Skip to main content
Search form
  • 00:00


  • 00:15

    JAROSLAW HAREZLAK: My name is Jaroslaw Harezlak.[Jaroslaw Harezlak, Professor of Epidemiology Biostatistics]I'm a professor of epidemiology and Biostatisticsat the Indiana University School of Public Healthin Bloomington, Indiana, USA.My major research interests are wearable computing,brain imaging, and the regularization methodsin statistics.In this video, I'll present and discuss a study

  • 00:38

    JAROSLAW HAREZLAK [continued]: using wearable devices to classifydifferent types of walking.So the title of my talk, my presentationis "Collection and Processing of raw AccelerometryData from Wearable Devices."In the next 15 or so minutes, I willaddress the issues of measuring physical activity objectively.

  • 00:60

    JAROSLAW HAREZLAK [continued]: I will discuss available devices.I will also discuss what the data will look like.And then I will go on with the actual studythat we performed here at Indiana University.I'll describe the study design, the analysisof the data collected from the study,including both feature extraction from the data,

  • 01:21

    JAROSLAW HAREZLAK [continued]: and classification of different working activities usingextracted features.In particular, I will concentrateon differentiating between walking and stair climbing.So how do we measure physical activity?Well, there are different ways.In the old times, we would fill outsurveys to assess how active we were during a day,

  • 01:47

    JAROSLAW HAREZLAK [continued]: during a period of time.Recently, a more popular approachis to use so-called acceleration measurements.So there is a whole field devotedto that is called actigraphy.So in this area what we do, we use devices, for example,

  • 02:07

    JAROSLAW HAREZLAK [continued]: like this one Actigraph GT3X+, to measure accelerationsin three dimensions.This is the physical space where we live in.And then such data are analyzed.So what do the data look like?

  • 02:27

    JAROSLAW HAREZLAK [continued]: As you can see on this graph, we havedata collected for a 24-hour period isshown in the top figure.It's very hard to see actually, to findthe features of this data.So what we usually do, we zoom in this data.So in the middle plot, we presented the data

  • 02:51

    JAROSLAW HAREZLAK [continued]: for one hour, for 60 minutes.And in the bottom plot, we presented the same databut just--we extracted it for just one minute.So as you can see, this data have a few featuresthat we will be using for our purpose.What we did, we were interested in differentiation

  • 03:14

    JAROSLAW HAREZLAK [continued]: of different walking activities.So what we did, we designed an experiment,this was with my former PhD student.We call the study, IU walking and driving study.Now, we had walking and driving portion in the study,but we concentrated, in this presentation,in this work on differentiation between different types

  • 03:35

    JAROSLAW HAREZLAK [continued]: of walking.Why walking and driving?Well, these are two most common activitiesthat are being done in general populationin our everyday life.Another motivation for our study was

  • 03:56

    JAROSLAW HAREZLAK [continued]: that it was very hard actually to obtainsuch data to validate the statistical methodologythat we're developing.So in the general study, we're differentiating,walking, stair climbing up, stair climbing down,and driving.

  • 04:17

    JAROSLAW HAREZLAK [continued]: And as I said here, will concentrateon walking differentiation on flat surface,walking up, and walking up.Now, the data that we collected in the study we designed,we did it in a way that it simulated free livingenvironment.So we designed the area where the study participants

  • 04:38

    JAROSLAW HAREZLAK [continued]: will be walking.But this was again, resembling the natural free livingenvironment.We designed walking and driving study.We enrolled participants in the study.We enrolled 32 healthy adults, 19 females, and 13 males.And we enrolled actually over an age range

  • 05:02

    JAROSLAW HAREZLAK [continued]: that's representative of a walking population.So the age range was between 23 and 54 years old.We collected data using four actigraph devices.This are GT3X+ accelerometers that I showed earlierin the video.We placed them on left wrist, left hip, left ankle,

  • 05:25

    JAROSLAW HAREZLAK [continued]: and right ankle.We collected this data at the sampling frequencyof 100 Hertz, meaning 100 observations per second,in three orthogonal axis.Now in order to differentiate to put timestamps in the datacollection itself, we ask participants to clap

  • 05:48

    JAROSLAW HAREZLAK [continued]: between different activities.So for example, if they were walking on the flat surfaceand then they were starting to walk up,we asked them to stop and clap three times.What that enables us to do is we see big spikes in the data.So in a sense we have an insight internal timestamp

  • 06:12

    JAROSLAW HAREZLAK [continued]: where we can create gold standard for the classificationof different activities.So going back to our research goals, once we collectedsuch data, we wanted to extract features that describeimportant aspects of walking.And then we're using those extracted featuresto build an interpretable classification model.

  • 06:35

    JAROSLAW HAREZLAK [continued]: Now, as you can see on this plot, in the top plot,we see three activities, we see walking downstairs, upstairs,and downstairs again.And between those three activities,you can see in gray, there are three clapping activities

  • 06:56

    JAROSLAW HAREZLAK [continued]: from the--well, just clapping hands.So that enables us to put gold standard in the time framehere.In the bottom plot, you see the summaryof the data from the free axis, is so-called vector magnitude.And it's basically the square rootof the vectors of the coordinates in three

  • 07:19

    JAROSLAW HAREZLAK [continued]: different directions.So going with our goal 1, we extracted meaningful featuresfrom the data.So that included vector magnitude, alsothe tri-axial view of the data.And then we use different statistical and signalprocessing methods to extract informationfrom this data in short time intervals.

