Nice! Following your project!

Hi, where did you purchase the following 

LED EOLD-1550-5255mW 1550nm LED, 

Photodiode PIN 35PD1M-TO

can you provide more information on the prices and where you bought them from???

That's such a nice project,but this bothers me a little bit.By what means can you get the training datasets?

Have you looked into reflectance-based chocometers?

Your goal is great, your chosen path also. 

When it will work, try a crowfunding campain to spread it. Lots of people are in need of this kind of device.

Just a tech question : would it be an optional external connection for smart phone ?

Congrats for getting featured on the main blog.

Thinking about your project got me wondering if applying some type of balm (oil, hand lotion, moisturizer) to the skin on the earlobe would make it more transparent. Moisturizing creams supposedly cause the skin to absorb moisture, which might reduce opacity caused by dried skin.

Any thoughts?

Thanks Peter, actually the water is the problem. It is the problem every time you think in infrared 1um, 5-8um, beyond 14um, water absorbs ever. I already think about plasmons amplification, but it is not easy to do it at home at all. And nanoparticles could be toxic!

Congrats on winning the HAD interim prize!

Thank you Peter!

Thanks to the Judges!

Hey cool, I was actually working on a similar project, but have not advanced very far. I do however have a built working(in a sense, no determining glucose levels yet) prototype. Runs on a Atmega328p.

There is one factor I didn't see you mention. The amount of skin will vary the attenuation of the light skewing the volume of blood present between the sensor the and LEDs. More skin less blood, vice versa. You need to know the volume to get the mmol/l.

I found some research papers(Ill try to dig them up) on using a green LED to determine how much skin is in the way.

My sensor has 8 LEDs, 2xGreen(571nm), 1xRed(631nm), 1xIR(940nm), and 2xNIR(1550nm).

I was having a lot of difficulties getting stable results and I really think its clever software that is needed for something like this. Learning algorithms specifically. I may be moving away from this idea and onto a different technique altogether though. I didn't want to hijack your project, anyone working on this stuff is good news to me(Type 1 D here). Wanted to give some input and show a similar design.

Cool!! Maybe we can share some data and experiences if you wanna!

I mentioned briefly about the volume of blood and water, which I think is more important than tissue absorption. I plan to measure this by 850 and 940nm wavelength, like an oximeter does. I read about another project using red and green led also, but I'm not sure if this wavelenghts really matter. But, it's only a guess. As you say, only a statistical or a machine learning algorithm could solve this.

Maybe the instabilities of your results cames from signal/to noise ratio, which in this case is really low as the papers says, what type of ADC you use?

Formally, I work with Long Wavelength Infrared (LWIR), wherein the background and signal are comparable or worse and if you not use a lock-in or another trick you will see nothing. Because of that, I will try very long integration time first and lock-in techniques after.

Machine learning is the wrong approach. It's fine for choosing the parameters of an existing model, but from the text here I get the impression that you don't have an established model yet.

If you take lots of data and can't easily correlate it, you could put it up on Kaggle (http://www.kaggle.com/) and people will compete for "best algorithm". Those people are really sharp, and you'd probably get a lot of response for something like this.

It occurs to me that you can detect a person's heart rate by noticing the slight change in color of the skin in videos. If you use a similar method and detect the heart rate in your system, you can correlate the changes in glucose reading with the periodic changes in color coming from the heart pulses - there may be a statistical inference you can make from that that helps eliminate noise.

At the very least, if you go the Kaggle route be sure to put up all the information you have. If there is a correlation, people would pick up on it.

The way a pulse oximeter works is by looking at the whole system blood and all subtracting the "stagnant" things like skin, fluids, and anything else that will absorb some light. The way this is done is by looking at the pulsatile waveform that is created from the pressure change in the blood. Essentially what you said about the change in light on a camera, this is already factored into this method. Heart rate is a byproduct of finding oxygen levels. Oxygen levels are a byproduct of finding glucose levels.

I somewhat disagree on what you said about Machine learning.  Machine learning is meant to correlate data or classify a set of inputs. There are already methods to finding blood glucose levels accurately in the world. You can run the new system along side a continuous glucose monitor and get literally days worth of training data for an SVM or alike. You have your inputs and outputs laid out for you.

Kaggle is a good idea if you wanted to go that route. 

