1

Synchronous Big Data Analytics for Personalizedand Remote Physical Therapy

Prasad Calyam, Anup Mishra, Ronny Bazan Antequera, Dmitrii Chemodanov,Alex Berryman, Kunpeng Zhu, Carmen Abbott, Marjorie Skubic

Center for Eldercare and Rehabilitation Technology, University of Missouri-Columbia, USA; The Samraksh CompanyEmail: {calyamp, skubicm}@missouri.edu, {akmm94, rcb553, dycbt4}@mail.missouri.edu, {alex.berryman, tony.zhu}@samraksh.com,

abbottc@health.missouri.edu

Abstract—With gigabit networking becoming economically fea-sible and widely installed at homes, there are new opportunitiesto revisit in-home, personalized telehealth services. In this paper,we describe a novel telehealth eldercare service that we developedviz., “PhysicalTherapy-as-a-Service” (PTaaS) that connects aremote physical therapist at a clinic to a senior at home. Theservice leverages a high-speed, low-latency network connectionthrough an interactive interface built on top of Microsoft Kinectmotion sensing capabilities. The interface that is built usinguser-centered design principles for wellness coaching exercisesis essentially a ‘Synchronous Big Data’ application due to its: (i)high data-in-motion velocity (i.e., peak data rate is ≈400 Mbps),(ii) considerable variety (i.e., measurements include 3D sensing,network health, user opinion surveys and video clips of RGB,skeletal and depth data), and (iii) large volume (i.e., several GB ofmeasurement data for a simple exercise activity). The successfulPTaaS delivery through this interface is dependent on the veracityanalytics needed for correlation of the real-time Big Data streamswithin a session, in order to assess exercise balance of the seniorwithout any bias due to network quality effects. Our experimentswith PTaaS in an actual testbed involving senior homes in KansasCity with Google Fiber connections and our university clinicdemonstrate the network configuration and time synchronizationrelated challenges in order to perform online analytics. Ourfindings provide insights on how to: (a) enable suitable resourcecalibration and perform network troubleshooting for high userexperience for both the therapist and the senior, and (b) realize aBig Data architecture for PTaaS and other similar personalizedhealthcare services to be remotely delivered at a large-scale in areliable, secure and cost-effective manner.

Index Terms—Smart Health Care, Personalized Remote Phys-ical Therapy, Synchronous Big Data, Gigabit Networking App

I. INTRODUCTION

Particularly in the case of eldercare, health and wellnessservices can be proactively delivered from remote clinic sitesto allow unobtrusive, continuous monitoring and personal-ized coaching in homes of seniors. Given the challengesbeing faced in providing smart health care services due tolimited facilities as well as human resources to large ge-ographically distributed senior populations, a new class oftelehealth services are essential to deliver healthcare on a real-time basis. With Gigabit networking becoming economicallyfeasible and widely installed in homes through city-supportedprograms such as Google Fiber in several US cities (e.g.,

This work was supported by National Science Foundation under awardnumbers: CNS-1346789, ACI-1246001 and ACI-1245795, and the Departmentof Energy under award numbers: DE-SC0001331 and DE-SC0007531. Anyopinions, findings, and conclusions or recommendations expressed in thispublication are those of the author(s) and do not necessarily reflect the viewsof the united states government or any agency thereof.

Kansas City), there are new opportunities to revisit in-home,personalized telehealth services. The opportunities include theconceptualization and implementation of Gigabit applicationsthat use the latest advances in sensing, high-definition video-based communication, and cloud computing. The Gigabitapplications within novel telehealth service compositions canenable: early interventions, better health outcomes throughfewer hospital visits, better one-on-one interactive care, andultimately significant cost savings in the relatively expensiveeldercare health market sector.

In the diverse field of eldercare, novel telehealth servicesin the context of “Physical Therapy” are particularly attractivefor both hospitals and senior patients. Physical and cognitivelimitations of seniors as they age can present difficulties forseniors to travel to a physical therapy clinic for on-goinginteractions with a therapist. In this paper, we describe a noveltelehealth service that we developed viz., “PhysicalTherapy-as-a-Service” (PTaaS) that connects a remote physical therapistin a clinic to a senior at their private home.

The system design and components of the service leveragea high-speed, low-latency network connection through aninteractive interface built on top of Microsoft Kinect motionsensing capabilities. We utilize the openly available KinectApplication Programming Interface (API), C# language andWindows Presentation Foundation (WPF) technologies on.Net Framework 4.5 in our interactive interface developmentenvironment. The service hardware components in additionto the Kinect sensors include measurement point appliancesfor application and network monitoring services providedby Narada Metrics [1]. We also use a cloud-hosted storageenvironment provided by Global Environment for NetworkInnovations [2] for synchronous analytics of the data streamsin the PTaaS session. Two separate interface variants havebeen developed, one for the therapist side, and another for thesenior side, in order to address the unique needs of the visualinteraction (e.g., the therapist can use voice commands) andexercise assessment (e.g., the therapist can generate exerciseactivity reports for discussion with the senior).

Our PTaaS interface that is built with user-centered designprinciples for in-home senior directed health and wellnesscoaching exercises is more than a typical videoconferencingapplication. It integrates 2D video feeds between the therapistand senior along with 3D sensing data to provide an immersiveexperience, and thus is essentially a ‘Synchronous Big Data’application. The Big Data characteristics are due to its: (i)high data-in-motion velocity (i.e., peak data rate is ≈400Mbps and average is over 100 Mbps), (ii) considerable variety(i.e., measurements include 3D sensing, network health, user

2

opinion surveys and video clips of RGB, skeletal and depthdata), and (iii) large volume (i.e., several GB of measurementdata for a simple exercise activity).

The successful PTaaS delivery through this interface isdependent on the veracity analytics needed for correlation ofthe real-time Big Data streams within a session, in order toquantitatively assess balance of the senior without any bias dueto network quality effects. More specifically, a therapist needsto be able to confidently assess whether non-ideal performancein the exercise forms of a senior is being impacted due to lag innetwork communications for the data-intensive interactive ses-sion, or in fact are due to the physical and cognitive limitationsof the senior owing to aging. Our analytics approach involvesa network quality estimation module whose output (i.e., greenor red indicator on the interface for network quality status)provides visual evidence to the senior and therapist regardingany undesirable network status that degrades interface userexperience and may cause any exercise mis-assessment.

We investigate our synchronous Big Data nature of ourPTaaS interface through experimental testbeds, both in a lab-oratory environment as well as in an actual testbed involvingsenior homes in Kansas City with Google Fiber connectionsand our university clinic site in Columbia, Missouri. Ourexperiment results clearly demonstrate the network configu-ration and time synchronization related challenges in orderto perform online analytics on the data streams in a PTaaSsession for different physical therapy activities such as balance,sway detection, and analysis of walking gait patterns. OurPTaaS interface test plan and experiment results discussionare focused around the synchronous big data analytics forideal network quality condition scenarios, and bad networkhealth quality scenarios. We also qualitatively study how animmersive telehealth setup for physical therapy with our PTaaSinterface compares with a face-to-face session, and explorethe pros and cons of our approach from user experience and‘therapy effectiveness’ standpoints.

