Context-aware Cardiac Monitoring for Early Detection of Heart Diseases

Abdur Forkan1,2, Ibrahim Khalil1,2, Zahir Tari1

1 School of Computer Science and IT, RMIT University, Melbourne, Victoria, Australia2 National ICT Australia (NICTA), Victoria Research Lab, Melbourne, Victoria, Australia

Abstract

The aim of this paper is to propose a scalable context-aware framework for early detection of several cardiovas-cular diseases by continuous monitoring using smart sen-sors and utilizing the strength of cloud computing. By con-stant sampling of ECG signal, vital signs, and activitiesour system detects possible symptoms of heart disease andalerts user by delivering context-aware service using flex-ible output modalities. A non-context-aware system thatmakes a decision based only on abnormal ECG signal cangenerate false alerts at high rate. Our proposed solutionaims to reduce that rate by bringing different contexts indecision making process. As a proof of concept, we devel-oped a simulated prototype to detect long term health riskof Premature Atrial Contraction (PAC), a common formof cardiac arrhythmia. The system can classify ECG sig-nals as PAC using appropriate feature selection and learn-ing algorithm. By tracking the stored context history andpersonal profile in the cloud database, our system detectssmoking habit, alcohol consumption, caffeine intake of theuser. It can also detect activities like stress, hypertension,and anxiety using different physiological parameters of theuser and capable of sending situational warning notifica-tions. Thus, this model can be a new mechanism for heartdisease detection.

1. Introduction

Cardiovascular disease (CVD) is the most commondisease in Australia especially for elderly people whichcaused 50,000 mortalities in 2008. While the treatment fa-cilities for cardiac arrhythmias [1] are inadequate in mostof the underdeveloped and developing countries, evenwestern countries do not have proper technology to detectdiseases at early stages. Aged people living alone in homedie from CVD more than other members of the population.The death rate can be minimized if the causes of any po-tential CVD are automatically detected before the patientis in high risk situation.

In recent years various cardiac monitoring systems havebeen developed for detecting abnormalities in ECG [2, 3].

Most of the systems rely on stand-alone single processingserver and analyze data locally. When abnormality is de-tected in ECG signals the system sends emergency alert tothe caregiver for immediate help. So the decision is madeupon only heart conditions of the user. There is no systemthat stores context as medical history and predicts symp-toms by looking at the present and the past contexts andactivities. But many of the cardiac arrhythmias which looknormal from an instant EGC observation can create severeheart diseases due to some long term harmful habits andmental illness. Our context-aware solution targets to detectthe occurrence of such situations. Upon discovery of suchthreatening patterns in context the system suggest patientsto control those influencing factors to correct the arrhyth-mias and avoid severe danger like heart failure.

We picked the example of Premature Atrial Contraction(PAC), which generally occur in a normal heart. PAC canbe triggered by alcohol, nicotine, caffeine, anxiety, fatigue,fever, and infectious diseases [1, 4]. In fact, PAC com-monly cause no symptoms and can go unrecognized foryears. PAC can be dangerous as an early sign of heart fail-ure or an electrolyte imbalance. After occurrence of PAC,the doctor needs to ask patient about the habits and med-ical history for diagnosis. In addition, medical tests arerequired to identify the symptoms properly. Continuouscontext monitoring can eliminate these manual steps andsimplify the jobs of health-care professionals. In summary,the advantages of having such system are:

• Many of the elderly patients who are not accompaniedby family members can have certain habits or behaviourswhich may cause serious heart diseases later without theirawareness. Our system will help those lonely elderlypeople by detecting potential cardiac syndromes at initialstage.• Traditional systems generally lack context-awareness intheir design and suffer from higher false alarm rate. Be-cause, often the assumptions based only on ECG abnor-mality do not hold true. Our proposed system uses context-aware approach for more robust estimation of actual inci-dents and thus reduces diagnostic delay as well as falsealarm rate.• The system will minimize the extra burden on health-

ISSN 2325-8861 Computing in Cardiology 2013; 40:277-280.277

Camera

TemparatureSensor

Monitor

Smart Phone

ECG Senosr

SpO2

BSN

LightSensor

Audio Detector

A person under Ambient Assisted Living (AAL)

BloodpressureSensor

RF Tag

RFID reader

Smoke Detector

Speaker

Local Server

Profile data(Disease,Medications,Medical data etc.)

Personal cloud server

ContextAggreagatordata

data

data

Context Providers

ECG dataclassifier

Activity Recognizer

LocationTracker

Context

Context Management System (CMS)

Contextand Activity History

Context Model

Servicemapping

Rules

PVC symptomdetection

PAC symptom detection

Service Providers

Heart attack symptom detection

Doctor

Emergency

Monitoring

ServiceRulesService

Context-aware service action ( e.g. warning, reminder, notification ) Emergency Alert

Reminder

Temp Sensor

Care-givers

AccelerometerMic

Figure 1. The Generic system architecture of the proposed cloud-based context-aware model

care professionals. They will just define the rules for thesymptoms in terms of causes and consequences. Theydon’t need to follow any manual process. The system willautomatically detect those causes from context knowledgeand will send immediate alerts to the doctors and care-givers.

