An Experimental Investigation Comparing Age-Speciﬁc and Mixed-Age

Models for Wearable Assisted Activity Recognition in Women

Pratool Bharti

, Arup Kanti Dey

, Sriram Chellappan

and Theresa Beckie

Dept. of Computer Science and Engineering, University of South Florida, Tampa, FL, U.S.A.

College of Nursing, University of South Florida, Tampa, FL, U.S.A.

Keywords:

Wearable Computing, Activity Recognition, Health Informatics, Machine Learning, Algorithms, Aging.

Abstract:

In this paper, we investigate the impact of age diversity on accuracy for activity recognition among women

with wrist-worn wearables. Using a sample of 10 elder women and 10 younger women, and by monitoring

ﬁve activities related to cardiac care (Running, Brisk Walking, Walking, Standing and Sitting), we show

that while personalized models are best, activities classiﬁcation based on age speciﬁc models are deﬁnitely

superior in terms of accuracy compared to classiﬁcation using mixed age models. We do so by a) extracting 11

features from inertial sensing data; b) reducing dimensionality using Linear Discriminant Analysis methods;

c) quantifying variance among features using Principal Component Analysis; d) clustering activities; and

ﬁnally e) comparing classiﬁcation accuracies of all activities for personalized, age-speciﬁc and mixed-age

models. We believe that our study is unique, and potentially important for superior healthcare for women, a

demographic that is largely underserved today across the world.

1 INTRODUCTION

Consistent physical activity is important for human

health across all ages. To cater to this need, billions

of dollars and signiﬁcant human resources have been

invested by industry and academia to advance the

ﬁeld of wearable assisted activity recognition. Chief

among these are wrist-worn wearables like the FitBit

band, Apple Watch, Samsung Gear etc., that are very

popular today. The global demand of wearable tech-

nologies is estimated to be around 200 million devices

in 2021 (Beaver, 2016).

1.1 State of the Art in Wearable Tech

w.r.t. Elders and Identiﬁed Gap

As of today, the wearable tech market is dominated

by sensory devices worn on the wrist for recogniz-

ing basic physical activities like walking, running, sit-

ting, standing and sleeping. These wearables typ-

ically come with pre-trained models, but do suffer

from accuracy problems to a certain degree. This

lack of accuracy is unavoidable, since each person

performs the same activity a little differently, and it

is virtually impossible for an algorithm in a wearable

to correctly identify all possible modes of diversity

across all humans. However, when it comes to el-

ders, both patients and physicians have very high ac-

curacy expectations, since physical activities are ex-

tremely important, but can simultaneously be strenu-

ous for elders, and so there is an expectation that every

“step” be counted, and counted correctly. However,

despite studies showing that aging causes diversities

in the way humans perform and perceive physical ac-

tivities (Borg, 1998), (Bar-Or, 1977), (Levy and My-

ers, 2004) there is no careful study yet on impact of

age diversities on accuracy in wearable assisted activ-

ity recognition.

1.2 Our Contributions

In this paper, we investigate the importance of con-

sidering age diversities in wearable assisted activity

classiﬁcation for women, and identify critical obser-

vations. Each physical activity we investigate in this

paper is vital for health, and speciﬁcally cardiac care.

We speciﬁcally focus on women subjects in this paper

since they are an underserved population in cardiac

care (Valencia et al., 2011), (Benz Scott et al., 2002),

and we can retain problem scope.

Our speciﬁc contributions are as follows. We con-

duct an experiment with 20 women subjects, where

10 subjects were younger in the age group of 21−26,

Bharti, P., Dey, A., Chellappan, S. and Beckie, T.

An Experimental Investigation Comparing Age-Speciﬁc and Mixed-Age Models for Wearable Assisted Activity Recognition in Women.

DOI: 10.5220/0007398003670374

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 367-374

ISBN: 978-989-758-353-7

367

and the other 10 subjects were older in the age group

of 65 − 75. Each subject was asked to wear a wrist

wearable, and they performed a series of ﬁve activi-

ties: Brisk Walking, Running, Sitting, Standing and

Walking. During this time, the accelerometer and gy-

roscope readings were collected from the wearable

device and later exported to high end server for fur-

ther processing.

