Sleep 2.2 Traditional Signal Processing: Sleep-Wake Classification

How raw ACC data is transformed into sleep insights, covering feature extraction, sleep classification algorithms, and our automated actigraphy reporting.

Video Breakdown:

00:00 - How Raw ACC Data Becomes Meaningful Sleep Information

01:01 - Feature Extraction: ZCM, TAT, and PIM Explained

05:50 - Sleep Propensity & Traditional Classification Algorithms

09:07 - Refining Sleep-Wake Classification for Accuracy

10:40 - Labfront’s Automated Actigraphy Reports & Future Enhancements

The second section I wanted to talk about, was traditional signal processing, and sleep wake classifications.
So here are the key steps in how raw acceleration data becomes meaningful sleep information. The process follows a clear progression. Raw acceleration data is first transformed into activity counts, which are then used to calculate sleep propensity, and then finally converted into sleep wake classification.
And you can imagine it's like a cascade where each step builds upon the previous one. And the approach resembles closely to broader machine learning strategy of feature extraction, algorithm implementation, and score refinement. And understanding that these intervening steps is really crucial because each transformation of points presents an opportunity for optimization. But also for potential error as well.
So let's focus on the first step, which is the feature extraction. So, traditional sleep actigraphy uses three main approaches to derive activity count, um, from raw data. First, there's zero crossing mode, or ZCM, which essentially counts how many times the signal crosses a set threshold, giving us a measure of movement frequency.
The second is the time above threshold, or TAT, which sums up the total duration, which is in seconds, when movement exceeds a certain level. And you can see that sort of demarcated by this threshold, and this is the total amount of time, that's accumulated over, uh, regarding the time above threshold.
And this gives some information about duration. And finally, we have Proportional Integration Mode, or PIM, which calculates the intensity of movements by integrating the area under the curve. And each of these methods captures different aspects of movements, and this includes the frequency, duration, and amplitude, respectively.
So, how does Garmin specifically implement the zero crossing mode? Their algorithm follows a pretty straightforward but precise rule. A zero crossing is counted and given a value of one only when the two conditions are met simultaneously. First, the time series must change sign between consecutive data points, and second, the acceleration value must exceed a specified threshold.
And so you can see this as an example here. This was marked as a ZCT of 1. It's because the value right before it had a different sign, and then the value itself was greater than the threshold. Anything not meeting these two conditions was given a value of 0. And this dual criteria, helps ensures that not only meaningful movement transitions are captured, but filters out some minor fluctuations that might not represent actual physical activity.
So why is there a threshold value? And this image illustrates why we need a threshold value in zero crossing calculations. Now the term zero crossing Originally came from counting when the signal crosses the y axis or the zero point. But the problem is that all accelerometers have white noise, which is shown here and zoomed specifically here.
The counterintuitive nature of zero crossing is that during periods of minimal movement, the white noise actually creates more zero crossing than periods of active movement, which occurs here. So you can see here, this is the total window zero crossing and then actually there's a little bit of a lag.
So this may probably correlate more with here. Because of the white noise, it's crossing the zero multiple times, and you can see that the zero crossing mode without a threshold is much higher. However, when there's significant movement, as you can see in the acceleration here, the zero crossing mode actually decreases.
So, you know, without a threshold, you get the paradoxical result of detecting more activity during still periods than during actual movement. This is why the dual criteria is really important.
Garmin's time above threshold implementation is really simple. When acceleration exceeds the set threshold, the TAT value is assigned 1 over 25 and this fraction isn't really arbitrary. It's directly related to the 25 Hz sampling rate. Effectively converting each above threshold data point into a time measurement of 1 25th of a second. So obviously, if you change the frequency sampling rate, to like 50 hertz, then this would have to be 1 over 50 seconds.
And then Garmin's approach to proportional integration mode is also straightforward. It simply sums all data points whose absolute values exceed the threshold. And since we're measuring energy here, the PIM measures are expressed in milligees, representing the intensity of movement.
An essential aspect of processing actigraphy data is that these activity counts aren't simply summed continuously. Instead, they're aggregated within specific time windows called epochs. In sleep actigraphy, these epochs are typically 30 or 60 seconds long. And this window sums, applies to all the activities that we mentioned previously, including ZCM, TAT, and Energy, which is the same as PIM. Note how the units are completely different for each of the three, settings, so ZCM is in counts, TAT is in time, um, and PIM or Energy is in milligees.
All right, so, now, sort of understanding how activity counts are derived, we can now talk about the next crucial transformation, converting the activity counts into sleep propensity scores.
