healthcare application using cloud platform

230 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 1, JANUARY 2013

Toward Ubiquitous Healthcare Services With a NovelEfficient Cloud Platform

Chenguang He, Xiaomao Fan, and Ye Li∗, Member, IEEE

Abstract—Ubiquitous healthcare services are becoming moreand more popular, especially under the urgent demand of the globalaging issue. Cloud computing owns the pervasive and on-demandservice-oriented natures, which can fit the characteristics of health-care services very well. However, the abilities in dealing with mul-timodal, heterogeneous, and nonstationary physiological signals toprovide persistent personalized services, meanwhile keeping highconcurrent online analysis for public, are challenges to the generalcloud. In this paper, we proposed a private cloud platform archi-tecture which includes six layers according to the specific require-ments. This platform utilizes message queue as a cloud engine, andeach layer thereby achieves relative independence by this looselycoupled means of communications with publish/subscribe mech-anism. Furthermore, a plug-in algorithm framework is also pre-sented, and massive semistructure or unstructured medical dataare accessed adaptively by this cloud architecture. As the testingresults showing, this proposed cloud platform, with robust, stable,and efficient features, can satisfy high concurrent requests fromubiquitous healthcare services.

Index Terms—Cloud computing, healthcare services, plug-inalgorithm, ubiquitous.

I. INTRODUCTION

ACCORDING to the World Health Organization report, theincidence of the suboptimal health status (SHS), which

is also called “the third state” (between health and disease),accounts for 75% of the world population [1]. Only in China,there are estimated 900 million people in such status [1]. Someof them would pay close attention to their health hoping to getpreventive health examination. Moreover, with the aging popu-lation growing, we also have to take measures to launch chronicdisease surveillance for elder people. Therefore, to satisfy thedemands of “SHS” groups and aging population, the healthcaresystem is proposed and becoming more and more popular, which

Manuscript received March 16, 2012; revised July 5, 2012; accepted August29, 2012. Date of publication October 5, 2012; date of current version Decem-ber 14, 2012. This work was supported in part by the Shenzhen Basic ResearchFunds for Distinguished Young Scientists under Grant JC201005270306A,in part by the Comprehensive Strategic Cooperation Projects for GuangdongProvince and Chinese Academy of Sciences under Grant 2011A090100025,and in part by the National Basic Research Program of China under Grant2010CB732600. Asterisk indicates corresponding author.

C. He is with the Key Laboratory for Health Informatics of Chinese Academyof Sciences, and with Shenzhen Institutes of Advanced Technology, ChineseAcademy of Sciences, Shenzhen 518055, China. He is also with the Uni-versity of Chinese Academy of Sciences, Beijing 100049, China (e-mail:cg.he@siat.ac.cn).

X. Fan is with the Key Laboratory for Health Informatics of Chinese Academyof Sciences, and with Shenzhen Institutes of Advanced Technology, ChineseAcademy of Sciences, Shenzhen 518055, China (e-mail: xm.fan@siat.ac.cn).

∗Y. Li is with the Key Laboratory for Health Informatics of Chinese Academyof Sciences, and with Shenzhen Institutes of Advanced Technology, ChineseAcademy of Sciences, Shenzhen 518055, China (e-mail: ye.li@siat.ac.cn).

Digital Object Identifier 10.1109/TBME.2012.2222404

can provide self-monitoring of healthy status, early warning ofdisease, and even an instant report of the physiological signalanalysis for individuals.

Until now, such health monitoring systems have been widelyimplemented with traditional management information system(MIS) mode, which could no longer meet all the requirementsof the healthcare system. For instance, Leel et al. [2] proposea service-oriented architecture (SOA)-based system to integratethe data from a personal health device and customize servicesfor users, but somewhat lack concurrency and persistent servicecapabilities; Kulkarni and Ozturk [3] developed a system namedmPHASiS, which claims an end-to-end solution. However, it re-lies on the particular programming language, Java, which willresult in less efficient of system due to the performance of Javavirtual machine (JVM). Nowadays, as an emerging state-of-the-art technology, cloud computing can provide pervasive, eco-nomical, and on-demand services for customers. Many piecesof research on medical information have been involved into thecloud [4]–[6] recently. It can fulfill the flexibility and scalabilityof service-orientated systems, just like the healthcare systeminfrastructure should have, which is very different from thetransaction-oriented MIS mode.