Subsequently, a) we extract 11 contextually rele-

vant features each from the three axes of accelerom-

eter and gyroscope sensory data, and analyze them

extensively; b) perform dimensionality reduction us-

ing Principal Component Analysis; c) perform clus-

ter analysis via Linear Discriminant Analysis; and d)

implement several machine learning algorithms for

classiﬁcation. We make several interesting ﬁndings.

We ﬁnd that among principal components, only very

few components contain large portions of variance in

datasets for the younger cohort, compared to the elder

cohort. The conclusion here is higher similarities in

activities when performed by younger women, com-

pared to elder women. The distribution of variances

in the mixed-age cohort was the worst, hence making

a case for age-speciﬁc models. After clustering, we

ﬁnd that the activity clusters on the younger cohort

are highly separable with minimal overlap, while the

elder only cohort had more than reasonable overlaps

among activities classiﬁed, and the mixed-age model

again performed the worst with signiﬁcant confusion.

Finally, we ﬁnd that classiﬁcation accuracies for the

age-speciﬁc models outperform mixed-age models by

an average of 20%. Models personalized for each in-

dividual are much more accurate.

To the best of our knowledge, our study is the ﬁrst

to comprehensively investigate impact of age dispar-

ities on accuracies in terms of wearable assisted ac-

tivities classiﬁcation, and present a formal need for

age-speciﬁc, or even better, personalized models. The

exclusive focus on women is a further salience of our

contributions.

2 RELATED WORK

In this section, we elaborate on important related

work in two broad topics: technology assisted activ-

ity recognition in general, and works that speciﬁcally

consider impact of diversities on wearable assisted

physical activities recognition.

2.1 Physical Activity Recognition using

Computing Technologies

Human Activity Recognition is a very well studied

topic, with some good surveys in (Avci et al., 2010),

anchez et al., 2008). There are broadly three classes

of work in this realm. The ﬁrst one is to use sen-

sors emplaced in the ambient infrastructure for ac-

tivity recognition. Typically, the sensors are video

cameras, WiFi receivers, PIR sensors, etc. Such sys-

tems have been used to detect activities like walking,

sitting, standing, running in (Bao and Intille, 2004),

gait study in (Lee and Grimson, 2002), activities per-

formed by healthcare professionals in clinical settings

in (S

anchez et al., 2008) and more. The second class

of work detects activities using only wearables em-

placed in different body positions like wrists, ﬁngers,

neck, feet and more. Among these, there are works

like (Bharti et al., 2018a) that recognizes basic ac-

tivities like walking, sitting, standing, running etc.,

and complex ones like cooking, cleaning etc. There

are also works that detect more ﬁne grained activi-

ties like self-harming activities (Bharti et al., 2018b);

walking upstairs, walking downstairs, taking elevator

up/ down, lying down (Jiang and Yin, 2015); lie-to-

sit, stand-to-lie, lie-to-stand, cycling, ironing clothes

etc. (Reyes-Ortiz et al., 2016). Finally, there are

works that use a combination of wearable sensors

and infrastructure sensors for recognizing activities

like cooking, cleaning utensils and many more (De

et al., 2015). From reading extensive related work in

this space, we observed that the elder demographic is

largely under-represented in existing studies today in

wearable assisted activity recognition, as summarized

in Table 1.

2.2 Impact of Diversities in Wearable

Assisted Activity Recognition

We ﬁnd that most works that look at wearable assisted

activity recognition for elder subjects are speciﬁc for

Fall Detection only (de la Concepci

on et al., 2017),

(Kaur and Kaur, 2017), (Wang et al., 2017). From the

perspective of detecting basic physical activities using

wearables, there are very limited works that consider

elder subjects. For instance, one very recent work in

(Alinia et al., 2017), evaluate three different types of

Fitbit activity trackers and concluded that these de-

vices are accurate when subject walks/ runs on tread-

mill, but the accuracy goes down when the subject is

walking with an assisted device, or walks very slowly,

which are representative with elders.

In fact, we are aware of only one work

(Del Rosario et al., 2014), where there was an at-

HEALTHINF 2019 - 12th International Conference on Health Informatics

368

Table 1: Related work in space of human activity recognition using wearables.

Publications Sensors Age group (in years.)

