Sleep 2.3 Piezoelectric vs Capacitive Accelerometers

Dr. Ahn compares piezoelectric vs. capacitive MEMS accelerometers in sleep actigraphy and their impact on movement detection and classification.

Video Breakdown:

00:00 - Introduction to Accelerometer Types

01:35 - Piezoelectric vs. Capacitive MEMS

03:48 - Why This Matters for Sleep Actigraphy

06:06 - Signal Preprocessing Capacitive MEMS data

08:06 - Processing Differences: FMpre vs. FMpost

All right, the next section is the piezoelectric versus capacitive accelerometer data.
Now, this is important because we need to talk about the hardware that makes actigraphy possible. The traditional classical version is the piezoelectric accelerometer. Now, these devices use a simple principle. You have a piezoelectric crystal, which is attached to the seismic mass, and then whenever there's a mechanical stress along the sensitive axis, the crystal or the piezoelectric element generates electric charge.
And when acceleration occurs, these crystals deform more materially, producing a measurable voltage that corresponds to the movement. And that's what these devices are measuring to assess acceleration of movement. So small movements generate small voltage signals, while large movements generates large voltage signals.
About 2010 or so, we saw a significant shift towards capacitive MEMS accelerometers. Now, these work quite differently. They use a movable plate that's usually attached to springs, with components, like shown here, that are positioned between fixed outer plates. This creates a capacitive relationship, or in this case, multiple capacitive relationships.
And as acceleration occurs, the distance between these plate changes, changing the capacitance. And the change in capacitance is then used to calculate the actual acceleration. Now, piezoelectric accelerometers have some advantages, and this is a table that I pulled from analog.com. It is generally has lower noise levels compared to the capacitive MEMS.
It also has a larger spectral range and is particularly good at larger and higher frequency ranges. However, the MEMS accelerometers are cheaper, have a longer battery life, and have electronic capabilities which make it readily easily integrated into electrical circuitry. One of this is the possibility to do a self-test so that it could sort of see if the component is still working, whereas that's not an option for the piezoelectric accelerometer.
They also have some integrative electronics in these MEMS accelerometers so that you can have digital outputs and interfaces which allow direct communication with microcontrollers or other digital systems. And so as a result, MEMS accelerometers have really sort of entered into the hardware space simply because it's better suited for low-power portable applications where space is limited. And so this is why you can appreciate that MEMS accelerometers has really taken over into the wearable space. And this is certainly true in the sleep actigraphy space.
The traditional actigraphy devices such as Philips and AMI Motionloggers both use piezoelectric accelerometers. But the newer contemporary devices, such as Geneactiv, Actigraph Leap, and Garmin VS4 all have the capacitive MEMS accelerometers instead. And this is just proof of that to show here. This is sort of the brochure that comes with Actiwatch, which is, as you know, no longer available. It has a solid state piezoelectiric component, whereas the contemporary devices such as Geneactiv has a MEMS device.
And this is also true for the newer ActiGraph, which also has the MEMS accelerometer hardware in it.
So why should this matter for sleep actigraphy? Well, there are two important differences in the way acceleration data is generated and formatted. First is the piezoelectric material only responds to changing stress.
So they cannot generate a constant signal response to steady forces like gravity. So there, as a result of that, there's no DC response while the MEMS accelerometer does. So in other words, you can assess gravity in MEMS accelerometer, but not in the piezoelectric accelerometer. The other difference is that because the capacitive sensors are smaller, you're able to compact or put three different axes into the same device.
So, capacitive MEMS have three axes, whereas the piezoelectric has only one. And you can see, obviously the differences in there, and you can also see the DC, or the direct component, um, uh, here which denotes gravity, and you can see, you know, when the person's sort of rate rotating their wrist, it's showing sort of changes in the values of the, the low frequencies here.
This is a crucial point that's overlooked because all our common sleep activity, actigraphy algorithms, such as the Cole-Kripke and Sadeh, were originally developed using data from piezoelectric accelerometers, specifically the AMI Motionlogger. And so this is the actual publication of the Cole-Kripke in 1992, and you can see that it used a Motionlogger again, and the Sadeh method back in 1994 also used the AMI Motionlogger.
I think what’s really important is that this is, just not a historical footnote because when we use modern capacity MEMS accelerometers with a three-axis data, we can't simply plug this data directly into these traditional algorithms. The signal characteristics are fundamentally different, requiring additional processing steps to make our modern data compatible with these classic algorithms.
So this is where this needs to be re-evaluated, this first step of how the acceleration raw data is transferred, transformed into activity count. So here are the steps that needs to be taken in order to do that. Now, um, so this is an example of raw data. This is, demarcated here as UF, UFX, UFY, UFZ.
What we need to do is to do two things: First, to eliminate the gravity factor, and you do this by applying a bandpass filter or moving,average, and you can see that that's sort of like the, the wandering baseline here. The other thing that we need to do is combine these three accelerations into one, and we do this by creating an amplitude or a vector magnitude.
So let's just say we take this raw filter and then we first apply a bandpass filter, and so we could do that here and which is transformed to F, F, X, Y, and Z. And so you can see this, this wandering baseline is eliminated, and you can see all these components. Then you could do step number two, which is the vector magnitude.
You get what's here called FMpre. And you get this type of looking data. Now let's just say you do the vector magnitude first and you get UFM. And then afterwards you apply the band pass filtering. You get a completely different, or you look, it looks very different. than the FMpre. So this is why, sort of, in the options for the Garmin Enhanced Actigraphy, you have the option for FMpre and FMpost.
