Oh yeah and you need a full laptop to do the same as the “stupid iphone”. And that stuff does not look wearable. Also the MBA morronspeak you didn’t seem to understand explains they’re actually evaluating compression algorithms: that compressive sensing stuff was recently featured in wired and looks quite cool to me. If you throw in a more efficient communication layer, you can only improve on the system after compression. And compression remains key …

My tip was explicit about the fact that what is undertaken here is an instance of compressed sensing. To make it short, the reason the draw on power is expected to be low is because the encoding of the signal is done following the theory of compressed sensing. The reason the iPhone is important is because the signal being compressed during the low power acquisition (because of compressed sensing) is then decoded on the iPhone which performs some real computations involving full scale optimization routines with wavelets. The iPhone is central to this set-up because it is used as a full scale computer, not just a screen.

Compressed Sensing is a paradigm shift because the encoding is very low power and most of the work is really performed after the signal has been acquired not during acquisition.

Igor.

I think the iPhone version looks a little hokey and contrived. Is there any legitimate reason to need an EKG (not EEG – no brainwaves here) on your phone?

If you’re a doctor surely you’ll lose your license for basing medical care on untested BS equipment like this.

M4CGYV3R,

The iPhone is really needed here for its computing power. The fact that it also has a nice ergonomic factor helps in making the point that the system is already miniaturized using off-the-shelf hardware.

Igor.

Goldscott,

You are right, the compression scheme is very simple yet efficient. For some numbers check slide 32 of the presentation I mentioned to rasz above (http://lts2www.epfl.ch/~vandergh/Download/Marseille-2010.pdf ).

As for the bluetooth use as opposed to other means, I wouldn’t know. I am sure they wanted to use an off-the-shelf (full scale) computer such as the iPhone in order to show the reduced complexity of the overall prototype. But as I said, I don’t know about that part.

Hope this helps,

Igor.

Goldscott,

You said:
“I just looked at the slide show. I’m not familiar with much of the numerical analysis (e.g. Gabor, Rudin, Osher, Fatemi, etc.), but I took a short graduate course on compression, so I see a small dictionary is being used with DWTs. I also don’t have any experience with sparse matrices, but it seems like the “sparseness” is being highly compressed and the “meaningful” data is getting a DWT run over it?”

Not quite but close. In the generic encoding scheme one would perform an encoding at the sensor level with the DWT dictionary. In the CS case, the acquisition is performed with a “sparse” dictionary that unlike the conventional version does not require iterations (no need for finding the best element in the dictionary as in the conventional case). However, in the CS case, the DWT dictionary is used on the receiver side with the iPhone, as it is used to combine the encoded data and the dictionary to provide the original signal.

In the conventional case, the compression effort is spent on the sensor side while in the CS case, the effort is spent on the receiver side.

“Also, what is the icyFlex2 and why haven’t I heard of this chip? Proprietary development at CSEM?”

I really don’t know about that part, sorry.

Igor.

rasz,

You said
“..Basically using Bluetooth for talking to iphone zeroes any advantages CS might have.”

Yes they are at that stage it seems but as I have said, the iPhone as a computer is a terribly important part of the whole set-up.

You also said:

“..Last but not least I dont think compressing EEG/EKG signals would be advisable when raw data is almost free. I can understand using CS in MRI where it offers faster acquisition times..”

You are making several assumptions here:
one is that this is the best CS can do. My hunch is that they can do better. They are also interested in changing the A/D converter at the sensor level, with what we have seen in CS,I am thinking we are going to see improvement there as well. Finally, you are making the assumption that one cannot make some observation about an EKG from its compressed data. Unlike conventional systems, CS has the particularity of providing a sound theoretical background that says otherwise. In short, while it is nice to see an EKG traces being reconstructed on the iPhone, one will eventually be capable of making an actual diagnostic from the compressed data before they are being reconstructed. All in all, I agree with the PI of the project that this prototype is likely to provide much better outcome in the future.

In the case of MRI and CS, the hardware was already there, so not much convincing had to be made at the hardware level: the procedure for acquiring data just changed. In most other cases including compressed sensing, we have to change the hardware. This is why CS always look like a small improvement at first sight (or a solution looking for a problem) because you are generally using off the shelf hardware that was not designed for the new purpose at hand (BT in this case). If you are interested, I have ompiled a list of CS hardware being built: https://sites.google.com/site/igorcarron2/compressedsensinghardware

Hope this helped,

Igor.

I would think that EEG waveforms would be trivial to compress, and bt sniff mode would give you very low power consumption (wake up every few seconds to burst the data).

Where are the schematics/notes on the EEG hardware itself? Is there anything interesting or unique about the actual data acquisition?

Rasz,

You are missing the point.

There is really nothing to optimize for CS. In fact, optimizing for the DWT scheme is far more expensive than the simple multiplexing occurring for CS. This is so because you have to spend much time and money trying to figure out how to design your sensor to perform the compression in the right kind of basis. In CS, you don’t care about these bases issues until you reconstruct the signal (i.e. much later)

There is also no rule for CS being 1/8 Nyquist, in fact, the CS rate is really dependent on how sparse the signal is. The sparser the signal the lower the rate. In fact, the specific case shown here is valid for one ECG electrode. The CS rate decreases when we add several electrodes as most of the other electrodes share some part of the signal together (joint sparsity). The overall effect of adding additional electrodes becomes more obvious with the utilization of CS.

