'61=-9:1;@7.$07,-:4)6,'61=-9:1;@7.$07,-:4)6,
1/1;)475576:'$1/1;)475576:'$
-8)9;5-6;7.4-+;91+)4758<;-9)6,
175-,1+)46/16--916/)+<4;@"<*41+);176:
-8)9;5-6;7.4-+;91+)4758<;-9)6,
175-,1+)46/16--916/
%5)9;.7/7/+758<;16/.9)5->793.79<6:<8-9=1:-,+4<:;-916/%5)9;.7/7/+758<;16/.9)5->793.79<6:<8-9=1:-,+4<:;-916/
)6)4@;1+:16>-)9)*4-6;-96-;7.&016/:)6)4@;1+:16>-)9)*4-6;-96-;7.&016/:
-*)62)679;0)3<9
'61=-9:1;@7.$07,-:4)6,
)91:0+0)6,9)<*-@
1+074):76:;)6;
'61=-9:1;@7.$07,-:4)6,
-:41-)04-9
'61=-9:1;@7.$07,-:4)6,
<6)4)637,1@)
'61=-9:1;@7.$07,-:4)6,
3<6)45<91-,<
&01:76.-9-6+-"97+--,16/1:*97</0;;7@7<.79.9--)6,78-6)++-::*@;0--8)9;5-6;7.4-+;91+)4758<;-9
)6,175-,1+)46/16--916/);1/1;)475576:'$;0):*--6)++-8;-,.7916+4<:17616-8)9;5-6;7.
4-+;91+)4758<;-9)6,175-,1+)46/16--916/)+<4;@"<*41+);176:*@)6)<;0791A-,),5161:;9);797.
1/1;)475576:'$79579-16.795);17684-):-+76;)+;,1/1;)4+75576:-;)4<91-,<
1;);176"<*41:0-9;;91*<;1761;);176"<*41:0-9;;91*<;176
79;0)3<9<*-@76:;)6; )04-9)637,1@) 7=-5*-9
%5)9;.7/7/
+758<;16/.9)5->793.79<6:<8-9=1:-,+4<:;-916/)6)4@;1+:16>-)9)*4-6;-96-;7.&016/:
47*)476.-9-6+-76%1/6)4)6,6.795);176"97+-::16/47*)4%"76;9-)4#)6),),71
47*)4%"
=)14)*4-);0;;8,?,7179/47*)4%"
7447>;01:)6,),,1;176)4>793:);0;;8:,1/1;)4+75576:<91-,<-4-(.)+8<*:
&0-'61=-9:1;@7.$07,-:4)6,)+<4;@0)=-5),-;01:)9;1+4-78-64@)=)14)*4-&0-'61=-9:1;@7.$07,-:4)6,)+<4;@0)=-5),-;01:)9;1+4-78-64@)=)14)*4-
"4-):-4-;<:367>"4-):-4-;<:367>07>!8-6++-::;7;01:9-:-)9+0*-6-B;:@7<07>!8-6++-::;7;01:9-:-)9+0*-6-B;:@7<
&-95:7.':-
&01:)9;1+4-1:5),-)=)14)*4-<6,-9;0-;-95:)6,+76,1;176:)8841+)*4-;7>)9,:!8-6++-::
"741+@9;1+4-:)::-;.79;0167<9&-95:7.':-
SMART FOG: FOG COMPUTING FRAMEWORK FOR UNSUPERVISED CLUSTERING
ANALYTICS IN WEARABLE INTERNET OF THINGS
Debanjan Borthakur
1
, Harishchandra Dubey
2
, Nicholas Constant
1
, Leslie Mahler
3
, Kunal Mankodiya
1
1
Wearable Biosensing Lab, University of Rhode Island, Kingston, RI-02881, USA
2
University of Texas at Dallas, Richardson, TX-75080, USA
3
Department of Communicative Disorders, University of Rhode Island, Kingston, RI-02881, USA
Email: (kunalm@uri.edu, lmahler@uri.edu)
ABSTRACT
The increasing use of wearables in smart telehealth system
led to the generation of large medical big data. Cloud and
fog services leverage these data for assisting clinical proce-
dures.IoT Healthcare has been benefited from this large pool
of generated data. This paper suggests the use of low-resource
machine learning on Fog devices kept close to wearables for
smart telehealth. For traditional telecare systems, the signal
processing and machine learning modules are deployed in the
cloud that processes physiological data. This paper presents
a Fog architecture that relied on unsupervised machine learn-
ing big data analysis for discovering patterns in physiological
data. We developed a prototype using Intel Edison and Rasp-
berry Pi that was tested on real-world pathological speech
data from telemonitoring of patients with Parkinson’s dis-
ease (PD). Proposed architecture employed machine learning
for analysis of pathological speech data obtained from smart
watches worn by the patients with PD. Results show that
proposed architecture is promising for low-resource machine
learning. It could be useful for other applications within
wearable IoT for smart telehealth scenarios by translating
machine learning approaches from the cloud backend to edge
computing devices such as Fog.