Cloud computing inherited the features of high-performanceparallel computing, distributed computing, and grid comput-ing, as well as the further development of these techniques toachieve location transparency to enhance the user experiencesover the Internet. More and more applications are migrated ontothe cloud platform, which can provide services for various ap-plication scenarios [7], [8]. Such a platform generally comprisesa number of diverse groups of components, no matter they areopen source or not. Consequently, how to design a model withoptional components to meet the needs of flexibility and scal-ability is one of the critical issue when constructing a cloudplatform. In the case of the solution for this healthcare system,we presented a six-layer deployment architecture and proposeda message queue (MQ) as cloud engine as well as fundamen-tal cloud storage based on the open-source components, whichcan be obtained freely and timely from Internet communities.Also, a plug-in algorithm framework was developed with pub-lish/subscribe mechanism to implement this healthcare systemfor the community people. The expandable feature of the pub-lish/subscribe communication model can meet the needs of ahealth cloud platform in high concurrent performance. Detailsabout the implementation are depicted in Section III after theanalysis of the practical requirements of healthcare in Section II.The results of performance test of the system are given inSection IV, and the Section V draws a conclusion and foreseesthe future work.

0018-9294/$31.00 © 2012 IEEE

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 1, JANUARY 2013 231

II. SYSTEM REQUIREMENTS ANALYSIS

A. Application Scenarios

This system is planned to serve hundreds of families residingin the city for health monitoring via Internet, and will sup-port thousands of potential customers including physicians andwhoever care about health of themselves in the near future.So, higher concurrent transactions and shorter response timeare the requirements to our system. Certainly, visualization ofthe results is also compute-intensive tasks usually with large-scale graphic drawing, which need to be stored and accessed inan efficient way to provide instant services to individuals. Themultimodal and nonstationary characteristics of special med-ical signals, such as electrocardiogram (ECG), photoplethys-mograph (PPG), and electroencephalograph (EEG), are verytrivial for storage and may lead to a synchronous problem onGlusterFS [9]. Hence, these large numbers of semistructuredor unstructured health records, as well as companied with triv-ial files, are all adapted to NoSQL [10] databases instead oftraditional relational ones. These fundamental characteristicsare very different from the features of grid computing whichaimed at special applications and difficult to operate for unpro-fessional users (e.g., scientific exploration projects). Further-more, the method of service delivery is another important factorwhich affects the usability of the system. In addition to a mobilephone as an appropriate front-end device [11], TV set is anotherfriendly interface for aging people. So, with the assistance ofset-top box, services would be delivered through a TV cableubiquitously.

B. Overview of the Data Flow

There are two main activities of users: upload the raw dataand browse the results of diagnosis. Upload activities are di-vided into two steps: the first is to transmit physiological data,which are collected by sensors, to the gateway through short-distance transmission by a wired USB or a wireless Bluetooth.The second is to deliver the data from the gateway to the cloudservers through IP networks. Set-top box (for USB) and mo-bile phone (for Bluetooth) play the important role of gatewayto relay data, respectively. We have developed three types ofinterfaces to users including TV cable, Wi-Fi/Ethernet, and cel-lular networks (2G/3G). All of these forms are interacted withcloud servers via Internet at the end. Browsing the results ona TV or a mobile phone is very straightforward. As aforemen-tioned, all of the physiological data should be collected in adistributed database cluster as the reservoir for data mining.Although the data may possibly be lost by accident (e.g., powerfailure) during operation, it must be restored from the databasesin cloud. Therefore, to avoid a single point of failure, we utilizea primary/secondary scheme provided by a database to maintainenough replicas of data. Keep-alive mechanism is also adoptedin this scheme and switch back and forth if the heart-beatsignal is missing between primary ones and secondary ones.Fig. 1 shows the overview of our ubiquitous healthcare servicesystem.

Fig. 1. System overview of ubiquitous healthcare service.

