Professional Documents
Culture Documents
learning to decide which algorithm should be applied to each situa- Table 1: Overview of all data recording setups for each partic-
tion. For example, Bagui et al. report an accuracy of 93 % with this ipant as a combination of MM, sensor position, and location.
approach [2]. For each setting, we made two recordings
With a similar strategy, we aspire to build a fexible activity
tracking system for older adults to (1) recognise their MM, and (2) Movement modality Sensor position Location
apply an algorithm for that specifc modality to accurately com- Inside
Mobility aid
pute the achieved levels of PA. In this study, we focus on the frst Outside
Cane
objective, and consider six MMs: walking with a cane, crutches, Inside
Wrist
walker, rollator, using a wheelchair, or walking without a mobility Outside
aid (Figure 1). Our goal is to investigate the possibility of identi- Inside
Mobility aid
fying the user’s MM based on sensor data from a single inertial Outside
Crutches
measurement unit (IMU). Since this is part of an exploratory study, Inside
Wrist
we recorded a dataset with only three younger participants (two Outside
persons in the age range of 25-35, and one in 60-65). This allowed Inside
Mobility aid
us to avoid exposing older adults to health risks in light of the Outside
Walker
ongoing COVID-19 pandemic. Next, we implemented a minimum Inside
Wrist
distance classifer, and used the dataset to verify the accuracy of Outside
our classifer’s prediction. Inside
Mobility aid
Outside
Rollator
2 DATASET Inside
Wrist
Outside
We gathered 3-axes accelerometer and gyroscope data from three
participants engaging in the six MMs we considered in this study. In Inside
Mobility aid
addition to these MMs, we wanted to distinguish between a person Outside
Wheelchair
being in motion or not. Therefore, we also collected data while Inside
Wrist
the participants were sitting down calmly. For each modality, we Outside
recorded sensor data in a combination of diferent setups: having Inside
No mobility aid Wrist
the IMU attached to either the participant’s wrist or to the mobility Outside
aid being used, and recording the data either inside (smooth foor) Sitting Wrist Inside
or outside (uneven pavement). For each specifc setup, we made
Table 2: Three derived types of IMU data
two recordings (Table 1): the frst one to train our classifcation
model, the second one to verify the accuracy of its predictions.
Type of data Symbol
Each recording is about 60 seconds long, and data is sampled √
at a rate of approximately 20Hz. We used an Adafruit Flora micro- ®
Magnitude of angular velocity (�) � �2 + � �2 + ��2 = �
controller and a BNO055 IMU to register all activity and send the
√
Magnitude of acceleration (®
�) ��2 + � 2� + ��2 = �
sensor data in real time to a laptop, where everything was stored
in a CSV fle. We recorded movement data for each setup of Table Magnitude of acceleration (® �) along �® · �® = ��
1. The entire data collection process took about three to four hours ®
the gravitational force’s direction (�)
per participant. None of the participants used mobility aids in their
Table 3: Overview of all considered, mostly statistical, base
daily lives. However, for this exploratory study, having experience
features
with all aids is less important, as we investigate whether each aid
afords a sufciently distinct pattern of movement to automatically
recognise it using a machine learning algorithm. Base feature Abbreviation
Frequency of the signal f
Minimum value of the signal min
3 IMPLEMENTATION OF THE CLASSIFIER Maximum value of the signal max
We opted to use a minimum distance classifer because of its simplic- The diference between max and min range
ity and low use of resources, e.g., memory and computing power, Average value of the signal mean
and trained a model based on a feature set that highlights the most Standard deviation of all values of the signal sd
pronounced diferences between parameters of MMs. We derived Median value of the signal median
three types of data samples from the IMU, i.e., angular velocity mag- First quartile, 25th percentile of all values Q1
nitude, acceleration magnitude, and magnitude of the acceleration
Third quartile, 75th percentile of all values Q3
component along the gravitational force’s direction (Table 2). For
The diference between Q3 and Q1 IQR
each of these data types, we considered 11 possible base features
Total energy of the data signal e
(Table 3). Thus, to train our classifer’s model, we can compose a
feature set from a selection of 33 diferent features in total.