(de la Concepci

on et al., 2017) 3-axis Accelerometer 19-48

(Jiang and Yin, 2015) 3-axis Accelerometer and Gyroscope 19-49

(Bharti et al., 2018a)

3-axis Accelerometer, 3-axis Gyroscope,

Barometer pressure sensor, Temperature,

Humidity, iBeacon, GPS 20-25

(Bharti et al., 2018b) 3-axis Accelerometer, 3-axis Gyroscope 25-30

(Reyes-Ortiz et al., 2016) 3-axis Accelerometer, 3-axis Gyroscope 19-48

(Alinia et al., 2017) FitBit wearables 21-31

(Wang et al., 2017) Wi-Fi signal NA

tempt to study the efﬁcacy of an age-speciﬁc model

and a mixed-age model for activity recognition using

smartphone sensors. The authors showed that accu-

racy is signiﬁcantly higher when a model is trained

and tested on age-speciﬁc datasets, compared to a

mixed-age datasets. Our work in this paper is related

to (Del Rosario et al., 2014), but there are compelling

differences. In our paper here, we employ a wrist-

worn device, which is more realistic and practical,

when compared to smart-phones for activity recogni-

tion. Secondly, the work in (Del Rosario et al., 2014)

does not look at statistical properties of features ex-

tracted (which will give more context to the results),

and does not not consider activities like brisk walking

and running, which we do in our paper, hence making

our contributions signiﬁcant and relevant.

3 EXPERIMENTAL SETTING

We now present our experimental set-up. We col-

lected a dataset for ﬁve ADL (Activities of Daily Liv-

ing) activities, namely Brisk Walking, Walking, Run-

ning, Standing and Sitting. Each one of these activ-

ities is vital for health across ages, and more so for

cardiac care. Ten younger (age range: 21-26) and ten

elder female participants (age range: 65-75) took part

in our experiments. All the younger subjects did all

of the activities, as did four elder subjects. Six el-

der subjects (in the higher age group) did not run (for

obvious reasons), but did all the four other activities.

Each activity was performed for 4 minutes by every

individual. Thus, around 20 minutes of data were

collected from each participant. A wearable device

called Shimmer (Burns et al., 2010) equipped with

tri-axial accelerometer and tri-axial gyroscope sensor

was worn by each subject on their right wrist. Data

was sampled from the accelerometer and gyroscope

sensor at a frequency of 50Hz (samples per second).

Data from both sensors were directly streamed and

stored to server for processing and analysis.

Note that the Shimmer device is widely used

in research today for its miniature size and power-

ful sensing/ computing/ wireless transmission abili-

ties. The central element of the platform is the low-

power MSP430F5437A microprocessor with 24MHz

clock rate which controls the operation of the de-

vice. The CPU has an integrated 16-channel 12-bit

analog-to-digital converter (ADC) which is used to

constantly sample and capture triaxial acceleration

signals from an in-built accelerometer in the unit.

These accelerometers have a range of ±16g (where

g is gravitational acceleration) and were sampled at

50Hz. Note that the frequency of most human ac-

tivities lie within range of 15Hz (San-Segundo et al.,

2016). As such, this sensor sampling rate is ideal for

our problem, since according to the Nyquist rule for

lossless reconstruction of a signal, it needs to be sam-

pled at a rate that is at-least twice its highest frequency

(Landau, 1967). Figure 1 shows one subject in our ex-

periments actually wearing the Shimmer wearable.

4 DATA PRE-PROCESSING,

FEATURE EXTRACTION AND

DIMENSIONALITY

REDUCTION

We now present details on how we pre-processed the

sensory data, followed by feature extraction, and re-

duction of data dimensionality for tractability.

4.1 Data Pre-processing

The ﬁrst step after data collection is the pre-

processing of raw accelerometer and gyroscope data

collected from the Shimmer wearable. Depending on

the orientation of device, gravity can inﬂuence the

readings on one or more of the components of ac-

celerometer data. To avoid this issue, Shimmer API

provides methods to sample linear acceleration di-

rectly and hence eliminating the inﬂuence of grav-

ity. Once the linear acceleration data is extracted, we

An Experimental Investigation Comparing Age-Speciﬁc and Mixed-Age Models for Wearable Assisted Activity Recognition in Women

369

further pre-process both the accelerometer and gyro-

scope data using an adaptive median ﬁlter (Hwang

and Haddad, 1995) to remove noise, and we further

feed the signal to a low pass ﬁlter using a 15Hz cut-off

order Butterworth ﬁlter to limit the bandwidth of

the signal to the frequencies common in human mo-

tion, hence removing high frequency noise.