Keywords: Dysarthria; Edge Computing; Fog Computing;
K-means Clustering, Parkinson’s Disease; Speech Disorders
1. INTRODUCTION
As described in [1] Fog is a new architecture for computing,
storage, control and networking that brings these services
closer to end users.In simple words, the decentralization of
services at the edge of the network is achieved.The compu-
tation and control closer to the sensors make the concept of
Fog a better alternative to the cloud.In our proposed architec-
ture of smart Fog, we leveraged the idea of Fog for speech
signal processing for telehealth monitoring. Speech signal
processing and Machine learning are fundamental blocks for
The research discussed in this manuscript was supported by National
Institute of Health Grant: R01MH108641.
detection and evaluation of speech disorders like dysarthria
in patients with Parkinson’s diseases that affects a significant
portion of the world population. Telehealth monitoring is
very effective for the speech-language pathology, and smart
devices like EchoWear [2] can be useful in such situations.
Several signs indicate the relationship of dysarthria, speech
prosody, and acoustic features. As authors in [3] mentions
dysarthria always accompanies patients with Parkinson’s
disease Characterized by the monotony of speech, reduced
stress, variable rate, imprecise consonants, and a breathy and
harsh voice Authors in [4] [5] suggested that extreme F0
variation and range in speakers with severe dysarthria exist.
Another important acoustic feature for dysarthria is the am-
plitude of the speech uttered by the patients with Parkinson’s
disease. In [6] authors mention about reduced vocal intensity
in hypokinetic dysarthria in Parkinson disease. This paper
presents a Fog Computing architecture, SmartFog that relied
on unsupervised clustering for discovering patterns in patho-
logical speech data obtained from patients with Parkinson’s
disease(PD). The patients with PD use smartwatch while per-
forming speech exercises at home. The speech data were
routed into the Fog computer via a nearby tablet/smartphone.
The Fog computer extracts loudness and fundamental fre-
quency features for quantifying pathological speech. The
speech features were normalized and processed with k-means
clustering. When we see an abnormal change in features,
results are uploaded to the cloud. In other situations, data is
only processed locally. In this way, Fog device could per-
form "smart" decision on when to upload the data to cloud
backend and when not. We developed two prototypes using
Intel Edison and Raspberry Pi. Both of the prototypes were
used for comparative analysis of computation time. Both
systems were tested on real world pathological speech data
from telemonitoring of patients with Parkinson’s disease. The
increasing use of wearables in smart telehealth system led to
the generation of huge medical big data [7, 8, 9, 10]. The
telehealth services leverage these data for assisting clinical
procedures. This paper suggests use of low-resource ma-
chine learning on Fog devices kept close to the wearable for
472978-1-5090-5990-4/17/$31.00 ©2017 IEEE GlobalSIP 2017
Fig. 1. Proposed SmartFog architecture for enhanced analytics in wearable internet of medical things. It is developed and
evaluated for telehealth application.
smart telehealth. For traditional telecare systems, the signal
processing and machine learning modules are deployed in
the cloud that processes physiological data. In our analysis,
we have chosen the average fundamental frequency(F0) in
hertz and average intensity in decibel for K-means cluster-
ing analysis. The algorithm efficiently clusters the unlabeled
data into groups of similarity that was done on the fog plat-
form.One use of this analysis can be for real time Parkinson’s
phenotypic sub-groupings based on the clusters.
2. RELATED WORKS
2.1. Telehealth and Associate Challenges
The Fog Architecture shifts computation, networking, and
storage to the edge of the network.Various authors have de-
scribed a different architecture for Fog. FIT as described
in [11] has the following components: (1) Smartwatch; (2)
Fog computer; and (3) Cloud backend. In the paper [12]
authors presents an effort to conceptualize WIoT concern-
ing their design, function, and applications.The paper [13]
demonstrates the Fog Data that is a service oriented archi-
tecture for Fog computing.This literature emphasizes the im-
portance and versatility of Fog computing.The challenges IoT
faces are described in [1] are the requirement of stringent low
latency, IoT applications such as gaming, virtual reality de-
mands this. The issue of Network Bandwidth and Resource-
constrained devices are another challenges to the emerging
field of IoT. Thus arises the importance of fog that distributes
computing, control, storage and networking functions closer
to the end user [1].
