2001/ps4.html
[4] https://www.smithsdetection.com/canary-biological-detection-...
It must've been smell, he sensed the deterioration from biochemical signals. Why he liked to be around death, I don't know, cats are strange!
I volunteered in the Steere House and met the cat, and my close family member was a higher-up in the org and personally witnessed the cat's abilities numerous times.
What the NEJM publishes is not always accurate or even true, there have been numerous examples of retracted studies so no need for the appeal to authority.
If it is true then someone should figure out the mechanism and prove it otherwise it might just as well be coincidence.
By the way, in medicine, you don't need to do a randomized controlled trial to be considered a "real study". Do you think the study that found out that HIV was killing gay men in San Fransisco, a case-series study without any trial or randomization, was not a "real study"? Case series and observations are valid real studies, though not gold standards for establishing broad truths.
If you get 20 Oscars and 20 normal cats, feel free to do the RCT, or you can take a moment and do some simple Googling before going full-on skpetic on a situation you don't know anything about.
What is funny is someone killed Oscar with a bed pan after the original story come out. Sure it also is probably not true (The story also sounds apocryphal), but we can pick our realities -
https://scienceblogs.com/insolence/2007/08/19/oscar-the-deat...
Source: I have a cat.
As a novel signal or biomarker, figuring out aromatic traces of diseases we didn’t even know about seems cool too. So in that sense this is interesting.
But no, diagnosing disease is obviously multimodal. And I’m not at all convinced this would even be the most important modality, even in the distant future.
"Riad Sarkis, a surgeon and researcher at Saint Joseph University in Beirut, is part of a French–Lebanese project that has trained 18 dogs. Sarkis used the best two performers for the airport trial in Lebanon. The dogs screened 1,680 passengers and found 158 COVID-19 cases that were confirmed by PCR tests. The animals correctly identified negative results with 100% accuracy, and correctly detected 92% of positive cases, according to unpublished results."
btw, 10% of passengers are covid positive, and i'd guess most are asymptomatic as most of them wouldn't be flying otherwise.
A mass spectrometer is not going to be able to distinguish a rich mixture of compounds. Nor can it distinguish molecules with the same elements.
But it's been done, and apparently might still be in limited use: https://en.wikipedia.org/wiki/Mass_spectrometry#Respired_gas...
That's why GC-MS/MS is a thing
...well, almost nobody. "You're gonna die because you didn't call your mother often enough."
The problem is, how do you properly incentivize patients, providers and payers to actually implement this? The payoffs are not immediate but are rather long term for everyone involved: it takes 6+ years to clearly see the significant mortality reduction payoff actually show up in the data. The downsides, on the other hand, are immediate (another trip to the hospital/imaging center, deaths due to false positives, increased costs, more work for clinicians).
For providers, provide the studies and incentives. Bonuses for lung cancer screening work, or % of patients screened (of eligible patients seen above age X, who haven't been screened, don't have a counter indication to screening, etc).
For payers, I think there's two classes of payers; long term payers like national health services; and short term payers like US health insurance where people can change insurance every year.
Long term insurers can (and should) be looking at their budgets and data (their own, but also studies like you mention) to find out what things to prioritize. Depening on your perspective, goals could be to maximize good years of life given the budget or to minimize future budget needs given the current budget.
Short term payers are going to have a hard time prioritizing long term issues. Reducing future costs for your insurance pool doesn't help you much when your pool can change every year. To incentivize this through economics, we would need to somehow increase the time exposure of the provider. Something along the lines of the insurance carried for a given year has partial responsibility for future years. Of course, that would be an administrative nightmare.
Imaging convenience is a bit chicken and egg.
Not sure how you get the bonuses you are describing to happen. It seems to me like no one really is caring enough to make it happen.
Edit: hastily posted an article without reading it... this is not correct.
There are no RCTs, but at this point it seems not likely to happen for mammography. Perhaps for DBT though. I am optimistic that DBT + AI assistance will produce significantly different results than screening using a single reader on mammography.
There are good arguments on both sides, but I think that we have a much better shot at turning the breast-cancer specific mortality boost into all-cause mortality boost by continuing to improve treatment, biopsy, screening etc.
It's also something of a UX issue. Nobody is going to write an app that rings the cancer alarm after one positive. People might write an app that sends raw results to your doctor or medical monitoring service, and you get an email requesting an appointment to discuss your latest status. And if your breath results say that you're gonna die in the next 10 days and there's nothing to be done (exaggerated for effect), well then... maybe it won't send an email after all.