  • 07:42

    JAROSLAW HAREZLAK [continued]: So for example, we use fast Fourier transform, weuse discrete wavelet transform.And we use also use features of the data in the time domainby correlating the time series from the three axis.So in total, we extracted 13 features.And then our classification was differentiating

  • 08:06

    JAROSLAW HAREZLAK [continued]: between walking, descending stairs, and ascending stairs.We did the classic kind of statistical methodswhere we did cross-validation to obtain actuallythe testing rates, the misclassificationrates from the data.And as you can see in the figure for the sample results,

  • 08:29

    JAROSLAW HAREZLAK [continued]: our classification was very good wherewe plot here in red the actual activity that was performed.And in blue, we plot our results from the algorithmthat we built.What have we done in our research?Well, our goal was to differentiate

  • 08:50

    JAROSLAW HAREZLAK [continued]: between walking on the level ground,and stair climbing both up and down.And we did it.We performed that activity with good accuracy.Now, we also compared the data collectedfrom different devices and of course our algorithm.And especially for the walking activitythat we're interested in, the data collected from the ankles

  • 09:15

    JAROSLAW HAREZLAK [continued]: provided the best classification accuracy.Now, what were the challenges that we faced?Well, from the experimental point of view,one of the challenges was that even though the participantsknew they will be participating in a walking activity, some

  • 09:39

    JAROSLAW HAREZLAK [continued]: of them--some of the participants didn't wear proper walking shoes.It's kind of very hard to walk in, let's say, high heels.Again from the experimental pointof view, the time synchronizationbetween different devices or among all four devices,it was challenging.We had to be very careful with both the initialization

  • 10:01

    JAROSLAW HAREZLAK [continued]: of the devices and then synchronizing the timeamong all of the devices.And the last challenge, the last important challengeis that segmentation of the signal,especially for the short bouts of activity, stillit's a challenging event which we are hoping

  • 10:24

    JAROSLAW HAREZLAK [continued]: to address in the future.So in terms of recommendations for future research,we have a few.For example, the most important recommendationis that we have to match the design of the studyto the achievable goals.For example, in our experiments, we

  • 10:45

    JAROSLAW HAREZLAK [continued]: were trying to-- our conditions were resembling the free livingactivity.However, it was still a designed experiment.If we give the devices to the participantsto collect data over, let's say, one week, wehave to pay special attention to the initializationof the devices and then syncing of the data from the devices

  • 11:07

    JAROSLAW HAREZLAK [continued]: to the computer.Another important thing here is usuallyin free living activity will not have gold standard labels.So for example, we will not be able to say with certaintywhat activity was actually performed.So the algorithms that we'll need to develop,will have to deal with so-called silver

  • 11:30

    JAROSLAW HAREZLAK [continued]: standard or even worse standards when we put labels on the data.And the last thing here is differentiationbetween so-called micro and macrolevels of physical activity.So for example, if we want to only assesshow much time we are spending sitting versus standing,

  • 11:54

    JAROSLAW HAREZLAK [continued]: the data collected and the processing of the datadoesn't have to be as carefully designed and executedand as we did here in differentiationof walking between kind of walking on a flat surface,and walking upstairs and downstairs.The last one, this are so-called micro-level.

  • 12:18

    JAROSLAW HAREZLAK [continued]: This is micro-level of physical activity, where we trulywant to get to, for example, cadence of walking,or the speed of walking.Now of course, all of this work wouldn'tbe possible without participationof the whole research team.So I would like to give special thanks to my former student,William Fadel, who is currently an assistant professor.

  • 12:40

    JAROSLAW HAREZLAK [continued]: To another faculty from my department XiaochunLi and my colleague Andrea Chomistek.Also a lot of this work was done with the helpof my undergraduate interns Stephen Albertson.And my collaborator from Johns Hopkins University,Jacek UrbanekThank you for watching, and I hope

  • 13:02

    JAROSLAW HAREZLAK [continued]: you'll have fun both collecting and processing datafrom devices like this, the actigraph.And this is a very important area these dayswhere we are truly trying to assess in an objective wayhow much physical activity actually we do perform.

  • 13:22

    JAROSLAW HAREZLAK [continued]: And with the devices currently available on the market,even the devices that we don't think aboutas measuring physical activity, let'ssay Apple Watch or our iPhones, or other smartphones,this data is collected on us.What we do with them it's up to us.

  • 13:43

    JAROSLAW HAREZLAK [continued]: Hope you'll have fun doing research using such data.Thank you.

  • 14:07

    JAROSLAW HAREZLAK [continued]: nbsp;


Jaroslaw Harezlak, Professor of Epidemiology and Biostatistics at the Indiana University School of Public Health, discusses a study using wearable devices to classify different types of walking, including measuring physical activity, available devices, a sample study design, and challenges faced.

Looks like you do not have access to this content.

Collection and Processing of the Raw Accelerometry Data From Wearable Devices

Jaroslaw Harezlak, Professor of Epidemiology and Biostatistics at the Indiana University School of Public Health, discusses a study using wearable devices to classify different types of walking, including measuring physical activity, available devices, a sample study design, and challenges faced.

Copy and paste the following HTML into your website