III. SYSTEM ARCHITECTURE AND IMPLEMENTATION

Considering the application requirements and the data of theinconsistent format, we improved our previous work [12] andproposed a hybrid model which simplified the one derived fromthe National Institute of Standards and Technology (NIST) [13]and Jericho models [14]. Being everything can be treated as aservice (XaaS) [15], we can also choose component as a ser-vice to take very efficient way to develop this specific privatecloud architecture. The philosophy of this private healthcarecloud inherited the software as a service (SaaS) of the NISTmodel and introduced the insource/outsource concept into ourcloud platform development. Outsource products include open-source components and commercial software, whereas the coredata mining models are developed by ourselves as an insourceproduct. The whole cloud platform comprised of several build-ing blocks according to functionality requirements. Fig. 2 showsthe six-layer architecture of the healthcare cloud platform whileeach layer’s content and function are interpreted as follows.

1) Service interaction is a top layer where users can inter-operate with the terminals such as 3G mobile phone andset-top box, and browser on a computer, etc., to collectand upload original physiological data, as well as down-load the analysis results from the cloud servers.

2) Service presentation can be regarded as an interface withvarious kinds of services such as wireless application pro-tocol (WAP), Web, image provider, etc. In addition, thenode balancer also works on this layer, which can providehigh concurrency capability. Some run-time informationis also reserved at this level.

3) Session cache stores the user sessions on one hand, whichmaintains the authentication and certification status at thefirst time when the user accesses the services, i.e., serviceinteraction through the presentation layer. On the otherhand, memory cache is adopted to expedite the result dataaccess. In fact, this session’s architecture is a big hashedkey/value table.

4) Cloud engine is a dispatcher to cooperate the compo-nents with each other to make the cloud running by us-ing message-driven mode via MQ. The functionalitiesof the management for the queue are provided by this

232 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 1, JANUARY 2013

Fig. 2. Six layers of a healthcare cloud platform.

layer, which is regarded as a critical core for schedulingtasks.

5) Medical data mining is clusters of algorithms, includingdata preprocessing, data analysis, mining algorithms, andvisualization processing. This layer can handle the datatransmitted from the front end and generate the resultsback to database. Other algorithms can also be easilyplugged in if needed.

6) Cloud storage provides data resources for the entirehealth cloud platform, including user information, vitalsigns, health records, and graphic data for processing.Physiological data collected from body area networks andmassive graphic data for distributed processing of suchdata-intensive tasks are the primary contents. Using thedistributed file storage system, like Hadoop DistributedFile System [16] with semistructural database, canprovide more extensible and efficient access. Moreover,sharding [10] is another major characteristic of this dis-tributed database to gain higher availability. Redundancyamong these pieces of shards and different views of thesame data provide the consistency to a large extent. Thismechanism can guarantee the integration of global dataand transparency to users.

There are two key components addressed here: the first one isMQ, which is a kind of component that provides asynchronouscommunication protocol to achieve the independence betweenmessage sender and receiver. Queue is utilized for storing the

Fig. 3. Plug-in algorithm framework.

event message generated by a sender (publisher), and all of thelisteners (subscriber) who are interested in this kind of event canfetch this message to process a predefined routine. All commu-nication parties should follow the same protocol (e.g., advancedmessage queue protocol), and utilize available MQ applicationprogram interface [17] to operate the queue. It is not neces-sary that two parties must know the existence of each other.After a message is inserted into a queue, it will be preservedand not deleted in the MQ until the corresponding subscriberreconnects to the system. Messages can be exchanged on thelevel of process, application, or even intercloud. Queue residesin cloud just as an engine to drive the system. Addition, differentmessage head identifies the analysis algorithm and indicates thesubsequent behaviors of cloud.

The second one is the plug-in algorithm framework devel-oped with respect to the extendibility of various services, whichis based on the publish/subscribe mechanism to provide cus-tomized functions conveniently, not only for healthcare butalso for other services. The whole system can reduce the mod-ule coupling by adopting this algorithm. For instance, variousdata mining algorithms, such as analysis of peripheral vascularfunction, instantaneous heart rate, and chaotic characteristicsof power spectral density, etc., should be adopted to performautomatically according to different analyzed signals. Everydifferent function can subscribe the different theme of mes-sage, which is classified by a message head. Accordingly, wedesigned an abstract core class named CoreStubClass, includ-ing a private attribute analysisKind, and an abstract methodhandleRequest(int). This class communicates with Message viaMQ, and the other kinds of concrete implementation classesextended the CoreInterface, e.g., SignalFilterCore, ECGAnaly-sisCore, and other data mining classes. Fig. 3 shows the classframework of the plug-in algorithm.