@marcelclaro - I am using a op-amp to a MCP3421. Probably not the best method, but I was going for some funding so I wanted to get a device together. My initial proposal to this project was to create something that did not give specific "values" of glucose levels, but more of a guideline for where the levels are. For example there are cases where you feel as though you are a hypo or hyper, but they are just false-hypo/hypers. Now say I'm riding my bike and I get my heart rate up, maybe get some adrenalin pumping, it can feel the same as a hypo - this can lead to a little anxiety about going low which can further give the effect of feeling low. So I have to stop my bike, get off and check my BG. Normally I'm not low and its just my brain thinking the symptoms are of me being low. I've essentially just wasted a test strip which is ~$0.50. Instead of doing this the devices could classify areas of Low, in Range, and High. I could check while still riding my bike. If the device says I'm in range, then I can relax and keep riding. If it says Low, then I can stop, check the actual value and see how much glucose I should take. Same goes for driving.

This doesn't just benefit active people either. The most beneficial reason is pretty much mentioned in your project. It could stop people from having to test in random situations. Sure they would still have to check before/after meals as usual, but it could eliminate "unnecessary" checks when you feel weird. Those cost savings would still be huge. Test strips, for the meantime unfortunately, are still the most accurate way of testing. This is due to the chemistry of how they work. 

P.S. I'm also working on some reverse engineering on continuous glucose monitors/sensors. The Dexcom system to be precise. 

I didn't know Kaggle, thank you for the advice.

@Sean Hodgins - I think MCP3421 is a good ADC but if you used only a op-amp and resistor for amplification, the real integration time was short. A Capacitive Transimpedance amplifiers could be more suitable for this application. The IVC102 (http://www.ti.com/product/ivc102) has the op-amp, transistors and capacitors in the same packcage, reducing the noise. I had good experiences with this component.

I agree with you, only a guideline is a huge advance. Maybe the requirement of precision delayed the FDA approval of bloodless devices (maybe a lobby, who can tell?). I'm very sensitive to hypos, so, for me, the most important is the night glucose level control. I have common overshoots at night or at early morning and it is really hard to predict, its depends on exercises, hormones and food consumption during the the day.

I guess is really hard to do reverse engineering on these continuos glucose monitors, maybe the most important components are ASICs not common ones. Personally, I don't like them because the attached sensor cause a extra pain and is annoying during exercises (and sex).

Oh for me, I would love to use a CGM. Costs don't allow me to at the moment. I have taken apart one of Dexcoms $700 transmitters that only last 6-8 months. It wasn't easy being completely encased in epoxy, a couple components went with it. Its just a TI CC2511. The sensors just measures current of an anode/cathode pair wire in the interstitial fluid. The rate of reduction is related to the amount of glucose in the fluid. That is why the sensors only last about 7 days, the reduction process stops working. Also with higher BG the sensors don't last as long. I want to make my own open source transmitter that would cost magnitudes less than theirs and have replaceable batteries, instead of someone having to pay ~$700 twice a year because their coin cell batteries ran out...

Have you seen this?

http://stephenblackwasalreadytaken.github.io/xDrip/

Or how about a closed loop system using a raspberry pi: 

http://diyps.org/2014/12/15/how-does-a-closed-loop-artificial-pancreas-work-when-you-diy-or-diyps-closed-loop-is-working/

Both are amazing projects. I'm developing some boards for the xDrip project to make it more accessible to people with less "maker" skills.

Ow! Cool projects!

It's unbelievable that they make money by this way! Judging by the Dexcom stocks price, they will continue doing this with support for a long time...

"I read about another project using red and green led also" which project is that? I'd like to know...

I attended a lecture in December where the developers created an ear sensor for measuring body information. I *believe* their blood sensor was looking at the ear canal and not the eardrum.

They were able to get pulse rate and pulse-ox directly. Then they noticed that their microphone could hear the heartbeat, and the difference in time of arrival of the sound versus the vascular pulse measured a ballistocardiogram.

Just thought I'd mention all this. If you have some spare brain power, you might want to look into measuring other things using LED light.

The lecture is below, the interesting bits start at about 11:00 into the video.

Also, if you hit a problem and want to talk to someone, I'll bet that if you asked Charlie Sodini (MIT researcher, lecturer from the video), he could put you in touch with researchers who have experience with ear sensor designs.

Just some info for your project, hope you find it useful.

Thanks Peter, It's a nice Lecture, very useful. I already seen some lectures from http://www.mirthe-erc.org/mirthecenter/research/ They use only mid-infrared, but it is the same philosophy

I've often thought that glucose monitoring could be effectively done using this method.

As a suggestion, if you can effectively monitor glucose, consider building/exploring a sensor into an earbud. You might get more consistent results from the blood supply in the eardrum than on a patch of skin.

Awesome idea! I don't know if the eardrum is thick or vascularized enough but I will investigate this method for sure, I already seen some experiments based on scattering, and this is an answer to my concerns about aesthetic. Earbuds are better than a weird thing in the earlobe

This is an awesome project! I look forward to seeing your project take shape.