Our findings from the PTaaS interface development andtestbed experiments in this paper provide insights on how to:(a) enable suitable resource calibration and perform networktroubleshooting to achieve good user experience for both thetherapist and the senior, and (b) realize a Big Data architecturefor PTaaS and other similar smart healthcare services to bedelivered at a large-scale in a reliable, secure and cost-effectivemanner. Thus, our work explores a relatively less exploredresearch problem of combining analytics of telehealth userexperience in conjunction with Big Data model characteriza-tion of a Gigabit application that is heavily network resourceintensive, while also being highly sensitive to network lags.It also represents an innovative and novel approach of usingbroadband and cloud technologies for ‘proactive’ healthcarethat is direly needed to minimize the rising health care costsin our society.

Our PTaaS development efforts are part of our engagementwithin the vibrant Gigabit application developer communitythat is growing through initiatives such as Mozilla Ignite andUS Ignite [3] being supported by the National Science Foun-dation within the USA. Specifically, the US Ignite organizationis fostering teams to develop the next-generation of Internetapplications that provide transformative public benefit in theareas of national priorities such as healthcare, public safetyand education.

The remainder of this paper is organized as follows: Sec-tion II describes related work. In Section III, we formallydescribe the PTaaS interface components and the synchronousBig Data as well as high-speed network connectivity issuesinvolved in our telehealth service development. Section IVdetails our human subject testing methodology and qualitativeresults within actual testbeds. Following this in the samesection, we also present controlled network results from oursynchronous Big Data analytics. Section V outlines a BigData architecture for PTaaS for large-scale delivery. Section VIconcludes the paper.

II. RELATED WORK

There have been earlier studies on remote physical therapyand other exercise systems. In [4], two different virtual realityapplications are proposed using Vicon and Kinect technolo-gies, for both local and remote physical therapy. Avatars ofthe patient and the therapist are created in real-time withinthe applications, and the communications between the patientand therapist is through each other’s Avatar models. In anotherrelated study [5], a motion capture application is proposed forremote orthopedics rehabilitation. This system uses a set ofgaming motion capture technologies (e.g., Wii, Kinect, andPlaystation Move) for exercise data collection in differentscenarios involving clinic and home sites. In remote sessions,the exercise data are not exchanged in real-time through avideoconferencing setup, but rather are sent over a networkfor offline (not real-time) visualization and feedback. Theoffline nature was pertinent in their case given the purposeof their application was to allow a patient to learn how toperform exercises correctly at the hospital, and later executethem at their home without any therapist by their side. Otherworks have investigated exercise coaching with vision througha socially assistive robot coach [6], pre-recorded videos withautomated coaching advice [7], the Wiimotes attached to limbsfor movement-training exercises [8], and wearable sensingwith iPods for tracking home balance exercises [9].

In contrast to these works, our PTaaS application approachinvolves a peer-to-peer communication model for sending andreceiving skeletal data (i.e., 2D skeletal visual showing jointlocations) as well as RGB, depth, and audio data in real-time,all using a high-speed network setup in a wide-area setting.Hence, our approach aims to establish a more realistic andhighly interactive/immersive communication between therapistand patient in the comfort of a clinic and private home,respectively.

In terms of remote access to video streams, earlier workshave proposed techniques to perform snapshot captures forstreaming over low-bandwidth network paths [10]. In addition,real-time video compression schemes for users connected toaccess networks with high latency and different available band-width levels have been proposed [11]. In [12], authors presentnetwork traffic prioritization and load balancing schemes toenable Skype sessions over portable devices with satisfactoryuser experience for telehealth applications such as video-consultation-call in a home area network. Our applicationdesign goals are different and we aim for video streams to bedelivered without compression over wide-area network pathswith low-latency characteristics, and take advantage of high-speed Gigabit networking between patient homes and a remotetherapist clinic. Our network-awareness work builds upon the

3

Fig. 1: Physical Therapy-as-a-Service System Setup

earlier work in [13], where real-time network measurementsare used for notifying users regarding network quality statusin order to guide them in their real-time interactions withremote scientific instruments. The difference in this workis that we use the network quality status to help improvethe remote physical therapy effectiveness in a interactivevideoconferencing context supplemented by 3D sensor data.

Our PTaaS application requires on-line and post data analy-sis of synchronous Big Data within a remote physical therapysession. Our literature survey indicated that there are fewworks such as [14] and [15] that deal with synchronousBig Data analytics with applications data streams. The au-thors in [14] analyze social network data and show howinteractive Big Data analysis can help formulate and assesshypotheses in a rapid and iterative manner. The work in [15]considers distributed system optimization strategies to enablesynchronous parallel processing of Big Data streams within acloud platform. Following the definition in [16], these worksas well as our work can be viewed as Big Data due tothe nature of the data that pushes us to look beyond tried-and-true methods that are in the current state-of-the-art. Webasically use the common data analysis steps of: (i) integrating,cleaning and transforming data, and (ii) actual data analysiswith aggregation and/or visualization - in order to solve uniqueanalytics challenges. Our unique analytics challenge can beseen in the time synchronization issues within the data streamsin order to real-time align the various activities between thetherapist and the patient. Our PTaaS data streams analysismethodology correlates the perceived user experience withthe various network quality metrics for different video streamconfigurations (i.e., skeletal, RGB, depth) with synchronizedaudio for helping the therapist verify the balance assessmenthypotheses without excessive processing delays.

It has been observed in earlier studies such as [17] and [18]that sensing data from Kinect is prone to accuracy errors. Theerrors relate to image distortions in the data capture processesdue to the factory pre-calibration issues that specially impact atthe depth data-type level. To counter this problem and increaseaccuracy in the data capture process, the authors in [17] relyon indoor control field to calibrate Kinect sensor along withthe accurate internal parameters of source and destination sidecameras and their relative pose. In [18], authors handle theerrors using a method that involves modeling the intrinsic andextrinsic parameters of the distortions in the depth image data.

In our PTaaS application context, these accuracy errors donot significantly impact effectiveness of the remote physicaltherapy sessions because our interactive interface applicationdoes not heavily rely on depth data accuracy. For properfunctioning of our interactive interface application, the RGBand Skeletal data quality and accuracy are more critical tominimize errors due to miscommunication in those data types.Also, our focus on RGB and Skeletal data addresses the needfor the physical therapist to capture the real representation ofthe subjects, timelines of interaction and follow-through whenperforming exercises. Moreover, in our recent work [19], ourexperiments with the PTaaS application in a controlled settingdemonstrated that Kinect skeletal models offer acceptableaccuracy for use in a remote physical therapy system setting.