The rest of this paper is organized as follows: section2 summaries the architecture of the proposed model. Thefunctional components are presented in Section 3. Sec-tion 4 briefly describes the prototype that has been builtfor evaluating the architecture. Finally, section 5 concludesthe work.

2. Architecture

Our proposed cloud-based context-aware model is de-picted in Figure 1. The Ambient Assisted Living (AAL)system consists of a target user, body worn sensors (ECG,SpO2, Blood Pressure sensor, accelerometers, etc.), envi-ronmental sensors (light, temperature, smoke, etc.), mon-itoring and communication devices ( RFIDs, cameras,speakers, microphone, monitor, etc.), smart phone, and lo-cal processing server.

ECG data, sensor data and application data generated inAAL system, and profile data on personal cloud storageare forwarded to Context Aggregator (CA) cloud by lo-cal server. Context providers (CPs) collect data from CA,classify raw data to high level context using classificationalgorithms and send back to CA as context. CA combinesall the contexts together in a model and delivers it to theContext Management System (CMS).

CMS stores context history in the cloud repository. Itcollects service rules from Service Providers (SPs) andlooks for anomalous events by rule matching between con-texts and services. The service rules are defined by expertssuch as doctors and emergency service providers. TheCMS is properly trained to classify a rule-based situationcorrectly.

When an anomalous situation is detected the CMS sendsappropriate context-aware actions to the user, his/her doc-tors and, emergency service providers according to the

rules defined for the situation.

3. System functionalities

The primary objective of the system is to recognizeprobable heart disease from long term observations andprovide runtime context-aware actions to help the targetuser. The amount of information that can be categorized ascontext is large. The contexts that are captured from a sin-gle AAL system can be massive. The smart device is notmuch resourceful to process and store this huge amount ofcontext. So we selected scalable cloud model for our sys-tem. The use of cloud computing escalate the capabilityof handling versatile contexts. From literature review ofexisting solutions we assumed the capability of detectingsome contexts and activities in our model. As our systemkeeps track of complex situational context so it requires tosustain following functionalities:

3.1. ECG monitoring

ECG waveforms help to determine many cardiac ar-rhythmias [5]. An ECG waveform has three basic waves:the P, QRS, and T. When the values of intervals do not fallwithin the expected ranges, different cardiac abnormali-ties can be detected. As instance, PAC is the change inheart’s normal rhythm sequence caused by an early extrabeat originating from the atria. After PAC occurs sinusrhythm usually resumes. The hallmark ECG characteristicof a PAC is a premature P wave with an abnormal config-uration when compared with a sinus P wave. The baselinerhythm is irregular here. Therefore, using these basic fea-tures an abnormal ECG is detected.

3.2. Vital signs monitoring

Using wearable sensors some vital signs (Figure 2) aremeasured continuously. These physiological parametersare important measurements to pick up some contexts.Stress [6], anxiety, fever and some other mental and phys-ical illness can be recognized using the observed values of

278

DrinkingWine

DrinkingCoffee

Smoking Taking Medicine

Premature Atrail Contractions

Abnormal ECG Occurance

Nicotine / Alcohol / Caffeine Cosumption

Vital Signs

BPResp.SpO2Pulse Temp

Sleep Food and DrinkWake Up Excercise

Daily Activities

Decision

Normal

Warning

Emergency

Moderate

Hypertension Stress Anexity

Low RespirationHigh BPFever

Habits and activitity history

Improper Sleep WeaknessLess ExcerciseFood Cosumption

Fusion of contexts from multiple sources

Figure 2. Decision fusion from multiple observations

these vital signs. This kind of illness has direct impacts onsome heart diseases.

3.3. Activity monitoring

Using wearable accelerometer, RFIDs and cameras incombination with ambient sensors and devices (Figure 1)it is possible to identify the daily activities of the patient.As example, using camera and smoke detector the smok-ing activity can be identified. Using user’s e-prescriptionon personal cloud database and RFID it is easy to detectthe medication [7] that user has taken. Moreover, some ofthe activities can be detected using sequence of primitiveactivities. Such as, making coffee [8] and drinking it af-terwards can be treated as drinking coffee activity (Figure2).

3.4. Activity logging

To obtain previous context information, each of the ac-tivity and related context is stored on cloud repository us-ing time-stamp value. The history log simplifies the taskof determining context like sleep and wake up time pattern,average exercise duration etc. (Figure 2). These long termfeatures are the reasons of many diseases. So, such activitylogs are useful for anticipating certain symptoms.

