Sleep 2.5 Unexplored Research Opportunities

Alternative feature extraction methods, gyroscopic data, posture tracking, and spectral frequency analysis to improve sleep measurement accuracy.

Video Breakdown:

00:00 - Introduction to Underexplored Research Opportunities

00:37 - Feature Extraction & Algorithm Impact

03:25 - Gravity & Posture-Based Approaches

06:47 - Temporal Information & Future Innovation

Okay, so the final category for this talk is going to be sort of underexplored research opportunities. And again, just to let you know, this algorithm that has been pursued in traditional methods really is a broader algorithm that follows a machine learning approach. and so there's no reason why we can't expand the possibility of feature extractions and also the type of algorithms that are out there.
Sort of a case like similar to some of the machine learning approaches that are out there. So this was a publication that was published in 2021 and what they did was actually took a number of available, feature extractions, not just the ZCM, PIM, or TAT, but almost like 30 plus possibilities here.
And they used both acceleration as well as gyroscopic information. You know, some examples here are like skewness, kurtosis, mean, or variance. And so you could do a combination of these things, and you could also use each of multiple of these things. and then apply different algorithms to each of these feature extractions so that you get multiple possibilities.
You can see here that the choice of your features does matter. and so you can see these are the features on the horizontal axis. The J is a marker of how useful or informative this data is. So it's basically sensitivity plus specificity minus one. So it's better, to have, higher values.
And you can see that, for instance, this gyroscopic mean did really well, whereas the skewness of the gyroscope did poor overall. So the choice of your feature matters, but also the choice of the algorithm upon which you use these features also matters. And so, in this example here for GP75, GP75 works well with this hidden Markov model.
But doesn't do well with this linear discriminant analysis. And interestingly, I think, um, according to this study, the, you know, the feature that worked best for the algorithm, was, were these features. And all of them were, interestingly, sort of the gyroscopic. So, gyroscope measures rotation, and acceleration does much more sort of the principal axis movements.
So, there are probably some rotational information that is really missing in our present methodologies that are sort of worth exploring for the future. As I mentioned, sort of the Sadeh method. So the Sadeh is the, the one that actually performed the best here, out of all the five, mechanisms, with an ROC score, which is closer to this square here.
And, you know, I think because the Sadeh method is just more nuanced, it sort of has standard deviation, mean, sort of intermediate levels. I think that's why it did a little better overall. We're able to extract multiple features simply because our Labfront platform does allow you to extract raw data.
Yeah, I, I think the most we can do is 25 hertz, actually, because when you get to 100 hertz, the amount of data is just way too much, and when it comes to data transfer and the amount of memory storage that's available on our watch.
Now, the other part that I wanted to mention about is the components of the fact that we are transforming our three-axis data that you obtain, from the capacitive MEMS to the piezoelectric model.
Now in our adherence to the traditional actigraphy methods, we're really discarding two valuable types of information provided by the modern MEMS accelerometers. First, we're losing detailed specific movement, data capture, and momentary acceleration changes. So, for instance, if, you know, for, because the z axis is perpendicular to your wrist watch face, the perpendicular, the z axis will reveal momentary changes, uh, Z axis will show that the action's like this.
If it's a Y, it's this. X, it's this. When we're combining all that information into one acceleration, we're really losing that information. The second thing that we're ignoring is the directional context, from the gravity factor that could tell us about the posture. So you could see here, it, you know, tells us how our, posture our arm is in position in relative to gravity. And again, in our process of really trying to, match the traditional actigraphy method, using that information.
What's really surprising to me, I haven't really done some extensive literature review, is how little research has explored utilizing the full potential of this three-dimensional or gravitational information available from the capacitive MEMS accelerometers.
We've essentially been forcing our rich, modern data. into the constraints of traditional actigraphy, like fitting a square peg into a round hole, rather than innovating new approaches that could leverage all this additional information. And I've really only found few studies that have taken advantage of this information.
This was a study that was published in 2018. Basically using an arm angle relative to gravity to create, a specific algorithm, which they've designed here. and, you know, I think, according to this, this heuristic algorithm, which is denoted, by blue,did, uh, better job when it, when it compared to sort of the sleep diary.
It had more of a Gaussian distribution, whereas the traditional methods, as mentioned before, it sort of miscalculates the times that are sort of quiet, wakefulness when there's still wakefulness is misinterpreted as sleep. So, you know, I think this algorithm, which it considers sort of the gravity or the posture movement actually did a better job.
Also you could see that the mean absolute error was better overall using this type of information compared to the traditional activity methods.
There is another study that also sort of used a different, approach. This, study actually used a device that had an IMU, or initial,motion units in both the, the, hand and the forearm, and then they could use information about the relative positioning of the hand relative to the form to determine the types of postures and poses that the child is taking during sleep.
And again, this all incorporates the rich 3-axis data that you get on accelerometers that are missing in our traditional actigraphy approach.
The other thing that we're missing when we're applying our, sort of standard, traditional actigraphy methods is a loss of temporal information. What we do is we basically take each epoch and then we're basically shrinking it down and we're just literally obtaining a sum of our feature extraction that we're focused on, whether it's zero crossing, TAT, or energy level.
And the order in which the acceleration happens, like the temporal information itself, is lost. And there are various ways that you can capture temporal information. One of that is actually obtaining spectral frequency or applying sort of like something like a Fourier transform or a wavelet transform to the actual acceleration data and to get sort of something, what's called here, like a frequency time window here. And you could see the type of frequencies that are sort of brightest, at each time point. And when you incorporate this spectral information, compared to just standard low resolution, resolution acceleration or using typical features that you use traditional actigraphy, you find that the traditional approach doesn't do as well in, multiple data sets that they approach here.
I think TBI was a, you know, sort of traumatic brain injury data sets. This was a healthy data set, but you could see that when the information, temporal information is incorporated here, it does better overall. And so, we are losing temporal information when we're taking a, taking an approach. and so, that's kind of the, the, the things that I think are missing components in our present approach to sleep research.
We are, again, open to collaboration. We are trying to enable these tools, including sort of raw data acceleration, so that we can really move the boundary forward towards exploring new algorithms and things that haven't been explored before, so that we can really improve our assessment of sleep in time.

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