III. PHYSICAL THERAPY-AS-A-SERVICE (PTAAS)

In this section, we first present the PTaaS applicationrequirements and system overview. Following this, we detailthe salient components of the PTaaS system and describe theuser-centered design principles. Next, we show how the PTaaSinteractive interface is essentially a synchronous Big Dataapplication. Lastly, we describe the importance of high-speednetwork connectivity for our PTaaS system and methods toreduce peak data rates for lower resolution modes of operation.

A. PTaaS System Overview

Figure 1 shows the system design for the immersive in-teractions between the Therapist and the Patient using anetwork overlay setup based on virtual link (Layer 2) tech-nologies. The overlay path is needed for fast data movementof video, audio, RGB, depth and skeletal data for real-timedisplay of gait and other movement parameters at both ends.Given that private home 1Gbps access network connections ofGoogle Fiber customers in Kansas City (or any other InternetService Provider) do not have public IP addresses, peer-to-peer networked applications such as our PTaaS interactiveinterface needs to rely on a overlay setup to bind the portsbetween the Therapist and Patient applications, and to allowcustom protocol communications. On each side, the interactiveinterface uses a Kinect device along with a local computer(specifications: Windows 7 64 bits, RAM 4 GB, HDD 500GB,Gigabit NIC) that is mounted on a mobile cart that has a largedisplay (specification: 1920x1080 px).

4

Fig. 2: PTaaS Interface Architecture and Components

For the out-of-band communications to support data pro-cessing with session-related data logs, we utilize regularInternet (Layer 3) protocols. The application and networkdata on both sides are processed locally for tasks such aspreliminary sanitization and meta data annotation (e.g., datafolder naming), and are immediately sent to a cloud-hosteddatabase for detailed data processing that help with later ac-tivity report discussions between the Therapist and the Patient.We utilized the GENI Rack at University of Missouri (MU)for the database, and also hosted the peer-to-peer applicationorchestration signaling co-ordination module also within theGENI Rack that is part of the overlay setup. For thesepurposes, the GENI Rack was configured with three virtualmachines provisioned using the ESXi hypervisor. Last butnot the least, the entire application and network measurementdata collection and aggregation as well as visualization areperformed with the Narada Metrics software [1] that we havepreviously developed as a end-to-end measurement framework.

B. PTaaS Components

1) Interactive Interface: Figure 2 shows the PTaaS interfacerelated software architecture, and the various components thatuse the Microsoft Kinect motion sensing capabilities. Weutilize the openly available Kinect Application ProgrammingInterface (API) within the Kinect SDK v1.8, C# languageand Windows Presentation Foundation (WPF) technologies on.Net Framework 4.5 in our interactive interface developmentenvironment. The main component that connects the Kinectsensor to our application is the ‘Kinect Service Module’that collects the various data sets during the session, andwrites it to the network interface for both local and remoteaccess. The individual data streams on both the sender-side andreceiver-side are accessed using the corresponding network IPaddresses and respective port numbers. The remote access en-ables the interactive videoconferencing feature of the interface,whereas the local access enables the preliminary processingand the data export into the public cloud-hosted database. Forthe local processing of raw Kinect data within the ‘Kinect

Data Processor’ component to analyze the saved activities andgenerate information for data correlations, we utilize the fileformats and processing features such as ‘rewind’ and ‘play’of the Kinect Studio software.

The Therapist and Patient each have a slightly differentsoftware configuration in order to suit the respective user-specific requirements. The screenshots of the ‘InteractivePTaaS Interface’ that we developed for the Therapist andPatient sides are shown in Figure 3. The Therapist side hasadditional real-time display of detailed sensing and networkquality data to help with the exercise assessments. To enableconvenient administration of exercises, the Therapist sideinterface has support for voice commands and keyboard-mouseinput handling that allows for start/stop markers within databetween exercise activities. Note that the Therapist sees alarger scale frame of the Patient’s view, shown on the leftpanel; whereas, the Patient sees both sides views in a equalsized panels side-by-side. The bottom right shows the skeletaldata visualization within the session. The network strengthstatus bar is either ‘green’ or ‘red’ in color based on theoutput of the network quality estimation module within the‘Active Data Processor’ explained in the following sub-section.This visual format was designed through multiple experimentsinvolving users, and by gathering feedback and refining theinterface over multiple iterations during the development stage.Overall, this visual format of the interactive interface helpswith effective involvement of both the Therapist and thePatient within the session, and provides a convenient methodto administer the different exercise activities.

2) Network Quality Estimator: Given that a robust net-work architecture with predictable end-to-end performance isessential for the interactive monitoring and coaching in thePTaaS session, we deeply instrumented the PTaaS systemto obtain measurements for network quality estimation. Bothactive measurements (i.e., end-to-end TCP throughput, round-trip time (RTT) delay, jitter and packet loss) and passive mea-surements (i.e., transfer rate on uplink of the local interface,download rate at remote interface) were collected using theNarada Metrics measurement framework.

5

(a) Therapist View

(b) Patient View

Fig. 3: PTaaS Interactive Interface Screenshots

We scheduled conflict-free active measurement tests withIperf and Ping tools using the Narada Metrics active measure-ment scheduler [1]. Iperf tool tests that are computation andchannel intensive are performed offline in a periodic samplingmanner with inter-sample times of 30 minutes to 1 hour, beforethe use of the PTaaS application on a network path. However,the lightweight Ping tool tests are performed inline with thePTaaS application with a sampling rate of 30 seconds. Thepassive measurements were in-line with the application trafficand were collected using the popular Wireshark packet capturetool, as well as the ‘Windows Network Interface’ countersviz., BytesReceivedPersec and BytesSentPersec.In order to keep the measurement tap to be non-intrusive, weexperimented with different settings of the sampling period ofthe interface statistics, and chose a sampling rate of 10 secondswithin the PTaaS interface.

The raw measurement data size for a simple session wasseveral GB for the active and passive measurements. Westored all of the summarized measurement data in JSONformat and allowed for publish (from the Interface to sendraw measurements) and subscribe (from the Interface to re-quest processed/correlated measurements) within a session. Inaddition, detailed metrics information, graphs and other drill-down features for measurement data analysis were providedthrough a customized Narada Metrics Dashboard shown inFigure 4. The measurement analysis through Narada Metricsgreatly helped in the resource provisioning and end-to-endperformance monitoring for the PTaaS system and helpedovercome several performance bottleneck challenges. For in-

Fig. 4: Customized Narada Metrics Dashboard for PTaaS

Fig. 5: Throughput of different PTaaS Interface Modes

stance, in a case relevant to network resource provisioning, wewere able to characterize how much bandwidth is consumedby the PTaaS interface for four application data streams: RGB,Skeleton, Audio and Depth. We were able to observe that theSkeleton and Audio data have minimal bandwidth consump-tion, the RGB data has medium bandwidth consumption, andthe Depth data was a heavy consumer of network bandwidthconsumption as shown in Figure 5.

