[george88b]

As an analytical chemist also with a degree in biochem who has worked extensively on blood analysis, this will not work at any meaningful rate of reliability. I sometimes get frustrated when the key part of some great breakthrough is hid behind a buzzword as if its a cure-all for the details. I would love to know how "machine learning" is going to just magically make this work at a reliable rate. I guess I am just a pessimistic lab rat.

[gone35]

Indeed it won't. Microscopy is hardly the bottleneck in hematology; and neither is cell counting, which is already carried on by automated analyzers for the most part anyway [1,2]. The problem is the need of differential staining: due to fundamental physical limits, no amount of machine learning can ever distinguish key hematocytes like lymphocytes from granulocytes in raw, unstained samples from microphotographs alone. So unless you use spectroscopy --and there's been some work done on that, eg [3,4]--, you need to spread, fix and stain your sample, each of which take a series of choreographed steps, reagents and considerable skill in controlled conditions to get (minimally) right [5] --hence the need for a lab.

So unless they attached a USB microspectrometer to the iPod, or streamlined the existing sample preparation process in a low-cost, fully-portable form; they are just solving the wrong problem.

[1] http://www.mlo-online.com/articles/201401/automation-in-hema...

[2] http://www.ncbi.nlm.nih.gov/pubmed/18550479

[3] http://www.opticsinfobase.org/abstract.cfm?uri=FiO-2008-FWD5

[4] http://cancerres.aacrjournals.org/cgi/content/meeting_abstra...

[5] http://mmserver.cjp.com/gems/blood/lh.6.1.houwen.pdf

[ihnorton]

The point that you miss is that microscopy is still a bottleneck in developing and rural areas. H&E staining is not that hard, but you are right that maintaining reagents and a clean lab is a logistical challenge.

Still, a lot of smart money says the smartphone-pic-to-clinician thing could have a big impact. See Foldscope, for example, which takes this idea to the next level: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjourna...

By the way, nice to see a Raman shout-out! Here's a slightly newer Raman paper with some nice pictures (compulsory open-access for the win): http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3806096/

[gone35]

You are absolutely right: the fact that microscopy is not sufficient doesn't mean it's not necessary in the first place. Besides, a portable microscope would definitely be a boon for other types of clinical work like urianalysis --as elij pointed out-- or plasmodia detection (for some species, at least). I stand corrected.

[elij]

Microscopy is also the limiting factor in _developed_ areas when it comes to urinalysis which I'm sure this solution can easily be adapted for.

[richardbrevig]

I was wondering about the need for staining. Is there potential that they could stain the blood sample and potentially get a result? I'm speaking relatively hypothetically.

This is at least more in the right direction than what I had seen previously.

I saw much worse a few months ago at a competition I was in. The winning team "created" a device (that looked like a USB key). They claimed that if you had a sore throat you could take a sample with a q-tip, insert into the device, and it would magically determine the presence of an infection. Those were their words. I was horrified and when I approached the organizers afterwards they didn't understand my explanation on why it was not possible. Indeed, after that time as I have spoken about it most people do not understand that it's not currently possible. Sci-fi blurs the realm of possibility for many and it seems reasonable to them. Back to the actual contest, mine was an "idea competition" and not a YC Hackathon.

Tanay's idea is leaps and bounds closer to the realm of possibility than the idea behind the other team I witnessed. For that, his age, and his other work on his startup clipped.me, I congratulate him and look forward to seeing him come up with something truly useful in the future.

[ihnorton]

The Oxford Nanopore Minion is "usb scale": https://www.nanoporetech.com/technology/the-minion-device-a-...

Not quite built in a weekend, but certainly within the realm of possibility.

[Brakenshire]

Just for context, Oxford Nanopore have struggled a fair bit with getting this device to work. They have recently released quite promising prototype devices, but the project is many years behind schedule. Albeit that isn't surprising given what a massive shift in technology it is.

[ihnorton]

Oh, sure, but no one thinks they are violating the laws of physics anymore. It's "just" engineering now to get the accuracy and sustained-read lengths high enough.

[vobios]

Although many years behind schedule, it's still far more advanced than any other sequencing technology on the market.

[gone35]

Yes, I think so. As others pointed out earlier, portable devices with immobilized reagents like pregnancy tests or glucose monitors already show the concept is (in principle) feasible; and there's been a lot of recent work dedicated to miniaturizing and integrating all steps of critical assays, including even the most elaborate ones [1]. It's pretty impressive stuff, so I expect a breakthrough anytime soon.

I definitely agree though it's a commendably well-thought-out project in itself, especially for the usual techno-bubbly standards of Silicon Valley hackatons of late.

[1] http://www.springer.com/cda/content/document/cda_downloaddoc...

[vobios]

Why is that not possible? A standard home pregnancy test is small, takes a few minutes, and is fairly reliable at detecting human chorionic gonadotropin (hCG). If you can find an antibody specific for a particular infection, you can make a similar test. It's mostly useless, but certainly possible.

[richardbrevig]

If you're asking about the competition I was in, with the "miracle device," I forgot to mention that they gave multiple examples of what it could detect. Sore throat, ear infection, etc. The pregnancy test you mention is determined by a single use hormone detection. It's a chemical reaction dependent on one specific hormone being detected, not multiple hormones or multiple antibodies. For comparison, a urinalysis strip has multiple markings on it each for a different test. And they're not reusable. Whereas this just "magically" knew that you had Streptococcus pneumoniae or Haemophilus influenzae present--no demo, no explanation of how.

[bermanoid]

Can you explain the actual difficulties?

In particular, do you question whether the quality of the image would be high enough, or whether the ML techniques can automate what a lab tech does while looking through a lens, or is the problem that seeing blood is not enough to diagnose much of anything with any certainty?

"Reliable rate" is relative, and something that I've found lacking in modern medical care in the US even when it's a dude in a lab coat looking at samples through a state of the art microscope...

[thecritic1729]

Yep, it was mentioned that they do it with 75% accuracy. That's too big of an error margin for it to be used in production.

[gotrecruit]

my shallow knowledge of machine learning tells me that the idea behind it is that initially it will suck at diagnosis and analysis, but over time the algorithm will learn and improve to a point where eventually it becomes actually good at it and even exceeds human capability.

[john_b]

Machine learning, in this context supervised machine learning, is a useful tool for deriving unintuitive relationships between different parts of complex data sets. To do this, there must be some discernible correlation between the parameters of interest that isn't subsumed within the noise of the system+measuring device(s).

In this case, those parameters would be the image data and whatever health parameter is of interest (e.g. white blood cell count). My initial skepticism, perhaps that of the parent comments as well, has more to do with whether the measurements are of high enough quality for any reliable analysis to be done. The app doesn't seem to require any background or contextual data either (though I haven't verified this). If not, false positives and negatives could be problematic.

Anyway, machine learning isn't a form of magic that can transform data with no meaningful sensitivity to something into a something that is sensitive to it.

[hiddencost]

That's a dangerous way of thinking about ML. Models aren't magic, they're a approximate hacks that end up working for a specific instance of a problem.

More data is always nice, but typically you see accuracy level off (diminishing returns). ML is a constant process of improving your data, increasing the amount of available data (not the same as improving your data), improving your features, and improving your model. No one thing is sufficient.