IV. TEST RESULTS AND ANALYSIS

As a result, Fig. 4 shows the upload interface on a TV whilecollecting user’s psychological signals for 2 min. Results of

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 1, JANUARY 2013 233

Fig. 4. Presentation of a user’s upload interface on a TV.

Fig. 5. Presentation of a user’s abnormal ECG on a browser.

TABLE IOVERALL STATISTICS OF UPLOADING AND BROWSING ACTIVITIES

analysis are viewed by a Web browser as shown in Fig. 5, whichcan also been shown on a TV. To perform stress and load tests onour system, we designed two test cases to verify the capabilitiesof cloud platform: simulating initial 100 users for uploadingactivity lasting for 10 min, as well as 300 users for browsinglasting for 30 min. Table I shows the overall of the testing statis-tics. Use four servers with the same configuration (2.93-GHzIntel Xeon CPU with 4G RAM) as front-end clusters to providevisualization services under the Tsung [18] testing environment.Tsung can generate many sessions and make concurrent requeststo a server automatically. It starts with the initial user count andreports the performance in detail when testing is over. A new

TABLE IISTATISTICS OF NETWORK TRAFFIC AND PERFORMANCE

TABLE IIISTATISTICS FOR READING PERFORMANCE OF AMAZON S3 [19]

HTTP request generated within a given interval of 0.02 s repre-sents that a new user connection happened, and a session willbe created according to the probability presented in the testingconfiguration file. The mean response time and count (for page,request, etc.) for the whole test are computed, which generated16 3011 and 57 3683 concurrent requests for the two activities,respectively, as shown in Mean Time and Count columns. Re-sponse time less than 1s is tolerable for user experience. Theactual number of concurrent users is nearly 30 000 during theexperiment as the Users_count column of Table II shows. It isacceptable to us.

By comparison, Table III shows the reading performance ofAmazon’s simple storage service (S3) with different page sizes(writing data to S3 takes about three times as long as readingdata) [19]. Ignoring the slight differences of hardware, the us-ability of our system is satisfied in terms of Bandwidth, which isat least double of that in S3 (compare the last row of Table II withthat of Table III). Response Time is also better obviously. Over-all, it demonstrates that this document-oriented cloud storagearchitecture is more appropriate for the environment of health-care services with a large number of trivial files, rather thannondocument-oriented ones such as S3. Furthermore, due toplatform capabilities of linear extendibility, we simply increasethe number of servers if there are more potential users’ requestsin the future.

V. CONCLUSIONS AND FUTURE WORK

In this paper, we propose a six-layer architecture of the pri-vate cloud platform associated with the technologies such asMQ, load balance, plug-in algorithm, and cloud storage. Thisplatform can manage the semistructure, unstructured, and het-erogeneous physiological signal data very well. Testing resultsshow that it can cope with the issues about high concurrent re-quests in ubiquitous healthcare service and dispose of tasks formassive physiological signals analysis rapidly, as well as owningrobust, instant, and efficient features. Particularly, it demands notmuch for hardware, and most of those integrated componentssuch as NoSQL databases and MQ are open-sourced and free,whereas the commercial portfolio of software is always very

234 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 1, JANUARY 2013

expensive. So, it is also an appropriate low-cost solution withrespect to the effectiveness. In the future, we will further explorethe security issues and add novel trustworthy module to protectuser’s privacy.

REFERENCES

[1] H. Ding, J. He, and W. Wang, “The sub-health evaluation based on themodern diagnostic technique of traditional chinese medicine,” in Proc.Workshop Edu. Technol. Comput. Sci., Shanghai, China, Mar. 7–8, 2009,pp. 269–273.