In a different case relevant to performance troubleshooting,using a “Patrol App” customization across multiple vantagepoints on the wide-area network with a unified dashboard ofthe various metrics across locations, we were able to identifyoptimal and often suboptimal application configurations insessions with various homes. As shown in Tables I and II,we were able to detect actual deployment scenarios wherebottlenecks due to mis-configuration of network ports (dueto duplex mis-match of Ethernet interfaces) that degraded theinteractivity and videoconferencing quality.

Based on the various measurement functions supportedwithin Narada Metrics, we are developing an intelligent ‘Net-work Quality Estimator’ module within the PTaaS system thatis able to aggregate and correlate the various active and passivemeasurements of application and network performance. Thiswill allow us to provide tangible performance intelligenceto the PTaaS interface, and ultimately to the Therapist formaking proper exercise form assessments. Our network qualityestimator at the basic form can be represented as a neuralnetwork as shown in Figure 6. Considering only the activemeasurements, there are four dimensions of end-to-end net-work path performance that include: throughput or available

6

TABLE I: MU Therapist side performance gathered from Narada Metrics

Activity under Test Available Bandwidth, Mbps Jitter, mlsec Loss, % RTT, mlsec Upload Rate, Mbps Download Rate, MbpsSingle leg stance 3.91 14.53 73.84 30.67 0.34 1.94

TABLE II: Kansas City Senior side performance gathered from Narada Metrics

Activity under Test Available Bandwidth, Mbps Jitter, mlsec Loss, % RTT, mlsec Upload Rate, Mbps Download Rate, MbpsSingle leg stance 4.9 6.30 0.86 15.75 1.3 3.16

Fig. 6: Model of Network Quality Detector

bandwidth Bav , delay D, losses in the network path L andpacket jitter J . Our neural network model for the activemeasurements can be expressed mathematically as:

φ = φ(wi,B ·Bav, wi,D ·D,wi,L · L,wi,J · J) (1)

where wi,B , wi,D, wi,L, wi,J are coefficients of different ap-plication modes (i = ¯1, d). If φ > 0, we can say that networkis not a factor in the assessment of Patient exercise perfor-mance. The opposite is true (i.e., network is a degrading andmisleading factor for Patient exercise performance assessment)for φ < 0; Note that φ = 0 is the baseline and φ can beany of the different functions such as sign or tanh. Ourapproach is to construct such individual models for differentmeasurement data sets at the application and network levels,and use a combined closed-form mathematical expression thatcan predict the network quality status for any given inputsof measurements in an actual deployment. Note that exactsolution of such an approach is outside the scope of this paperand is part of our future work. With regards to our currentwork, our network quality estimator uses a set of thresholdsbased on empirical performance history to output either a‘green’ or ‘red’ status signal displayed on the PTaaS interfacein the ‘Network Strength’ panel shown in Figure 3.

C. Synchronous Big Data ModelThe network bandwidth consumption results in Figure 5

show how the PTaaS interface has high velocity when thereis application data-in-motion. We have consistently observedpeak data rates of ≈400 Mbps and average data rates of over100 Mbps for simple exercise activity tasks with the PTaaSinterface between an actual Therapist and Patient in a closed-network setting. In addition, given the considerable variety ofdata involved in a PTaaS session such as 3D sensing, networkhealth, user opinion surveys and video clips of RGB, skeletal

and depth data - we can appreciate the various quantitativeand qualitative information that needs to be correlated to makesense of the data. The user opinion surveys form more of thequalitative data of user’s quality of experience (QoE), howeverwe can adopt a ‘Mean Opinion Score’ (MOS) ranking scaleof 1 - 5, where 1 rating is Poor and 5 is Excellent.

Further, the data volumes we have seen in a PTaaS sessionare quite high, and are several GB of measurement data fora simple exercise activity. For instance, 1 second of video ata resolution of 1280 x 960 pixels for depth at 30 frames persecond generates approximately 140 MB of data. Given that atypical PTaaS session lasts for several tens of seconds or evenminutes as shown in Figure 7, the data generated quickly canexceed several GB of measurement data. If we configure thePTaaS interface to collect gait, bed and motion sensor data toextend it for home fall-detection and gait analysis and healthchange alerts as done in previous work [20]- [22], the datavolume can be significantly large and again on the order ofseveral tens of GB (we noticed worst case data collection canbe ≈150 GB per hour).

Based on the above discussion, we can argue that the real-time PTaaS interface requirements, and application resourceconsumption behavior can be characterized as being of the‘Synchronous Big Data’ nature. Consequently, it requires usto integrate new methods to integrate, clean and transform datathat is in different scales and is collected at different samplingrates (or frequencies) for actual data analysis with aggregationand/or visualization. Moreover, our need to solve uniqueanalytics challenges described in Section II is a guideline forthe veracity needed within our data analysis for correlation ofthe real-time Big Data streams within a session for differenttime frames, in order to quantitatively assess balance of thesenior without any bias due to network quality effects. Asshown in Figure 7, the time frames can relate to differentexercise forms such as Single leg stance (eyes open), Tandemwalk (eyes open), and Tandem walk (eyes closed). We needto divide our analytics for different time frames depending onthe data nature, and use significant computation power to lookat meaningful ways to query data and correlate the individualanalyses to find any network effects that impact the PTaaSexercise assessments, e.g., when a Therapist observes a Patientis having balance problems during the exercises.

D. Importance of High-speed Network Connectivity

Herein, we discuss the importance of having a high-speedbandwidth, low latency network connection for our PTaaSapplication. We have found that a high-speed network con-nection is a critical need for proper usability of our interactiveinterface, and to minimize the patient risk due to miscom-munication of exercise activity with the therapist. In addition,the large quantity of data generated by our PTaaS application

7

in the order of severe tens of GB just within a single PTsession requires us to use a robust and high-speed networkinginfrastructure for real-time data import into a remote databasefor synchronous Big Data analytics explained in the previoussubsections. However, we have explored a few methods toreduce the peak data rate and average data needed by our inter-active interface application modes of operation. In our currentapplication implementation, our PTaaS application can be usedin a low resolution mode, and switched a high resolution modeif enough end-to-end session network bandwidth is available.We can also reduce the peak data rate if we turn off thedepth data transmission in the interactive interface application,when there are network bandwidth limitations. We consciouslydo not compress the video or reduce the Frames Per Secondbecause we need to deliver high-quality video to show detailedmovements and facial expressions in the session between theTherapist and Patient. Another reason we do not compress thevideo is our original intent in this study to fully utilize thenew paradigm of Gigabit network connections to residentialhomes through offerings such as Google Fiber, and therebyexplore benefits for immersive PTaaS application delivery.