These two provide very different results, and it's the order by which you are doing these filtering or this processing that matters. So, whether you do the filtering first, in FMpre or filtering afterwards and FMpost can really change results. Now, here you could kind of see here that, the gravity vector magnitude is one.
And so there's another option is just to simply eliminate, this one, establish zero and then to create an absolute value. And so there's another option here, which is called UFNM.
So the question is, which one should you use? Should you use FMpre or FMpost? And I wanted to go a little bit of the intricacy of this because it gives you an idea of this is not being really that straightforward here.
So, this is, an individual ,I have to thank Dr. Huang who is part of our team. He placed an ActiWatch on, while also placing an MEMS accelerometer on the same wrist, and you can see here that, you know, how this data sort of looks very similar to this, and what we do is that we can apply the FMpost, and we get this result.
Now, what if we do the same thing, but we do FMpre. So instead, what we do is apply a filter first and then apply the amplitude vector, uh, or, and then you can see again that the results are different. First of all, the FMpre tends to all be positive because the act of this vector magnitude makes everything positive.
The other thing is these small motions that you see here are shown up here but not shown up on the FMpost. Now why is that? It's because the FMpost has almost a preferential bias towards the axis that is more oriented towards gravity. So you could see here that, you know, the value for the z axis is 800 and to 900.
So it's almost like because the watch is positioned like this, it's almost like your wrist, is positioned like that. And so what FMpost does is that it sort of, it sort of, biases these motions that you see, in this z axis because it is aligned with gravity and it shows up here. Whereas the motions that you see in the ones that are in the other, axes are not appearing here. And so, that is an important differentiation, between the two.
This is just another example, again, courtesy of Dr. Huang, also part of our data science team. He actually wore these, devices while he was asleep. with the lights on, with the video recording. So really amazing work by Dr. Huang. You could see here that, um, you know, have this particular movement. It was a very subtle movement where the finger moved along the z axis. And what you could see here is that, basically, a lot of the movement is here and in FMpre as well as inFMpost, you can see that it is, identified.
However, during the twitching, when the finger twitches along the y axis, you see it still in the piezoelectric and then you also see it in the three-dimensional MEMS, but then the process of filtering, the FMpre sees it, but the FMpost attenuates it significantly because this movement is not aligned along the gravity vector.
So I think there's some important differentiation. The piezoelectric accelerometer is sensitive along all the X, Y, and Z directions, whereas the FMpost is sensitive in the gravity direction. FMpre is sensitive to all three directions. Another interesting thing is, you know, there are some movements such as rotational movements.
This is FMpost processing as the wrist is changing along its axis. And, and you can see that this appears here. However, in FMpre,I mean, sorry, so the FMpost sort of doesn't denote this rotation that well, but whereas FMpre, you see the rotation much better. So in summary, FMpre is sensitive to small motions, including rotational movement, whilst FMpost has a directional bias along gravity vectors.
So you could potentially use this to your advantage if you wanted to, to sort of delineate sort of the different type of activities that your study participants may experience. Now this difference between FMpre and FMpost actually gets even more complicated when you're considering the activity count transformations.
So let's just say for instance here the z, you are taking zero crossing modes to the FMpost data. And you can see sort of the FMpost data here in blue, the zero crossings marked as green crosses in the threshold, and you can see that the final, zero crossing counts matches pretty well with the actual magnitude of the acceleration.
However, if in FMpre, When you apply the zero crossing mode, you can see the magnitude increase here, but because these values don't cross this threshold value, your activity count, or your zero crossing, can actually falsely be low. So, when you're doing, when you're applying the zero crossing mode with FMpre, it really is something that you should be cautious about.
And when you were talking about energy level, they both work pretty well, but again, FMpre is, as mentioned before, more sensitive to smaller movements, so then the values overall are greater for energy or PIM, and also, the same for time above threshold, that FMpre is more sensitive to small movements.
This is an example of, you know, an actual data of a person walking at a particular pace and you can see the FMpost in green and FMpre in orange. And again, as mentioned before, at least for FMpre when it comes to zero crossing mode, you have a falsely reduced level. And so you could see this in this setting.
In FMpre during energy or proportional integration mode, FMpre is larger in FMpost, and the reason it's larger is probably because you have more horizontal movement. Now remember, FMpost is more biased towards movements that's vertical or aligned with the gravity vector, or there may be a rotational movement.
So you could sometimes see this difference and then differentiate the type of movement that you're dealing with here. And then you can see FMpre saturates pretty quickly pretty much here.
And so in summary, FMpre is sensitive to small motions through rotational movements. It can be relatively incompatible with zero crossing mode and easily saturated with time above threshold.
FMpost has some directional bias along the gravity vector and so there's some, metric compatibility here in terms of the type of acceleration signal that you get and the activity metrics, and, you know, this is something that we need to consider, as you're doing your study.

Course episodes

Sleep 2.1 Intro to Sleep Actigraphy

Sleep 2.2 Traditional Signal Processing: Sleep-Wake Classification

Sleep 2.3 Piezoelectric vs Capacitive Accelerometers

Sleep 2.4 Signal Processing Configurability: Parameter Settings

Sleep 2.5 Unexplored Research Opportunities