After noise removal, the next step is to determine

a sliding window size for the signals to attempt clas-

siﬁcation. A window size, W = 150 samples (ap-

proximately 3 seconds) with 50% overlap was used

to create a new database that was used as the train-

ing/testing data for activity classiﬁcation, also sug-

gested by prior related work on human activity recog-

nition (Banos et al., 2014) it is found that 2-5 seconds

window works best for human activity recognition us-

ing accelerometer data. Hence, we conducted our ex-

periment with window length from 2 to 5 seconds

having 0.5 second intervals and found that window

length of 3 seconds is working best for our problem.

Subsequently, the segmented window is forwarded to

next steps for feature extraction and selection.

4.2 Feature Extraction

We extracted 11 features (Table 2) for each window

along three axes of tri-axial accelerometer and tri-

axial gyroscope sensor. Each of the features we ex-

tract is contextually relevant, and extensively used in

the literature for classifying physical activities (Gupta

and Dallas, 2014), (De et al., 2015), (Sousa et al.,

2017). For example, variance, entropy and mean

crossing rate features have very good discriminating

power to classify activities like running, walking and

standing. The energy feature is very useful for dis-

criminating more intense activities like brisk walk-

ing and running. Features like variance, entropy and

mean crossing rate also can give vital information

about activities of interest to this paper. Maximum

frequency, skewness, percentile, min and max fea-

ture values captures information from signals that dif-

ferentiate the activities of interest to this paper from

other physical activities that humans perform. As

such, a total 66 features were computed for our prob-

lem, which comes from 11 features in Table 2 for ac-

celerometer and gyroscope across all the axes.

4.3 Dimensionality Reduction with PCA

and LDA

Processing high dimensional data sets (66 features

for our problem) can be noisy, harder to visualize

and computationally demanding. To ensure accuracy

and tractability, dimensionality reduction is a popu-

Table 2: Feature set.

Norm =

∑

i=1

)

+ (a

)

+ (a

)

Variance =

∑

i=1

− µ)

Max = argmax

i∈{1,2..N}

)

Min = argmin

i∈{1,2..N}

)

Entropy = −

∑

i=1

(log p

)

Max reduced Mean = ( argmax

i∈{1,2..N}

) − ¯a

Mean crossing rate = Count o f signal crossing mean

in each window

Spectral energy =

f s/2

∑

f =0

|a[ f ]|

Maximum Frequency = argmax

i∈{1,2..N}

FFT (a

)

Mean absolute Deviation =

∑

i=1

− µ|

IQR = 3rd Quartile

median

− 1st Quartile

median

Figure 1: Shimmer wireless sensor device worn on a partic-

ipant’s wrist.

lar approach, for which we use two techniques in this

paper

. The ﬁrst is Principal Component Analysis

(PCA), where the overall goal is to identify ‘prin-

cipal components’, each of which quantify a notion

variance among features, while also being orthogo-

nal to other components (Mika et al., 1999) The sec-

ond technique is Linear discriminant analysis (LDA),

where the overall goal is to ﬁnd a linear combina-

tion of features that characterizes or separates two or

more classes of objects or events. The resulting com-

Both techniques we use in this paper are well estab-

lished, and so we don’t present details.

HEALTHINF 2019 - 12th International Conference on Health Informatics

370

bination may then be used as a linear classiﬁer (Wold

et al., 1987).

5 DATA ANALYSIS OF ELDER VS

YOUNGER ACTIVITY

RECOGNITION

In this section, we show the impact of age-speciﬁc,

mixed-age and personalized model development for

classifying physical activities. We do so by a) demon-

strating that variance of features are concentrated

among fewer PCA components only for younger spe-

ciﬁc model compared to elder speciﬁc model, both of

which are better than mixed-age models; b) results

from clustering activities that demonstrate superior

performance for younger group, followed by the el-

der group and the mixed age group; c) classiﬁcation

results from many machine learning algorithms that

validate the above results; and d) ﬁnally demonstrat-

ing that models personalized for each user have much

higher accuracies compared to any model that mixes

data from subjects.