IV. PERSONALIZED PATIENT-CENTRIC STUDY

In this section, we first describe our human subject se-lection and the testing methodology within actual testbeds.Following this, we present controlled network results fromour synchronous Big Data analytics that shows the networkconfiguration and time synchronization related challenges fordifferent physical therapy activities and interface resolutions.

A. Testing MethodologyWe recruited 5 healthy senior participants from 4 different

senior homes in the Kansas City areas where Google Fiberaccess network connections were available. Participants (ourhuman subjects) were chosen to represent those who wouldbenefit from a smart remote in-home physical therapy sys-tem as part of wellness coaching activities. Fortunately, theMissouri’s regional backbone network operated by MOREnetwas directly peering with Google Fiber, and our MU campuswas connected at 1 - 10 Gbps speeds on different networkinterfaces to MOREnet. Thus, we were uniquely positioned toconduct different experiments with our Gigabit application onactual testbed settings and with real users.

We were assisted by a professional Physical Therapist onthe MU campus in Columbia, Missouri to define the exerciseprotocols that involve upper body and lower body motionfor the PTaaS sessions, and to work with the seniors andour PTaaS interface in our testing activities. Precautions weretaken in the protocol to limit any risks such as physicalor cognitive injury in the human subject interactions withthe PTaaS interface. We strictly followed our campus IRBapproved guidelines in our recruitment and discussions relatedto test plan administration with the 5 human subjects.

The main testing goals were related to evaluation of theusability, user satisfaction and therapy effectiveness on boththe Therapist and the Patient sides using a survey shown inTable III. More specifically, we are interested in studyinghow well the Therapist is able to effectively evaluate thevarious survey categories, and how well the seniors are ableto follow the Therapist instructions when using the high-definition videoconferencing feature of our PTaaS interface.

We considered requirements that are needed for care coordi-nation in response to health issues such as the need for theTherapist to be able to see high-resolution skin color, facialexpressions and fine-level motion trends of the remote senior.

B. Qualitative Usability Results

We conducted two sets of experiments, and asked the Ther-apist and the study participants representing the Patients to fillout the PTaaS user survey forms. In the first experiment set,we were interested in qualitatively studying how our ‘PTaaSinterface with 3D sensing data as well as videoconferencingcapabilities’ would compare to ‘plain 2D video chat or evenjust with a telephone connection’. The Therapist was able tounequivocally conclude that just audio or audio-video 2D inreal-time data is not very helpful for remote physical therapyrelated performance evaluation of a Patient. The 3D depthsensor data using Microsoft Kinect and the user-centered paneldesign in our PTaaS can provide important quantitative datasuch as 3D joint locations and joint angles to calculate gait andother performance parameters accurately. We explicitly remarkthat the accurate parameters information is critically neededto make a remote physical therapy session more informativeabout body dynamics of a Patient and can truly be effective forwellness coaching through better monitoring of Patient status.It also provides the Therapist with a richer ability to offerverbal, auditory and visual cues to perform correct movementswithin exercise forms.

In the second experiment set, we were interested in qualita-tively studying the effectiveness of ‘remote physical therapy’versus ‘in-person physical therapy’ when Patients perform instatic and dynamic balance programs. Both the Therapist andthe Patient concluded that both configurations were equallyeffective. However, given the technology advantages in ob-taining more quantitative information regarding the Patientexercise performance in the PTaaS session, the Therapistconcluded that the remote physical therapy session with arobust high-speed network connection could provide a trulyimmersive user experience. Experiment configurations withregular Virtual Private Network (VPN) software packages ledto unusable sessions, and were rejected for use by both theTherapist and the Patients.

The Therapist also noted that audio quality was a veryimportant feature, and any issues with feedback, backgroundnoises or echo could make the remote physical therapy sessiona failure, even if all of the visual elements had no impairments.The video quality during a remote physical therapy sessionwith PTaaS interface was found to be affected due to thecontrast of background or floor color, Patient or Therapist’sclothing color, room lighting conditions, and network quality.One of the salient conclusions was that the users involved in aremote physical therapy session are advised to wear clothing(especially footwear) that is contrasting to the backgroundor floor color for more optimal user experience and therapyeffectiveness.

In the in-person physical therapy, accurate visual recog-nition of personal movements is harder without objectivemeasurement devices, and thus remote physical therapy withdevices such as the Kinect was more advantageous. Further,the Therapist preferred the remote physical therapy over in-person physical therapy in cases where the Therapist needed

8

Fig. 7: Synchronous Big Data representation

TABLE III: PTaaS User Satisfaction and Interaction Effectiveness Survey Considerations

A. Technical Quality of Video (rate 1/poor to 5/excellent)1. Video artifacts: Look for video artifacts around the remote users head and shoulders (e.g., blocks, image distortions such as tiling, or

out-of-focus areas).2. Sharpness: Observe whether object edge details and fine lines can be distinguished.3. Contrast, brightness, and color saturation: Observe whether the images are too dark or too white with a washed out appearance.4. Color depth: Look for color banding in the backgrounds and on the remote users faces.5. Stability: Evaluate whether images are stable with no motion in the background due to video artifacts, or video jitter noise.6. Background Clarity: Evaluate whether the background is out of focus; whether it is rich in color and texture.

B. Technical Quality of Audio (rate 1/poor to 5/excellent)1. Audio Clarity: Evaluate how clear the audio is, and whether echo noise occurs.2. Audio Stability: Evaluate whether the audio quality is consistent or whether dropout interruptions or other audio degradation occurred,

including audio and lip movement being out of sync.C. User Satisfaction and Effectiveness (rate 1/strongly disagree to 5/strongly agree)

1. I am able to satisfactorily use the video conferencing system with confidence.2. The video conferencing system is readily available and can connect to remote sites of interest.3. The video conferencing system is easy to use.4. I can communicate effectively to complete an agenda with the remote person using video conferencing.

D1. Remote Interaction (rate 1/strongly disagree to 5/strongly agree) *Physical Therapist1. My movements are accurately represented in the interactive coaching interface.2. I am able to see the movements that the remote senior user is performing.3. The remote senior user is able to follow my demonstrated movements.4. I am able to see how my movements compare to the movements of the remote senior user.5. I would be interested in using this system in the future for my clients who need rehabilitation services.