To train our minimum distance classifer, we select a feature
set, group all participants’ data per MM, and compute the mean
Flexible Activity Tracking for Older Adults Using Mobility Aids ASSETS ’22, October 23–26, 2022, Athens, Greece
Actual
Wheelchair
Crutches
Rollator
Walker
Sitting
No aid
Cane
(c) (d)
Cane 67 0 8 8 0 0 0
Crutches 25 92 0 8 0 0 0
Predicted
Rollator 0 0 92 0 0 8 0
Walker 0 0 0 84 8 0 0
Wheelchair 0 8 0 0 92 0 0
No aid 8 0 0 0 0 92 0
Sitting 0 0 0 0 0 0 100
4 RESULTS
(e) Here, we consider four diferent subsets of our data to verify our
classifer. For each of these cases, we obtained a feature set through
Figure 1: Mobility aids supported by the fexible tracking forward selection, i.e., iteratively adding features that lead to the
system: (a) cane, (b) crutches, (c) walker, (d) rollator, (e) biggest increase in prediction accuracy until it no longer increases.
wheelchair. The sixth movement modality, i.e., walking with- Table 4 shows the resulting confusion matrix when we consider all
out a mobility aid, is not depicted here data recordings where participants wore the IMU on their wrist,
and train our classifer model with 16 selected features. The mean
prediction accuracy for this setup is 88 %. When the IMU is attached
to the mobility aid instead of the participant’s wrist, the prediction
accuracy is 82 % with four features selected to train the model
(Table 5). Note that no mobility aid is used for sitting and walking
value for each selected feature of all MMs (see Algorithms 1 and 2). without an aid. Therefore, only for those MMs, we included the data
Once a model is trained, the algorithm predicts the user’s MM by recordings during which the IMU was attached to the participants’
comparing the feature values of a given data sample to the mean wrist (see Table 1).
feature values of all MMs stored in the model. The smallest distance In a setting where an older adult would use no mobility aid to
between these values results in the classifer’s MM prediction. To walk short distances inside, and a rollator or a walker for longer dis-
compensate for potential diferences in range and variance among tances outside, the classifer achieves a prediction accuracy of 96 %
the selected features, we used the Mahalanobis distance instead of (Table 6). Similarly, if a person only walks either with or without a
the Euclidean distance [9]. cane, the classifer always predicts the correct MM (Table 7).
ASSETS ’22, October 23–26, 2022, Athens, Greece Vargemidis, et al.
Table 5: Confusion matrix showing the classifer’s predic- distinguishing two diferent activities [2], we found that predictions
tions as a relative ratio (in %) (IMU attached to mobility aid, are still accurate when we include up to seven MMs, including MMs
except for No aid and Siting). The mean accuracy is 82 %, using fve mobility aids. Note that in a more realistic setting, one
and all results in this table are obtained using the following older adult is unlikely to use all considered types of mobility aids
features: ���� , ���� , ��,�3 , ��,����� on a regular basis. We can leverage this knowledge, e.g., by let-
ting users select which aids they regularly use through a one-time
Actual setup. This will further increase the classifer’s prediction accuracy,
Wheelchair
mounting to 96 % or higher.
Crutches
Rollator Additionally, the results suggest that the IMU’s location also
Walker
Sitting
No aid
impacts the prediction accuracy. Here, we found more favourable
Cane
Rollator 0 0 50 0 0 0 0
Walker 8 8 0 92 0 0 0 This initial study allowed us to investigate the feasibility and poten-
Wheelchair 0 8 33 8 33 0 0 tial of a fexible tracking system. An important limitation is that we
No aid 0 8 0 0 0 100 0 recruited three participants only, and no older adults were involved
Sitting 0 0 17 0 17 0 100 for recording our dataset. We took this decision because the work
was carried out during the COVID-19 pandemic, and we did not
Table 6: Confusion matrix showing the classifer’s predic- want to put older adults at risk of exposure to collect preliminary
tions as a relative ratio (in %) for a subset of MMs (only wrist- data. Future studies need to verify whether the results found in
worn IMU). The mean accuracy is 96 %, and all results in this this study still apply to similar setups with older adults, specifcally
table are obtained using the following features: ����� , � � , those who are familiar with using mobility aids in their daily lives,
��,�3 , ��,�� and what adjustments would need to be made. Moreover, these re-
sults allow for continuing the fexible tracking approach, i.e., linking
Actual a MM prediction to a tailored algorithm for accurately computing
the user’s level of PA, e.g., step count or distance travelled.