5.1 Difference in Variance in Each

Category

For all the 66 (2 sensors×3 axes×11 features) fea-

tures listed in Table 2 for our problem, we applied

PCA to transform the feature set into orthogonal PCA

components. Recall that PCA components have vari-

ances of features in decreasing order. If the data is too

noisy or have many orthogonal features, then PCA

components have variances distributed across more

components. When data is more consistent, then most

of the variances are present in very few PCA compo-

nents only. In Figure 2, we immediately see that a sig-

niﬁcant portion (i.e., 90%) of the variances are present

in only one principal component in the left most ﬁg-

ure (i.e., the younger cohort), compared to the other

two cohorts, which are 76.2% and 72.4% respectively.

To quantify further, for the case of the younger cohort,

only 7% of variance are distributed among other com-

ponents other than the principal component. But for

the elder and mixed-age cohorts, we see 20% and 22%

of variances are distributed among components other

than the principal component. This validates our in-

tuition that features of the younger only cohort have

the least amount of noise. The features among the el-

der only cohort is more noisy, but it is worse than the

mixed-age cohort.

5.2 Clustering Activities

Now that we understand the impact of age-speciﬁc

models on noise, we leverage Linear Discriminant

Analysis (LDA) to create compact clusters for each

activity, while keeping center of each cluster as sepa-

rated as possible. The LDA approach outputs a list of

transformed components for our problem scope. Out

of these, the ﬁrst two components, namely LDA1 and

LDA2 are used to draw the clusters for each activity

for younger, elder and mixed-age dataset. Using just

two components helps in better visualization, with-

out losing the generality of our results. Clusters are

plotted in Figure 3. As we can see in the ﬁgure, for

younger dataset, each activity cluster is well separated

except brisk-walking and walking. Even these two

activities have less than 10% of samples overlapping.

The reason we believe is very minimal difference in

these two activities which are hard to distinguish by

using just 2 sensors. On the other hand, the activities

in the elder dataset can be seen as more spread apart

from the center of each cluster, indicating more noise.

Furthermore, there are more overlapping points. Es-

pecially brisk-walking with walking, and sitting with

standing have more than 70% samples overlapped.

Running cluster is also not as well separated in elder

dataset as it is compared to the younger dataset. The

mixed-age dataset performs the worst overall.

5.3 Cross-validation Classiﬁcation

Accuracy

Finally, to delve deeper towards understanding the

impact of age of activities classiﬁcation, we lever-

age state-of-the-art classiﬁcation techniques to clas-

sify the activities within younger, elder, and mixed-

age cohort. Unlike in the previous clustering scenario,

here we use 4 LDA components to classify activi-

ties. Classiﬁcation results using the cross-validation

approach for each algorithm for younger, elder and

mixed cohorts are shown in Figure 4. The machine

learning algorithms used are Linear Support Vector

Machines (Suykens and Vandewalle, 1999) , Sup-

port Vector Machine with Radial Basis Function Ker-

nel (Scholkopf et al., 1997), K-nearest Neighbors

(Larose, 2005), Decision Trees (Safavian and Land-

grebe, 1991) and Random Forests (Breiman, 2001).

Note that the metrics we use, Precision, Recall

and F1-score are standard metrics in evaluating per-

formance of machine learning algorithms for classiﬁ-

cation. The precision is the ratio of correctly classi-

ﬁed positive instances to the total number of instances

classiﬁed as positive. Recall is the ratio of correctly

classiﬁed positive instances to the total number of

An Experimental Investigation Comparing Age-Speciﬁc and Mixed-Age Models for Wearable Assisted Activity Recognition in Women

371

Figure 2: Variance distribution in PCA components for Younger (on left), Elder (on middle) and Mixed (on right) age group

dataset.

Figure 3: Activity cluster for elder (on top), younger (on

middle) and mixed (on bottom) dataset.

positive instances. The F1-score balances precision

as recall, and is given by 2 ×

Precision×Recall

Precision+Recall

We see from Figure 4 that while the algorithms do

yield similar classiﬁcation performance (Recall), we

can see that the Younger only model performs much

better, compared to the Elder only model, while the

Mixed-age model performs the worst overall in all

cases. Most confusion happens between walking and

brisk-walking, which is understandable.

5.4 Personalized Classiﬁcation Model

for Elder People

With the diversities in activities as performed by el-

ders compromising classiﬁcation accuracy (when data

from all elders was used for modeling) as presented

above, we now attempt to investigate if personalized

models can help improve classiﬁcation accuracy for

the elder cohort. In Figure 5, we present classiﬁcation

accuracy results based on the Random Forest learn-

ing algorithm (which performed the best). As we can

see, the accuracy of all activities is much higher when

a model is trained for each elder subject exclusively

- with an average improvement of 30% in activities

classiﬁcation.