2.2. Big data and Telehealth
Tele-Health utilizes the recent developments of Big Data in
the context of biomedical and healthcare.Fields like medical
and health informatics, translational bioinformatics, sensor
informatics etc can avail the benefit of the personalized infor-
mation from a diverse range of data sources[14].Authors in
[13], proposes, validates and evaluates Fog Data architecture
for Fog computing. The proposed architecture is a low power
embedded computer that carries out data mining and analy-
sis on data collected from various wearable sensors used for
telehealth applications.[15] mentions about European project
’PERFORM’ that is a sophisticated multi-parametric system
FOR the continuous effective assessment and monitoring of
motor status in Parkinson Disease and other neurodegener-
ative diseases. It provides a telehealth system for remote
monitoring of Parkinson Patients.The paper also summarizes
the technical performance of the system and the feedback re-
ceived from the patients in terms of usability and wearabil-
ity.We in our work used Parkinson speech data analysis for
our proposed smart-fog framework.
2.3. Wearable Internet of Thing for Telehealth
IoT Device that interacts with the fog node is composed of
sensors that are capable of collecting and transmitting data
via wireless means.IoT allows handling of objects remotely
across the network. The versatility of IoT makes it more suit-
able for smart grids, smart homes, smart cities and wearable
health monitoring systems.This paper focuses on the health
aspect of IoT.Integration with the internet offers IoT devices
an IP address for better communication. Big data and Inter-
net of Things work collectively, and we tried to leverage this
relationship in our proposed architecture. We used Raspberry
Pi and Intel Edison as Fog computing device for the analysis
discussed in this paper. Fog Interface as described in [11, 16]
is a low-power embedded computer that acts as a smart in-
terface between the smartwatch and the cloud. It is used for
collection, storage, and processing of the data before sending
features to secure cloud storage. Raspberry Pi is used as Fog
device for this work. The Raspberry Pi is a series of credit
card-sized single board computers that has gained much pop-
ularity owing to its small size and multipurpose utility. It has
ARM compatible central processing unit and on-chip graphic
processing units.
473
Fig. 2. K-means clustering plot
2.4. Fog Computing
Cloud computing provides shared computer processing and
data analysis, in other terms Cloud is a hub of computing
resources such as computer networks, servers, storage, and
services. The availability of high-capacity networks, low-
cost computers, and storage devices makes cloud a highly
demanded service for the users seeking for high computing
power. Cloud can interact with the IoT device via the fog
node. This paper concentrates on the side of fog computing,
which allows users a higher computing power at the instru-
ment end.Reliance on fog will help cut the costs associated
with the Cloud to an extent.
3. FOG-BASED LOW-RESOURCE MACHINE
LEARNING
3.1. Feature Extraction
Feature engineering is the initial step in any machine learning
analysis. It is the process of proper selection of data metric
to input as features into a machine learning algorithm. In K-
means clustering analysis, the selection of features that are
capable of capturing the variability of the data is essential for
the algorithm to find the groups based on similarity. Our sub-
jects were patients with Parkinson’s disease and the features
chosen were the average fundamental frequency (F0) and Av-
erage amplitude of the speech utterance. Speech data from the
patients with Parkinson’s disease were collected. For analy-
sis, 164 speech samples were considered.These samples com-
prised of sound files with utterances as a short /a/, a long /a/,
a normal then high pitched /a/, a normal then low pitched /a/
and phrases. The feature extraction is done with the help of
Praat scripting language [17]. For pitch , the algorithm per-
forms an acoustic periodicity detection on the basis of an ac-
curate autocorrelation method. For calculating the intensity
the values in the sound are first squared, then convolved with
a Gaussian analysis window. The intensity is calculated in
decibels.
3.2. K-means Clustering
K-means clustering is a type of unsupervised learning, that
is used for exploratory data analysis of no labeled data [18].
K-means is a method of vector quantization and is quite ex-
tensively used in data mining.The goal of this algorithm is to
find groups in the data, the number of groups represented by
the variable K. The algorithm works iteratively to assign each
data point to one of K groups based on the features that are
provided. The input to the algorithm are the features, and the
value of K. K centroids are initially randomly selected, then
the algorithm iterates until convergence. This algorithm aims
to minimize the squared error function J.
J =
K
X
k=1
X
i∈c
k
||x
i
− m
k
||
2
Where Euclidean distance is chosen between the data point
and cluster center. Feature Engineering is an essential part
of this algorithm. Authors in [19], uses optimized K-means,
that clusters the statistical properties such as the variance of
the probability density functions of the clusters extracted fea-
tures. In [20] the authors have used clustering on a database
containing feature vectors derived from Malay digits utter-
ances. The features extracted in [20] were the Mel-Frequency
Cepstral Coefficients (MFCC). In our work, we have chosen
the average fundamental frequency and average intensity as
features extracted from the speech files for applying K-means
clustering.