3.5. Symptom detection

As mentioned earlier the rules of the symptoms are de-fined by the experts and stored in service provider’s cloudrepository. The medical knowledge is necessary to de-tect any suspicious condition. The rules are also variedamong individuals. When an abnormal ECG is detectedand classified the CMS picks the rules associated with thatarrhythmia. If the context matched with a rule then CMS

considered this as abnormal situation and take appropriateactions to help the user to overcome the problem. Some ofthe rules are presented Figure 3.

Figure 3. Some rules for detecting abnormalities

3.6. Context-aware service action

When any anomaly is detected in the ECG signal aswell as in context information, the CMS selects specificservices based on the importance of the situation. As inFigure 3, if PAC is detected for a patient who does nothave heart disease but have habit of drinking too much cof-fee the context-aware system can send a warning messageto the patient. But, if PAC is detected for a patient hav-ing heart disease and with other abnormal vital signs (e.g.High BP) then this is a severe case and immediate emer-gency alert is send to the doctor.

4. Prototype and results

We have built a small prototype in java which is capableof making context-aware decisions when any anomaly incontext is detected. We picked 50 records from MIT-BTHarrhythmia database. Each of the ECG signal is classifiedbased on average length of PR interval, QRS complex, QTinterval and RR intervals. So, using these features a PACsignal can be detected [9]. A screen-shot of PAC detectionprototype is shown in Figure 4.

We have generated 30 different rules like in Figure 3.Then we have created different patient profiles and ran-domly generated some contexts for each of the patient. Theprofile and the rules are stored in remote database. After-wards, different ECG signal is attached with the profile ofthe patient. When the total value set matched with a rule,corresponding context-aware action was reported. Some ofthe observed results are presented in Figure 5. There wasno false negative situation which could threaten the condi-tion of the patient being monitored. A non-context-awaresystem would not able to detect such situations.

279

Figure 4. The screen-shot of java prototype to distinguishbetween normal sinus rhythm and PAC

Figure 5. Experimental observations

5. Conclusion

In this paper we have presented a framework that canprovide remote monitoring facilities for heart disease de-tection using context. The system classifies the heart con-dition of the patient not only based on ECG analysis butalso taking into account of different contexts and activi-ties. The proposed system is expected to be useful to detectmany cardiac arrhythmias precisely. A small prototype isdescribed as proof of concept for the model. In the future,we plan to develop more modules to support the detectionof other arrhythmias which vastly depends on user’s con-text for diagnosis.

Acknowledgments

The authors wish to acknowledge the support of RMITUniversity and NICTA for funding the research work pre-sented in this paper.

References

[1] Wilkins LW. ECG Interpretation Made Incredibly Easy.Springhouse Publishing Company, 2005.

[2] Mendoza G, Tran B. In-home wireless monitoring of phys-iological data for heart failure patients. In Engineering inmedicine and biology, 2002. 24th Annual Conference and theAnnual Fall Meeting of the Biomedical Engineering SocietyEMBS/BMES Conference, 2002. Proceedings of the SecondJoint, volume 3. IEEE, 2002; 1849–1850.

[3] Gay V, Leijdekkers P. A health monitoring system usingsmart phones and wearable sensors’. International Journalof ARM 2007;8(2):29–35.

[4] Cardiac health - providing information, risk assessmentand advice about your heart health, 2011. URLhttp://www.cardiachealth.org/.

[5] Owis MI, Abou-Zied AH, Youssef A, Kadah YM. Study offeatures based on nonlinear dynamical modeling in ecg ar-rhythmia detection and classification. Biomedical Engineer-ing IEEE Transactions on 2002;49(7):733–736.

[6] Sun FT, Kuo C, Cheng HT, Buthpitiya S, Collins P, GrissM. Activity-aware mental stress detection using physiolog-ical sensors. In Mobile Computing, Applications, and Ser-vices. Springer, 2012; 211–230.

[7] Suzuki T, Oyama Y, Nakauchi Y. Intelligent medicine casesystem with distributed rfid readers. In Engineering inMedicine and Biology Society (EMBC), 2010 Annual Inter-national Conference of the IEEE. IEEE, 2010; 344–347.

[8] Wang L, Gu T, Tao X, Lu J. A hierarchical approach to real-time activity recognition in body sensor networks. Pervasiveand Mobile Computing 2012;8(1):115–130.

[9] Sufi F, Khalil I. Icardiogrid: Ecg analysis on demand to de-tect cardiovascular abnormalities. In ITAB, 9th InternationalConference on Information Technology and Applications inBiomedicine. ITAB, 2009; 5–7.

Address for correspondence:

Abdur Rahim Mohammad ForkanSchool of Computer Science and ITBuilding-14, Level-10, Room-06, RMIT University,City Campus, 414-418 Swanston Street, Melbourne,VIC-3001, Australiaabdur.forkan@rmit.edu.au

280