Rollator
Walker
Sitting
No aid
7 CONCLUSION
Rollator 83 0 0 0 In this exploratory study, we examined fexible tracking as an ap-
proach to make activity tracking more accessible to older adults
Predicted
Walker 0 100 0 0 using mobility aids. As an initial step, we implemented and evalu-
No aid 17 0 100 0 ated a minimum distance classifer, achieving prediction accuracies
of more than 82 %. Future work should focus on evaluating this ap-
Sitting 0 0 0 100
proach with movement data of older adults, and linking predictions
to algorithms that can accurately track activity.
Table 7: Confusion matrix showing the classifer’s predic-
tions as a relative ratio (in %) for a subset of MMs where we REFERENCES
consider only sitting, and walking with or without a cane [1] Hyun-Sung An, Gregory C. Jones, Seoung-Ki Kang, Gregory J. Welk, and Jung-
(only wrist-worn IMU). The mean accuracy is 100 %, and all Min Lee. 2017. How valid are wearable physical activity trackers for mea-
results in this table are obtained using the following features: suring steps? European Journal of Sport Science 17, 3 (March 2017), 360–368.
https://doi.org/10.1080/17461391.2016.1255261 Publisher: Routledge _eprint:
��,�3 , ��,�� https://doi.org/10.1080/17461391.2016.1255261.
[2] Sikha Bagui, Xingang Fang, Subhash Bagui, Jeremy Wyatt, Patrick Houghton,
Joe Nguyen, John Schneider, and Tyler Guthrie. 2022. An improved step count-
Actual ing algorithm using classifcation and double autocorrelation. International
Sitting
No aid
gait analysis for early diagnosis and therapy monitoring in Parkinson’s disease.
No aid 0 100 0 In 2011 Annual International Conference of the IEEE Engineering in Medicine and
Biology Society. 868–871. https://doi.org/10.1109/IEMBS.2011.6090226
Sitting 0 0 100 [4] Lynne Clay, Megan Webb, Claire Hargest, and Divya Bharatkumar Adhia. 2019.
Gait quality and velocity infuences activity tracker accuracy in individuals
post-stroke. Topics in Stroke Rehabilitation 26, 6 (Aug. 2019), 412–417. https:
//doi.org/10.1080/10749357.2019.1623474 Publisher: Taylor & Francis _eprint:
5 DISCUSSION https://doi.org/10.1080/10749357.2019.1623474.
The results suggest that predicting diferent MM is possible, and [5] Conor Cunningham, Roger O’ Sullivan, Paolo Caserotti, and Mark A. Tully.
2020. Consequences of physical inactivity in older adults: A systematic re-
accuracies above 82 % can be achieved with straightforward ma- view of reviews and meta-analyses. Scandinavian Journal of Medicine & Sci-
chine learning algorithms. While Bagui and colleagues report on ence in Sports 30, 5 (2020), 816–827. https://doi.org/10.1111/sms.13616 _eprint:
Flexible Activity Tracking for Older Adults Using Mobility Aids ASSETS ’22, October 23–26, 2022, Athens, Greece
https://onlinelibrary.wiley.com/doi/pdf/10.1111/sms.13616. community dwelling older adults. Geriatric Nursing 36, 2 (March 2015), S21–S25.