4. RESULTS & DISCUSSIONS
For our analysis we have chosen speakers with 164 speech
samples with utterances that are a short /a/, a long /a/, a nor-
mal then high pitched /a/,a normal then low pitched /a/ and
phrases. The features were chosen are average fundamen-
tal frequency and intensity. Feature extraction is done us-
ing praat [17] an acoustic analysis software and using Praat
474
scripts that use standard algorithms to extract pitch and in-
tensity mentioned in the discussion above. The results are
shown in the form of plots.The k-means clustering analysis
is done on Python programming language.The plots below
show the Clusters of the speech data samples used in the
analysis.Different colors represent different mutually exclu-
sive groups. The analysis is done with 2, 3 and 4 number of
clusters, i.e. the value of k chosen as 2 and 3 and four respec-
tively.
Figure (a) shows the K-means clustering plot for 2 clusters
shown with different colors.The python script is run on Rasp-
berry Pi and Intel Edison to generate the results.
Figure (b) displays the k-means cluster plot for 4 Clusters
designated with four different colors in a 3D plot.Each ob-
servation belongs to the cluster with the nearest mean in k-
means clustering.We have used k-means for feature learning
performed in the fog device.
Figure(c) shows the k-means clustering plot for 3 clusters
with different colors in 3D.The python script run on Rasp-
berry Pi and Intel Edison was used for generating the results
displayed in the figure.
4.1. Performance Comparison
The Raspberry Pi provides a low-cost computing terminal.
The Edison is a deeply embedded IoT computing module.
There is a difference of processor speed and power consump-
tion in Edison and Raspberry Pi . The Machine Learning al-
gorithms were run on both of the devices and their Run time,
average CPU usage and Memory usage have been calculated.
Fig. 3. A comparison of Intel Edison and Raspberry pi.
Figure 3 shows the comparison of Intel Edison and rasp-
berry Pi fog devices.The ideal system will minimize runtime,
maximize CPU usage, and use a modest amount of memory.
The raspberry Pi either outperformed or matched the Edison
in each of this criterion. The raspberry Pi was not capable
of generating a graphical output for this type of analysis in a
real-time response threshold of 200ms. However, without a
need for complex graphics, the raspberry Pi was able to reach
the threshold clocking in at 160ms.
5. CONCLUSIONS
Fog computing emphasizes proximity to end-users unlike
cloud computing along with local resource pooling, reduction
in latency, better quality of service and better user experi-
ences. This paper relied on Fog computer for low-resource
machine learning. As a use case, we employed K-means
clustering on clinical speech data obtained from patients with
Parkinson’s disease (PD).Proposed Smart-Fog architecture
can be useful for health problems like speech disorders and
clinical speech processing in real time as discussed in this pa-
per.Fog computing reduced the onus of dependence on Cloud
services with availability of big data.There will be more as-
pects of this proposed architecture that can be investigated in
future.We can expect Fog architecture to be crucial in shap-
ing the way big data handling and processing happens in near
future.
6. ACKNOWLEDGEMENT
Authors would like to thank George and Anne Ryan Institute
for Neuroscience for their support and help.
7. REFERENCES
[1] Mung Chiang and Tao Zhang, “Fog and iot: An
overview of research opportunities,” IEEE Internet of
Things Journal, vol. 3, no. 6, pp. 854–864, 2016.
[2] Harishchandra Dubey, Jon C Goldberg, Moham-
madreza Abtahi, Leslie Mahler, and Kunal Mankodiya,
“EchoWear: smartwatch technology for voice and
speech treatments of patients with parkinson’s disease,”
in Proceedings of the conference Wireless Health. ACM,
2015.
[3] Shunan Zhao, Frank Rudzicz, Leonardo G Carvalho,
César Márquez-Chin, and Steven Livingstone, “Auto-
matic detection of expressed emotion in parkinson’s dis-
ease,” in IEEE ICASSP, 2014.
[4] Tiago H Falk, Wai-Yip Chan, and Fraser Shein, “Char-
acterization of atypical vocal source excitation, tempo-
ral dynamics and prosody for objective measurement of
dysarthric word intelligibility,” Speech Communication,
vol. 54, no. 5, pp. 622–631, 2012.
[5] Rupal Patel, Katherine C Hustad, Kathryn P Con-
naghan, and William Furr, “Relationship between
prosody and intelligibility in children with dysarthria,”
Journal of medical speech-language pathology, vol. 20,
no. 4, 2012.
475