And this is where we encounter one of the more striking examples of stagnation in activity research because the methods for this transformation have remained really unchanged over the past three to four decades, or three decades and there are several commonly used algorithms in sleep actigraphy.
The key insights behind these algorithms is the recognition that sleep has a physiological inertia. So we don't typically flip between sleep and wake states instantaneously. And this is why simply applying a threshold to the activity count itself wouldn't work. It would result in unrealistic sporadic wake periods.
Instead, algorithms such as Cole-Kripke, which was introduced back in 1992, uses a moving average approach to smooth out the sleep propensity over time. Specifically, it takes weighted values of the zero crossing mode score across seven epochs. To arrive to a sum, or S, which provides sleep or wake propensity, or actually in this case it's actually wake propensity, because the higher this value, the more, like, more likely that you're awake.
So a value that's lower than 1 indicates that you're asleep, while a value greater than 1 suggests awake time. Now, there has been some modification to this approach, including this particular algorithm which was published in deSouza in 2003, but the core concept has remained largely unchanged since its introduction three decades ago.
The other commonly used method is Sadeh algorithm. which was introduced just two years after Cole Kripke in 1994. It offers a more nuanced approach to sleep propensity calculations, but also uses zero crossing mode scores. Its formula is notably more sophisticated and incorporates three distinct calculations across different time windows.
It considers, and this is the formula, it considers the mean ZCM scores from both five epochs. before and after the current epoch. So it takes the mean across the 11 epoch windows here. The second value is the standard deviation of the five preceding epochs, including the one that's current. And it also counts the epochs with intermediate ZCM values between 50 to 100 within an 11 epoch window value.
So, it's, it's an interesting algorithm, and, uh, it's an inference what we've seen before from the Cole-Kripke. So, the lower values indicate wake periods while positive values suggest sleep. and this combination of varied calculations across different epoch windows might contribute to a more robust sleep classification method, as we'll see later.
However, in general, the Cole-Kripke method has been validated for a broader range of populations, including those with sleep disorders such as OSA, while the Sadeh algorithm has been tested in pediatric populations, including preschool children, where it has good sensitivity but lower specificity. Now, the final step of transformation is the score refinement stage, where the calculated sleep propensity is modified further to arrive at a more physiologically plausible sleep wake pattern.
This was first proposed by Webster in 1982, and this step addresses a specific challenge in actigraphy and this is the misclassification of quiet wakefulness as sleep. And this usually typically occurs in two key scenarios. First, during the pre sleep period, where stillness before actual sleep onset can be misinterpreted as sleep.
And so there are various rules for this scenario, which are detailed here. In this example, because there's more than four minutes of wake time, this one minute of sleep is then converted into wake. The second scenario is during extended nighttime awakenings where brief periods of immobility might be incorrectly classified as short sleep bouts, within a longer wake period.
So the rules are again detailed here and in this case, because there's more than six minutes on each side, I mean, six, less than six minutes of sleep surrounded by more than 15 minutes of wake, then this is reclassified as wake. So, it's important to recognize that this final transformation stage is considered optional.
It's not always pursued, but, you know, that just wanted you to be aware of that possibility.
So at Labfront, we are working to automate these algorithms for your specific datasets and to produce actigraphy reports on a nightly basis. And this is a sample one night sleep report in CSV or Excel format, which we can provide to the researchers.
It gives you the parameter settings in the upper left. Minute to minute classification of sleep on the bottom left, which includes the Cole-Kripke and Sadeh method with the actual activity, uh, sleep propensity scores in each of, sort of minute to minute and then we have the summarization of the results on the right, in both tabular and plot form.
And you can see here the, the differences in the, in the Cole-Kripke and the Sadeh methodologies. Now Philips Actiwatch actually uses their own proprietary algorithm, which does not use ZCM but uses PIM or the energy level. We have actually reversed reverse engineered the algorithm a bit, but we will need some time to validate it further.
We, you know, can consider if there is significant demand out there, to possibly sort of recapitulate some of the results that the Phillips Actiwatch devices are giving. The other thing is also we may incorporate sample re-scoring as an option, also within this, sort of actigraphy reports.
These are settings that we could sort of preset based on your preferences, whether you want the AMI Motionlogger, uh, or the, uh, plus or minus the Philips ActiWatch. And this is something that, you know, just continue to be in touch with us to see how we can help you from that regard.

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