D2. Remote Interaction (rate 1/strongly disagree to 5/strongly agree) *Senior1. My movements are accurately represented in the interactive coaching interface.2. I am able to see what movements the remote physical therapist is demonstrating.3. I am able to perform the movements that the remote physical therapist is demonstrating, to the best of my ability.4. I am able to see how my movements compare to the movements of the remote physical therapist.5. I would be interested in using this system in the future if I have a need for rehabilitation services.

to compare self to remote Patient’s movements. Such a com-parison is impossible in real-time for an in-person physicaltherapy, whereas the side-by-side visual panels in remotephysical therapy provide a better feature in such cases. Moregenerally, the Therapist and the Patients expressed that specificmovement parameters such as range of motion, sway, reactiontime and timing of movements were captured in real-time withour PTaaS interface. This immediate feedback to the Patientand Therapist was critical to modify the Patient movementsvia Therapist demonstration and verbal instructions. Lastly,both the Therapist and the patients showed a high level of

enthusiasm and acceptance of the PTaaS interface despite themany underlying technological issues and changes made to theexperiments as our study progressed.

C. Controlled Network ExperimentsIn this section, we describe our objective methods for

collecting results that will provide an evaluation of the real-time data, computational resources, and network connectivitynecessary for the proper functioning of the PTaaS interface.More specifically, we show how ‘bad network’ and ‘goodnetwork’ scenarios impact the network quality estimation,

9

Fig. 8: General workflow of objective testing experiments

Fig. 9: Users’ average MOS for different application modes,bad network scenario

as well as the PTaaS interface usability and effectivenessdemands.

In our objective testing studies, our design of experimentsinvolved three types of tests with different application modesof high or low resolution RGB video, and with or withoutdepth data in the session as shown in Table IV. Each testingmodes has two parts, one corresponds to an ‘easy activity’(single leg stance) and the other corresponds to a ‘difficulttask’ (tandem walk), and we conducted two tries per each ac-tivity. Moreover, for each experiment, we particularly capturethe exercise activity-specific parameters, start time of activitiesusing a timer, and the reaction time of the Patient in orderto find the time delays associated with the data sent over thetimer socket connection between the Patient and the Therapist.The reaction time in each test is particularly important tounderstand how much time it actually takes for a Patient tostart a specific activity during a PTaaS exercise session afterthe timer starts.

The general workflow diagram of our objective testingexperiments is shown in Figure 8. To synchronize both sides,we utilized the widely-used and conventional Network TimeProtocol (NTP) [23]. However, as we will explain later, NTPdid not provide enough precise time synchronization, espe-cially for good network conditions; its worst-case accuracyis more than 100 milliseconds. Hence, we plan to explore

other technologies such as Precise Time Protocol (PTP) [24]in our future PTaaS interface development iterations. In theinteractions of the Therapist and the Patient using our PTaaSinterface, we changed the application mode, and collectedthe various outputs from the data sources to perform a time-synchronized correlation analysis between the MOS rankingsthat indicate user satisfaction, and the application as wellas network performance data. For a given network qualitycondition observed, we show how we update the weightsof the network metrics by also considering the user opinionscores and categorize the condition as ‘bad’ (red bar in PTaaSinterface) or ‘good’ (green bar in PTaaS interface).

1) Bad Network Scenario Results: In order to study thebad network scenario, we used the faulty network configura-tion described in Table I. We additionally used the VyattavRouter [25] at the software-level for the overlay network(Layer 2) path that further introduce overheads and per-formance bottlenecks for proper functioning of our PTaaSinterface between the Patients in Kansas City and Therapist inColumbia. Our end-to-end available bandwidth in the overlaynetwork channel as reported by Narada Metrics was a sus-tained ≈ 45 Mbps.

Figure 9 shows the users’ average MOS rankings (combinedfor Therapist and Patient) for the different application modes.We can see that the users found the low resolution mode tobe more usable, and the high resolution with depth mode waspractically unusable. The survey notes from the users indicatedthat the higher score for low resolution mode was due to thelower number of frame freeze events and inter-frame jitter,even though the picture quality was inferior in comparisonto the other two high resolution application modes. It isexpected that high resolution with depth mode that uses themost network resources has a larger impact on user experience,and thus suggests that it is not advisable to use high resolutionwith depth modes on medium to low network quality conditionpaths. Another interesting observation from the results inFigure 9 is that - in the cases where the PTaaS users maynot always have high-bandwidth connections such as GoogleFiber, there is an opportunity to consider other user-centereddesign and development approaches for the PTaaS interface tosee how the video panels or sensor data visualizations can beimproved to make the PTaaS session reasonably effective, evenwithout a immersive experience expected on a high-bandwidth

10

TABLE IV: PTaaS Test Plan

Mode-1: Low Resolution Mode-2: High Resolution Mode-3: High Resolution with DepthData Considered: RGB Video 640x480 + Kinect Audio +

Kinect Skeletal DataRGB Video 1280x960 + Kinect Audio+ Kinect Skeletal Data

RGB Video 1280x960 + Kinect Audio+ Kinect Skeletal Data + Depth Data

and low-latency network connection.To study the time synchronization issues in the bad network

scenario, we start with high resolution with depth applicationmode. The main network metrics of throughput and RTT werein the range of average 30 Mbps and 20 mlsec, respectively forthe first and second activities. During the same time periods,the average jitter and loss measurements were 9 mlsec (half theaverage RTT value; shows intermediate network congestion)and 75 %, respectively. For this mode, the ideal baselinethroughput requirements are 100-120 Mbps in average asobtained from our experiments in a controlled network settingwith a 1 Gbps connection. Figures 10 and 11 show throughputand RTT functions for two tries of the first activity. The threevertical dot lines with time labels on these figures are: (i) timewhen the therapist started the exercise activity (i.e., at 00.000sec), (ii) time when patient side interface started timer, and (iii)patient reaction time. We can notice that for both the tries, thenetwork conditions had similar average statistics for most ofthe time, except they were a little bit higher for the throughputat the beginning of the second try. The resultant sway for thefirst activity (i.e., for single leg stance) had the value of 3 inaverage for both the tries as shown in Figure 12.

For this set of experiments, we update weights of highresolution with depth mode w3,B , w3,D, w3,L, w3,J , so that ourfunction will be φ < 0 (poor); this is indicative of low user ex-perience scores for Bav = 30Mbps,D = 20mlsec, L = 75%and J = 9mlsec. We also update the corresponding weightsw1,B , w1,D, w1,L, w1,J of the low resolution mode, so that ourfunction will be φ ≥ 0 (good); this is indicative of the fact thatthe user experience is not relatively affected in this mode forBav = 20Mbps,D = 20mlsec, L = 68% and J = 16mlsec.For this low resolution mode, the ideal baseline throughputrequirements are ≈30 Mbps in average as obtained from ourexperiments in a controlled network setting with a 1Gbps con-nection. Lastly, for the high resolution (without depth) mode,we update the corresponding weights w2,B , w2,D, w2,L, w2,J ,so that our function will be φ = 0 (acceptable); this isindicative of the partly satisfactory user experience scores forBav = 20Mbps,D = 20mlsec, L = 70% and J = 11mlsec.