6 DISCUSSIONS AND

CONCLUSIONS

In this paper, we investigate the impact of age-speciﬁc

and personalized algorithms on accuracy for classi-

fying physical activities among women using a wrist

worn wearable device. Using only accelerometer and

gyroscope sensory data, and leveraging state of the

art data mining, and classiﬁcation algorithmic tech-

niques, we demonstrate that models focusing on a

younger cohort (21-26 age group) were superior in

terms of activities classiﬁcation compared to an elder

HEALTHINF 2019 - 12th International Conference on Health Informatics

372

Figure 4: Comparison between accuracy (recall) of activities classiﬁcation for elder, younger and mixed data using different

classiﬁers.

Figure 5: Personalized classiﬁcation accuracy using Ran-

dom Forest on elder data.

only cohort (65-75 age group), with the mixed age

cohort performing the worst. We ﬁnally also showed

that for the elder only cohort, a personalized model

performed much better.

REFERENCES

Alinia, P., Cain, C., Fallahzadeh, R., Shahrokni, A., Cook,

D., and Ghasemzadeh, H. (2017). How accurate is

your activity tracker? a comparative study of step

counts in low-intensity physical activities. JMIR

mHealth and uHealth, 5(8).

Avci, A., Bosch, S., Marin-Perianu, M., Marin-Perianu, R.,

and Havinga, P. (2010). Activity recognition using

inertial sensing for healthcare, wellbeing and sports

applications: A survey. In Architecture of computing

systems (ARCS), 2010 23rd international conference

on, pages 1–10. VDE.

Banos, O., Galvez, J.-M., Damas, M., Pomares, H., and Ro-

jas, I. (2014). Window size impact in human activity

recognition. Sensors, 14(4):6474–6499.

Bao, L. and Intille, S. S. (2004). Activity recognition

from user-annotated acceleration data. In Interna-

tional Conference on Pervasive Computing, pages 1–

17. Springer.

Bar-Or, O. (1977). Age-related changes in exercise percep-

An Experimental Investigation Comparing Age-Speciﬁc and Mixed-Age Models for Wearable Assisted Activity Recognition in Women

373

tion. Physical Work and Effort G. Borg (ED), pages

255–256.

Beaver, L. (2016). The smartwatch report: Forecasts,

adoption trends, and why the market isn’t living

up to the hype. https://www.businessinsider.

com/smartwatch-and-wearables-research-\

forecasts-trends-market-use-cases-2016-9.

Benz Scott, L. A., Ben-Or, K., and Allen, J. K. (2002). Why

are women missing from outpatient cardiac rehabilita-

tion programs? a review of multilevel factors affect-

ing referral, enrollment, and completion. Journal of

Women’s Health, 11(9):773–791.

Bharti, P., De, D., Chellappan, S., and Das, S. K. (2018a).

Human: Complex activity recognition with multi-

modal multi-positional body sensing. IEEE Transac-

tions on Mobile Computing, pages 1–1.

Bharti, P., Panwar, A., Gopalakrishna, G., and Chellappan,

S. (2018b). Watch-dog: detecting self-harming activi-

ties from wrist worn accelerometers. IEEE Journal of

Biomedical and Health Informatics, 22/3:686–696.

Borg, G. (1998). Borg’s perceived exertion and pain scales.

Human kinetics.

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Burns, A., Greene, B. R., McGrath, M. J., O’Shea, T. J.,

Kuris, B., Ayer, S. M., Stroiescu, F., and Cionca, V.

(2010). Shimmer–a wireless sensor platform for non-

invasive biomedical research. IEEE Sensors Journal,

10(9):1527–1534.

De, D., Bharti, P., Das, S. K., and Chellappan, S. (2015).

Multimodal wearable sensing for ﬁne-grained activity

recognition in healthcare. IEEE Internet Computing,

19(5):26–35.

de la Concepci

on, M.

A., Morillo, L. M. S., Garc

ıa, J.

A., and Gonz

alez-Abril, L. (2017). Mobile activ-

ity recognition and fall detection system for elderly

people using ameva algorithm. Pervasive and Mobile

Computing, 34:3–13.