Finally, you also make the point that Nyquist is free. Maybe so but if you have to devise an AI system on top of this sensor in order to produce some type of alert (as you know some people should have an ECG all day long for some diseases), then you are going to devise an algorithm that classifies this signal. And clearly you want that signal to be as small as possible so that any SVM/Classification algo can be fully tested. The Nyquist rate might buy you the ability to reconstruct perfectly the signal (without any notion of sparsity) but it will be a hindrance for classification purposes (especially with several electrodes) and will need to be done off-line (away from the sensor). On the other hand, CS buys you the ability to, for little additional power, perform classification on-the-fly, next to the sensor.

I hope this helped,

Igor.

Andrew,

Check the accuracy in the slides I pointed to earlier in this thread.

Pseudoscience ? wow.

Anyway as I said earlier, part of the idea is to eventually have a classification system as close as possible to the sensor. Another is to have many more electrodes than just one. A Nyquist rate signal could not possibly allow for classification without a bulky and power hungry overhead and would eventually be an issue for a larger set of electrodes.

>Compressed Sensing is a paradigm shift because
>the encoding is very low power and most of the >work is really performed after the signal has been >acquired not during acquisition.

As a paradigm shifts go, this one needs a cane and a hearing aid. Unless you think that compression after data acquisition and post-processing on the other side of the link after decompression is “new”.

The exact same paradigm shift occurred when someone discovered that you could use motors to play tapes of very long and complex telegraph messages in high-speed bursts, for decoding on the other side using the reverse process. Of course, that was circa a few years before WW1, so you’d hardly be expected to have seen your paradigm back when it had a full head of hair and could still get more women than Maurice Chevalier.

And, you probably missed it again when that Turing fellow was trying to bust a move before the big apple, when we used computers to decode radio transmissions. Or when we needed to decode pressure wave data while we practiced giving cancer to half of Nevada.

Like a bad movie vampire, your paradigm hung out with Von Braun and those silly rabbits at Lockheed, before graduating to doing secret squirrel spy stuff with bubbly torpedoes and bouncing recon images off satellites.

But like shag carpet, your paradigm will be back on the scene faster than lady gaga can peel her upper lip clear of hair, and so I would like to congratulate you on your clever iphone hack.

But in the meantime, go buy an ipad so you can find an appropriate audience for your hype. They’ll appreciate it. They dig anything.

Andrew,

It’s good to have BS detectors up.

I think the reason you don’t see traces and reconstruction is really that the slide were focused for a different audience. But in CS, the error is well controlled based on the information one has on the sparsity of the signal. I am sure that in the future, the authors will also show real and reconstructed waveforms (I am sure they will read this thread at some point in the future) On slide 30 of these slides:
http://lts2www.epfl.ch/~vandergh/Download/Marseille-2010.pdf

The good and very good limits are judgements made on how the data compares to the uncoded one. The DWT scheme (conventional compression scheme using wavelets) is the best one available, while the CS scheme is worse, but the time needed to run the encoding required for the DWT is 5 times more than for CS. The expectations would be for the encoding power level to also be in that ratio but the total power number given is only slightly giving an advantage to CS (must be the radio).
The point of the presentation is to show preliminary result and the need to know what you are doing when implementing this CS technique. Unlike “Fractal Compression” :-), CS is a different way to perform sampling. It is known that it is not the cure-all solution but rather a different way of doing things. The community at large is still trying to figure out situations where it brings an enormous advantage over current undertakings. MRI is clearly a good example where no hardware change was required that benefited from it, other areas are taking a little longer as hardware needs to be changed (the project shown here is a good example)

So I think the advantage is really that if you choose a simpler encoding hardware as well as a better radio you could really see CS shining through over both no encoding and the lossy DWT approach. The approach will eventually scale very well with multiple sensors as CS will provide additional theoretical arguments for higher compression. Eventually, the thought is that several sensors and a lightweight classification system are the targets to aim for. Again this is a work in progress.

As an outsider to the project, the most surprising to me was the fact that the iPhone could reconstruct the signal in real time. See, up until three/four years ago, this is a feat that just could not be achieved, see for instance the second figure in this blog entry I wrote a while back on the performance of reconstruction algorithms:

http://nuit-blanche.blogspot.com/2009/12/cs-are-we-hitting-ceiling.html

The convergence of tremendous algorithm improvement in the past few years and the rise of small handheld computers like the iPhone makes this ECG system a very unique project

I hope I answered some of your questions, for more, you guys can ask directly the folks working on that project (icyflex,…). As a work in progress, I am sure they will update us on their next iteration which will make the CS system a more obvious winner to different trade studies.

Cheers,

Igor.

Andrew,

I have responded to you but the comment awaits moderation. Sorry.

Abe N.

I am not involved in the project.

Cheers,

Igor.

Kris,

I am not sure how to answer that question but in the DWT case, teh reconstruction involves one matrix vector multiplication, whereas in the CS case you have at least an order of magnitude more matrix vector multiplications.

Hope this helps,

igor.

Wier cool!!!

Wim,

The point is that power level for compression in CS is expected to be very low compared to conventional compression (DWT). So the idea is to compress using CS techniques (lower power than conventional systems) and then send as few as possible information over the wire (because the information is compressed i.e. as opposed to no compression).

Where it gets interesting is that CS is not an optimal compression system compared to DWT. In effect, more data has to be sent over the wire with CS than with DWT. DWT on the other hand cost five times as much energy as CS for the compression step.

At the hardware level, the compression step is very different for CS and DWT, one is not just a by-product of the other, the architecture has to be changed. Hence the trade studies…of which this project is part of.

Hope this helped,

Igor.