[2] S.-H. Leel, J. H. Song, J.-H. Ye, H. J. Lee, B.-K. Yi, and I. K. Kim,“SOA-based integrated pervasive personal health management system us-ing PHDs,” in Proc. 4th Int. Conf. Pervasive Comput. Technol. Healthcare,Munich, Germany, Mar. 2010, pp. 1–4.

[3] P. Kulkarni and Y. Ozturk, “mPHASiS—Mobile patient healthcare andsensor information system,” J. Netw. Comput. Appl., vol. 34, pp. 402–417, 2011.

[4] N. Botts, B. Thoms, A. Noamani, and T. A. Horan, “Cloud computing ar-chitectures for the underserved: Public health cyberinfrastructures througha network of HealthATMs,” in Proc. 43rd Syst. Sci. Conf., HI, Jan. 2010,pp. 1–10.

[5] Z. Rashid, U. Farooq, J.-K. Jang, and S.-H. Park, “Cloud computing awareubiquitous health care system,” in Proc. 3rd E-Health Bioeng. Conf., Iasi,Romania, Nov. 2011, pp. 1–4.

[6] C. O. Rolim, F. L. Koch, C. B. Westphall, J. Werner, A. Fracalossi, andG. S. Salvador, “A cloud computing solution for patient’s data collectionin health care institutions,” in Proc. 2nd eHealth, Telemed., Social Med.Conf., St. Maarten, The Netherlands, Feb. 2010, pp. 95–99.

[7] J.-T. Hsu, S-H. Hsieh, P-H. Cheng, S-J. Chen, and F-P. Lai, “Ubiqui-tous mobile personal health system based on cloud computing,” in Proc.IEEE Region 10 Conf. TENCON, Bali, Indonesia, Nov. 2011, pp. 1387–1390.

[8] B. Jit, J. Maniyeri, K. Gopalakrishnan, S. Louis, P. J. Eugene, H. N. Palit,Y. S. Foo, L. S. Lau, and X. Li, “Processing of wearable sensor data onthe cloud—A step towards scaling of continuous monitoring of health andwell-being,” in Proc 32nd Conf. Eng. Med. Biol. Soc. (EMBS), BuenosAires, Argentina, Aug./Sep. 2010, pp. 3860–3863.

[9] GlusterFS. (Mar. 2012). [Online]. Available: http://www.gluster.org/about

[10] M. Stonebraker, “SQL databases v. NoSQL databases,” Commun. ACM,vol. 53, no. 4, pp. 10–11, Apr. 2010.

[11] C. G. Scully, J. Lee, J. Meyer, A. M. Gorbach, D. Granquist-Fraser,Y. Mendelson, and K. H. Chon, “Physiological parameter monitoringfrom optical recordings with a mobile phone,” IEEE Trans. Biomed. Eng.,vol. 59, no. 2, pp. 303–306, Feb. 2010.

[12] F. Miao, X. L. Miao, W. H Shangguan, and Y. Li, “MobiHealthcare sys-tem: Body sensor network based m-health system for healthcare applica-tion,” E-Health Telecommun. Syst. and Netw., vol. 1, no. 1, pp. 12–18,2012.

[13] A. Samba, “Logical data models for cloud computing architectures,” IEEEIT Prof., vol. 14, no. 1, pp. 19–26, Jan./Feb. 2012.

[14] Jericho Forum, “Cloud cube model: Selecting cloud formations forsecure collaboration version 1.0,” Jericho Forum Specification, Apr.2009.

[15] H. E. Schaffer, “X as a service, cloud computing, and the need forgood judgment,” IEEE IT Prof., vol. 11, no. 5, pp. 4–5, Sep./Oct.2009.

[16] T. White, Hadoop: The Definitive Guide. Sebastopol, CA: O’ReillyMedia, 2009, ch. 3, pp. 41–73.

[17] MQ API Guide. (Mar. 2012). [Online]. Available: http://www.rabbitmq.com/api-guide.html

[18] Tsung. (Apr. 2012). [Online]. Available: http://tsung.erlang-projects.org/user_manual.html

[19] M. Brantner, D. Florescu, D. Graf, D. Kossmann, and T. Kraska, “Buildinga database on S3,” in Proc. ACM SIGMOD Int. Conf. Manag. Data,Vancouver, BC, Canada, Jun. 9–12, 2008, pp. 251–263.