Del Rosario, M. B., Wang, K., Wang, J., Liu, Y., Brodie,

M., Delbaere, K., Lovell, N. H., Lord, S. R., and Red-

mond, S. J. (2014). A comparison of activity classi-

ﬁcation in younger and older cohorts using a smart-

phone. Physiological measurement, 35(11):2269.

Gupta, P. and Dallas, T. (2014). Feature selection and ac-

tivity recognition system using a single triaxial ac-

celerometer. IEEE Transactions on Biomedical En-

gineering, 61(6):1780–1786.

Hwang, H. and Haddad, R. A. (1995). Adaptive median

ﬁlters: new algorithms and results. IEEE Transactions

on image processing, 4(4):499–502.

Jiang, W. and Yin, Z. (2015). Human activity recognition

using wearable sensors by deep convolutional neural

networks. In Proceedings of the 23rd ACM interna-

tional conference on Multimedia, pages 1307–1310.

ACM.

Kaur, R. and Kaur, P. D. (2017). Review on fall detection

techniques based on elder people. International Jour-

nal of Advanced Research in Computer Science, 8(3).

Landau, H. (1967). Sampling, data transmission, and the

nyquist rate. Proceedings of the IEEE, 55(10):1701–

1706.

Larose, D. T. (2005). K-nearest neighbor algorithm. Dis-

covering Knowledge in Data: An Introduction to Data

Mining, pages 90–106.

Lee, L. and Grimson, W. E. L. (2002). Gait analysis for

recognition and classiﬁcation. In Automatic Face and

Gesture Recognition, 2002. Proceedings. Fifth IEEE

International Conference on, pages 155–162. IEEE.

Levy, B. R. and Myers, L. M. (2004). Preventive health be-

haviors inﬂuenced by self-perceptions of aging. Pre-

ventive medicine, 39(3):625–629.

Mika, S., Ratsch, G., Weston, J., Scholkopf, B., and

Mullers, K.-R. (1999). Fisher discriminant analysis

with kernels. In Neural Networks for Signal Process-

ing IX, 1999. Proceedings of the 1999 IEEE Signal

Processing Society Workshop., pages 41–48. IEEE.

Reyes-Ortiz, J.-L., Oneto, L., Sam

a, A., Parra, X., and

Anguita, D. (2016). Transition-aware human activ-

ity recognition using smartphones. Neurocomputing,

171:754–767.

Safavian, S. R. and Landgrebe, D. (1991). A survey of de-

cision tree classiﬁer methodology. IEEE transactions

on systems, man, and cybernetics, 21(3):660–674.

San-Segundo, R., Montero, J. M., Barra-Chicote, R., Fer-

nandez, F., and Pardo, J. M. (2016). Feature extrac-

tion from smartphone inertial signals for human activ-

ity segmentation. Signal Processing, 120:359 – 372.

anchez, D., Tentori, M., and Favela, J. (2008). Activity

recognition for the smart hospital. IEEE intelligent

systems, 23(2).

Scholkopf, B., Sung, K.-K., Burges, C. J., Girosi, F.,

Niyogi, P., Poggio, T., and Vapnik, V. (1997). Com-

paring support vector machines with gaussian kernels

to radial basis function classiﬁers. IEEE transactions

on Signal Processing, 45(11):2758–2765.

Sousa, W., Souto, E., Rodrigres, J., Sadarc, P., Jalali, R.,

and El-Khatib, K. (2017). A comparative analysis of

the impact of features on human activity recognition

with smartphone sensors. In Proceedings of the 23rd

Brazillian Symposium on Multimedia and the Web,

pages 397–404. ACM.

Suykens, J. A. and Vandewalle, J. (1999). Least squares

support vector machine classiﬁers. Neural processing

letters, 9(3):293–300.

Valencia, H. E., Savage, P. D., and Ades, P. A. (2011). Car-

diac rehabilitation participation in underserved pop-

ulations. Journal of cardiopulmonary rehabilitation

and prevention, 31(4):203–210.

Wang, Y., Wu, K., and Ni, L. M. (2017). Wifall: Device-

free fall detection by wireless networks. IEEE Trans-

actions on Mobile Computing, 16(2):581–594.

Wold, S., Esbensen, K., and Geladi, P. (1987). Principal

component analysis. Chemometrics and intelligent

laboratory systems, 2(1-3):37–52.

HEALTHINF 2019 - 12th International Conference on Health Informatics

374