2) Good Network Scenario Results: In order to study thegood network scenario, we setup a closed-network with avail-able bandwidth of ≈300 Mbps via a VLAN extension (Layer2) between two rooms within our MU Engineering Buildingbetween two different floors. Figure 13 shows the users’average MOS rankings marked separately for the Therapistand the Patients for the three application modes. We cansee that the users found the high resolution (without depth)mode to be the most usable, and the lower resolution modeto be not as much satisfactory. In comparison to the MOSrankings of low resolution mode for the bad network scenario,we see the relative MOS effect, where the users had betterexperience with other modes and were willing to consider thelow resolution mode to be not their desired mode for the PTaaSsession. Interestingly, the high resolution with depth mode isquite similar to the high resolution mode, and the therapist

(a) First try

(b) Second try

Fig. 10: Throughput for first activity in high resolution withdepth mode, bad network scenario

ranking is slightly lower due to the more visual consumptionof the interface information necessary in the exercise activities.Thus, on a good network quality path, it is advisable to useboth the high resolution, as well as the high resolution withdepth modes as they provide a high ranking perceived userexperience.

To study the time synchronization issues in the good net-work scenario, we start with the low resolution applicationmode. The main network metrics of throughput were 45 Mbpsin average for first activity, and 35 Mbps for the second activityas seen in Figure 14. During the same time periods, the RTTvalues were 3 mlsec and 2 mlsec in average for first and secondactivities, respectively as seen in Figure 15. In addition, thejitter and loss measurements were 1 mlsec and 0 % for bothactivities. The resultant sway data of these activities are shownin Figure 16 (note that the patient and therapist data are shownon the same plot). We can observe that the shift between thetwo data in time is almost 0 for the first activity. Hence, we cansay that the Therapist will see the Patient performance withoutany delay, similar to the case of in-person physical therapy

11

(a) First try

(b) Second try

Fig. 11: RTT for first activity in high resolution with depthmode, bad network scenario

user experience. However, the second result looks unintuitive,as it seems that the Therapist observed the Patient movementsbefore the Patient actually performed them. The cause ofthis phenomenon is due to the time synchronization accuracyissues within the NTP protocol, and thus we plan to furtherexplore the opportunity to consider other time synchronizationstrategies for our PTaaS interface design and big data handlingapproaches in future work. Further, from Figure 17 for the highresolution with depth mode sway data, we can observe thatsimilar trends occur due to the time synchronization accuracyissues within the NTP protocol.

For this set of experiments, we update weights of lowresolution mode w1,B , w1,D, w1,L, w1,J , so that our functionwill be φ > 0 (good); this is indicative of the relativelylower user experience scores with equal sway trends on bothsides, which are still in the good performance range forBav = 35Mbps,D = 3mlsec, L = 0% and J = 1mlsec. Wealso update the corresponding weights w3,B , w3,D, w3,L, w3,J

for the high resolution with depth mode, so that our functionwill be φ > 0 (good); this is indicative of the high userexperience scores with equal sway trends on both sides, whichare in the good performance range for Bav = 100Mbps,D =6mlsec, L = 0% and J = 2mlsec. Lastly, for the highresolution (without depth) mode, we update the correspondingweights w2,B , w2,D, w2,L, w2,J , so that our function will beφ > 0 (good); this is indicative of the highest user experience

(a) First try

(b) Second try

Fig. 12: Patient sway for first activity in high resolution withdepth mode, bad network scenario

Fig. 13: Users’ average MOS for different application modes,good network scenario

12

(a) First activity

(b) Second activity

Fig. 14: Throughput for low resolution mode, good networkscenario

scores recorded int his mode with equal sway trends onboth sides, which are in the good performance range forBav = 50Mbps,D = 3mlsec, L = 0% and J = 1mlsec.

V. PTAAS BIG DATA ECOSYSTEM

Our PTaaS interface is one of the major applications withina practical health care environment that is coupled with BigData analytics. Our study findings through actual network andcontrolled network scenarios provided valuable insights on thenetwork, compute and storage resource calibration issues, aswell as the performance troubleshooting issues that involvenetwork effects as well as time synchronization accuracy.Our future work is to explore the architectures for PTaaSthat could be applied to other similar personalized healthcareservices that are remotely delivered at-scale (for hundreds oreven thousands of customers simultaneously) through a cloudcomputing platform and various Big Data analytics tools in areliable, secure and cost-effective manner.

Figure 18 illustrates the various architecture componentsnecessary for at-scale delivery of PTaaS within an actualclinical environment. We envisage two layers of services thatsupport the PTaaS through RESTful APIs that are commonlyused in cloud platforms viz.,: (i) Secure Compute and NetworkServices, and (ii) Data Services. In the first layer, componentsare needed for different processing pipelines that handle theworkflow as well as security issues for data-in-motion (e.g.,

(a) First activity

(b) Second activity

Fig. 15: RTT for low resolution mode, good network scenario

SSL), and the access control as well as encryption to thedata-at-rest. The access control is possible through FederatedIdentity and Access Management frameworks that have webinterfaces for the owners of the PTaaS system to be able toadd, delete or modify permissions of the different data queryand analysis tasks.

A major challenge in the infrastructure configuration forsecure services can be related to the compliance issues (i.e.,pertaining to FISMA Moderate/HIPPA) that are paramount inthe health care related technology infrastructures. In the secondlayer, the components are essentially related to data servicesthat operate in virtualized storage environments that are againfederated across multiple private/public cloud platforms. Last,but definitely not the least, the deep and continuous monitoringwith frameworks such as Narada Metrics as shown in thisstudy, are fundamental to query data from multiple sources,and feed it to tools that perform the Big Data analytics in asynchronous or asynchronous manner as per the health careuse case needs.

VI. CONCLUSION AND FUTURE WORK

In this paper, we presented a 3D real-time interactive systemviz., PhysicalTherapy-as-a-Service (PTaaS) that can be used bya Therapist to remotely monitor Patient performance within anexercise balance assessment program over high-speed networkconnections. Using the features in the PTaaS interface, weshowed how the Therapist was able to monitor Patient status,

13

(a) First activity

(b) Second activity

Fig. 16: Sway for low resolution mode, good network scenario

offer verbal, auditory and visual cues to perform correctexercise movements, all through a synchronous Big Data an-alytics approach we adopted. Our veracity analytics approachcorrelated the high volume, velocity, and variety of data byaddressing challenges relating to network configuration andtime synchronization within wide-area overlay network pathsbetween the Therapist and the Patient. Our results showed thatwe were able to distinguish scenarios where there was bias dueto network quality effects, and thus enabled the Therapist tosuccessfully quantify and assess balance of the senior studyparticipants we recruited in private homes.