[6] Andreas Ejupi, Matthew Brodie, Yves J. Gschwind, Daniel Schoene, Stephen Lord, https://doi.org/10.1016/j.gerinurse.2015.02.019
and Kim Delbaere. 2014. Choice stepping reaction time test using exergame tech- [13] Tifany Sears, Elmer Alvalos, Samantha Lawson, Ian McAlister, L. Eschbach, and
nology for fall risk assessment in older people. In 2014 36th Annual International Jennifer Bunn. 2017. Wrist-Worn Physical Activity Trackers Tend To Underesti-
Conference of the IEEE Engineering in Medicine and Biology Society. 6957–6960. mate Steps During Walking. International Journal of Exercise Science 10, 5 (Aug.
https://doi.org/10.1109/EMBC.2014.6945228 2017), 764–773. https://digitalcommons.wku.edu/ijes/vol10/iss5/12
[7] Orestis Giotakos, Katerina Tsirgogianni, and Ioannis Tarnanas. 2007. A virtual [14] Salvatore Tedesco, Marco Sica, Andrea Ancillao, Suzanne Timmons, John Barton,
reality exposure therapy (VRET) scenario for the reduction of fear of falling and and Brendan O’Flynn. 2019. Accuracy of consumer-level and research-grade
balance rehabilitation training of elder adults with hip fracture history. In 2007 activity trackers in ambulatory settings in older adults. PLOS ONE 14, 5 (May
Virtual Rehabilitation. 155–158. https://doi.org/10.1109/ICVR.2007.4362157 2019), e0216891. https://doi.org/10.1371/journal.pone.0216891 Publisher: Public
[8] Paul Kooner, Taran Schubert, James L Howard, Brent A Lanting, Matthew G Library of Science.
Teeter, and Edward M Vasarhelyi. 2021. Evaluation of the Efect of Gait Aids, [15] Dimitri Vargemidis, Kathrin Gerling, Vero Vanden Abeele, Luc Geurts, and Katta
Such as Canes, Crutches, and Walkers, on the Accuracy of Step Counters in Spiel. 2021. Irrelevant Gadgets or a Source of Worry: Exploring Wearable Ac-
Healthy Individuals. Orthopedic Research and Reviews 13 (Jan. 2021), 1–8. https: tivity Trackers with Older Adults. ACM Transactions on Accessible Computing
//doi.org/10.2147/ORR.S292255 (TACCESS) 14, 3 (Aug. 2021), 28. https://doi.org/10.1145/3473463 Publisher: ACM
[9] Konstantinos Koutroumbas and Sergios Theodoridis. 2008. Pattern Recognition. PUB27 New York, NY, USA.
Academic Press. [16] Rachel K. Walker, Amanda M. Hickey, and Patty S. Freedson. 2016. Advantages
[10] Jonas Lauritzen, Adolfo Muñoz, Jose Luis Sevillano, and Anton Civit. 2013. The and Limitations of Wearable Activity Trackers: Considerations for Patients and
usefulness of activity trackers in elderly with reduced mobility: a case study. Clinicians. Clinical Journal of Oncology Nursing 20, 6 (Dec. 2016), 606–610.
Studies in Health Technology and Informatics 192 (2013), 759–762. https://doi.org/10.1188/16.CJON.606-610
[11] Denys J.C. Matthies, Marian Haescher, Suranga Nanayakkara, and Gerald Bieber. [17] Clint Zeagler. 2017. Where to wear it: functional, technical, and social considera-
2018. Step Detection for Rollator Users with Smartwatches. In Proceedings of the tions in on-body location for wearable technology 20 years of designing for wear-
Symposium on Spatial User Interaction (SUI ’18). Association for Computing Ma- ability. In Proceedings of the 2017 ACM International Symposium on Wearable Com-
chinery, New York, NY, USA, 163–167. https://doi.org/10.1145/3267782.3267784 puters. ACM, Maui Hawaii, 150–157. https://doi.org/10.1145/3123021.3123042
[12] Tara O’Brien, Meredith Troutman-Jordan, Donna Hathaway, Shannon Armstrong,
and Michael Moore. 2015. Acceptability of wristband activity trackers among