We conducted experiments and described results in bothqualitative and objective testing cases, which showed how thePTaaS interface is usable and effective for remote therapyinvolve real-time analysis of specific movement parameterssuch as range of motion, sway, reaction time and timingof movements. The real-time analysis fostered immediatefeedback to the Patient and Therapist, who then used thefeedback to modify the exercise movements via further Ther-apist demonstration and verbal instructions. Both the usersatisfaction and interaction effectiveness information capturedthrough opinion scores showed a high-level of enthusiasm andacceptance by the Patients and Therapist, despite technologicalissues and changes made to the system as the study progressed.

Our future work is to further prove the clinical validityof our PTaaS application by extensively testing and refiningour system with larger groups of Patients and Therapists.

(a) First activity

(b) Second activity

Fig. 17: Sway for high resolution with depth mode, goodnetwork scenario

Fig. 18: Big Data architecture for PTaaS at-scale

14

Our planned future work is to also evaluate our system withgold standard physical therapy assessment systems that useforce-plates and accelerometers. Further, our PTaaS interfaceintegration plans with a Big Data eco-system can decrease thefrequency with which Patients would have to travel to a clinic,while still being able to receive one-on-one interactive carefrom a Therapist. Ultimately, our efforts for remote monitoringwith our PTaaS interface are well suited for transition ofa Patient to total independence with their home exerciseprogram, as part of a therapeutic program or even within healthand wellness programs relating to the area of fall prevention.

REFERENCES

[1] P. Calyam, L. Kumarasamy, C. Lee, F. Ozguner, “Ontology-based Se-mantic Priority Scheduling for Multi-domain Active Measurements”,Springer Journal of Network and Systems Management (JNSM), 2014.(http://www.naradametrics.com)

[2] M. Berman, J. Chase, et. al., “GENI: A Federated Testbed for InnovativeNetwork Experiments”, Elsevier Computer Networks, 2014.

[3] US Ignite Initiative - http://www.us-ignite.org[4] C. Camporesi, M. Kallmann, J. Han, “VR Solutions for Improving

Physical Therapy”, Proc. of IEEE Virtual Reality Conference, 2013.[5] D. Tacconi, R. Tomasi, C. Costa, O. Mayora, “A System for Remote

Orthopedics Rehabilitation”, Proc. of IEEE PervasiveHealth, 2013.[6] J. Fasola, M. Mataric, “A Socially Assistive Robot Exercise Coach for the

Elderly”, Journal of Human-Robot Interaction, Vol. 2, No. 2, pp. 3-32,2013.

[7] S. Obdr?lek, G. Kurillo, E. Seto, R. Bajcsy, “Architecture of an AutomatedCoaching System for Elderly Population”, Proc. of MMVR, 2013.

[8] Y. Chen, Y. Hung, “Using Real-time Acceleration Data for ExerciseMovement Training with a Decision Tree Approach”, Proc., IEEE Con-ference on Machine Learning and Cybernetics, 2009.

[9] K. Huang, P. Sparto, S. Kiesler, D. Siewiorek, A. Smailagic, “iPodfor Home Balance Rehabilitation Exercise Monitoring”, Proc., IEEESymposium on Wearable Computers, 2012.

[10] L. Mize, R. Klenke, J. McCollum, “Snapshot Capture from Live High-definition Video Stream for Transmission over Low-bandwidth DataLink”, Proc. of IEEE SoutheastCon, 2010.

[11] L. Tan, S. Lau, T. Chong, “Enhanced Compression Scheme for High-latency Networks to Improve Quality of Service of Real-time Applica-tions”, Proc. of Asia-Pacific Symposium on Information and Telecommu-nication Technologies, 2010.

[12] M. Shah, K. Jinman, M. Khadra, D. Feng, “Enhancing Home AreaNetworks to Facilitate Telehealth Services: Test-bed Scenario with Video-consultation-calls (VCC)”, Proc. of IEEE International Conference onOrange Technologies, 2014.

[13] P. Calyam, A. Kalash, R. Gopalan, S. Gopalan, A. Krishnamurthy,“RICE: A Reliable and Efficient Remote Instrumentation CollaborationEnvironment”, Journal of Advances in Multimedia, 2008.

[14] J. Heer, S. Kandel, “Interactive Analysis of Big Data”, ACM XRDS:Crossroads Magazine, 2012.

[15] G. Jung, N. Gnanasambandam, T. Mukherjee, “Synchronous ParallelProcessing of Big-data Analytics Services to Optimize Performance inFederated Clouds”, Proc. of IEEE Cloud, 2012.

[16] A. Jacobs, “The Pathologies of Big Data”, Communications of the ACM,Vol. 52, No. 8, pp. 36-44, 2009.

[17] P. Gui, Y. Qin, C. Hongmin, Z. Tinghui, Y. Chun, “Accurately calibratekinect sensor using indoor control field”, IEEE Workshop on EarthObservation and Remote Sensing Applications, 2014.

[18] J. Chow, D. Lichti, “Photogrammetric Bundle Adjustment With Self-Calibration of the PrimeSense 3D Camera Technology: MicrosoftKinect”, Proc. of IEEE Access, 2013.

[19] A. Mishra, M. Skubic, C. Abbott, “Development and PreliminaryValidation of an Interactive Remote Physical Therapy System”, Proc. ofIEEE Engineering in Medicine and Biology Society, 2015.

[20] E.Stone, M. Skubic, “Fall Detection in Homes of Older Adults Using theMicrosoft Kinect”, IEEE Journal of Biomedical and Health Informatics,Vol. 19, No. 1, pp. 290-301, 2015.

[21] E. Stone, M. Skubic, “Unobtrusive, Continuous, In-Home Gait Mea-surement Using the Microsoft Kinect”, IEEE Transactions on BiomedicalEngineering, Vol. 60, No. 10, pp. 2925-2932, 2013.

[22] M. Skubic, R. Guevara, M. Rantz, “Automated Health Alerts Using In-Home Sensor Data for Embedded Health Assessment”, IEEE Journal ofTranslational Engineering in Health and Medicine, Vol. 3, pp. 1-11, 2015.

[23] D. Mills, “Network Time Protocol (NTP), IEEE Trans. on Communica-tions, Vol. 39, No. 10, pp. 1482-1493, 1991.

[24] A. Vallat, D. Schneuwly, “Clock Synchronization in Telecommunica-tions via PTP (IEEE 1588)”, Proc. of IEEE Frequency Control Sympo-sium, 2007.

[25] B. Gillian, “VYATTA: Linux IP Routers”, 2007.(http://freedomhec.pbworks.com/f/linux ip routers.pdf)