Development of a portable ECG and electronic stethoscope device for screening cardiovascular disease in rural locations.

Date
2018-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: Cardiovascular disease (CVD) is currently the number one cause of death worldwide, resulting in 17.7 million deaths in 2015 (World Health Organisation, 2017). In South Africa, five people suffer a heart attack every hour, placing CVD as the second deadliest disease in the country, after HIV/AIDS (Pillay-van Wyk et al., 2013). Studies predict that by 2030, CVD will be responsible for more deaths in developing countries than the total combined fatalities of HIV/AIDS, malaria and tuberculosis (Beaglehole and Bonita, 2008). Medical equipment required by cardiologists to diagnose cardiovascular disease is expensive and only available at larger hospitals in major cities within South Africa. This presents a significant challenge for the 35% of South Africans residing in rural areas who require medical attention (The World Bank, 2017). Subsequently, many patients in rural areas live with lingering cardiovascular problems. This thesis entails the design and development of a point-of-care device capable of screening for cardiovascular disease in rural locations in Africa. The device consists of an electrocardiogram (ECG) and electronic stethoscope capable of recording electrical bio-signals and heart sounds, respectively. The ECG consists of a reduced lead set that includes limb leads avL, avR, avF, I, II, III and precordial leads V2 and V4. The data recorded using the ECG can be used to autonomously identify patients with potential cardiovascular disease using machine learning techniques. Furthermore, the potential for reconstructing a full 12 lead ECG recording from a reduced lead set using machine learning is also investigated. Data acquired from the Physikalisch-Technische Bundesanstalt (PTB) online database was used to train the machine learning models. A deep pattern recognition neural network (DPRNN) was used to diagnose patients with normal or abnormal cardiac function. Additionally, a focus time-delay neural network (FTDNN) was used to reconstruct precordial leads V1, V3, V5 and V6 from the reduced lead set. The machine learning models were tested on 70 subjects recorded using the device in a clinical study conducted at Tygerberg Hospital. The classification method utilised first order features consisting of ECG amplitudes, intervals and segments, second order features derived from wavelet entropy and Shannon’s energy, as well as unsupervised features generated using stacked denoising autoencoders. The classification model, tested in the clinical trial, produced an accuracy, sensitivity, specificity and area under the curve (AUC) of 85%, 83%, 87% and 0.85, respectively. The ECG lead reconstruction produced acceptable root-mean-square error (RMSE) values of 181 to 266 µV, and excellent Pearson r correlation values of 0.91 - 0.95, for the reconstructed precordial leads. All correlation values were statistically significant at p « 0.01. The results obtained in this study compare favourably with an initial retrospective study as well as prior studies done in the research field. This evidence supports the possibility of deploying a low-cost portable device capable of referring patients with potential cardiac abnormalities, in rural locations, to hospitals for further examination.
AFRIKAANSE OPSOMMING: Kardiovaskulêre siekte (KVS) is huidiglik die voorstaande oorsaak van lewensverlies wêreldwyd, en het 17.7 miljoen lewens geëis in 2015 (World Health Organisation, 2017). In Suid-Afrika lei vyf mense aan ’n hartaanval elke uur, wat beteken dat KVS die tweende mees dodelike siekte in die land is, naas HIV/VIGS (Pillay-van Wyk et al., 2013). Studies voorspel dat KVS teen 2030 vir meer afsterwings verantwoordelik sal wees in ontwikkelende lande as die totale afsterwings van HIV/VIGS, malaria en tuberkulose saam (Beaglehole en Bonita, 2008). Die mediese toerusting wat benodig word deur kardioloë om kardiovaskulêre siektes te diagnoseer is duur en slegs beskikbaar by groter hospitale in hoof stede binne Suid-Afrika. Dit lei tot ’n beduidende uitdaging vir die 35% van die Suid-Afrikaanse bevolking wat in afgeleë areas bly en mediese hulp moet bekom (The World Bank, 2017). Gevolglik is daar menige pasiënte in afgeleë nedersettings wat lei aan voortdurende kardiovaskulêre kwale. Hierdie tesis onderneem die ontwerp van ’n punt-van-behandeling toestel wat in staat is om ondersoek in te stel vir kardiovaskulêre siektes in afgeleë liggings oor Afrika. Die toestel bestaan uit ’n elektrokardiogram (EKG) en elektroniese stetoskoop wat in staat is om elektriese bio-seine en hart klanke afsonderlik op te neem. Die EKG bestaan uit ’n verminderde probe terminaal stel wat die ledemaat probe terminale avL, avR, avF, I, II, III en voorhartversterkingsprobe terminale V2 en V4 bevat. Die data wat deur die EKG opgeneem word, kan gebruik word vir outomatiese identifisering van pasiënte wat potensieel lei aan kardiovaskulêre kwale deur middel van masjienleer tegnieke. Verder, word die potensiaal om ’n vol 12 terminaal EKG opname te herbou vanuit ’n verminderde terminaal stel deur middel van masjienleer ook ondersoek. Data wat verkry is vanaf die Physikalisch-Technische Bundesanstalt (PTB) aanlyn databasis was gebruik vir opleiding van die masjienleer modelle. ’n Diep patroon herkenning neurale netwerk (DPHNN) was gebruik om pasiënte te diagnoseer met normale of abnormale hart funksie. Daar is ook ’n bykomende fokus tyd vertraagde neurale netwerk (FTVNN) gebruik vir die herkonstruksie van die voorhartversterkingsprobe terminale V1, V3, V5 en V6 vanuit die verminderde stel probe terminale. Die masjienleer modelle was getoets op 70 opnames wat deur die ontwerpte toestel opgeneem is gedurende ’n kliniese studie by Tygerberg hospitaal. Die klassifikasie metode maak gebruik van eerste orde kenmerke, bestaande uit EKG amplitudes, intervalle en segmente, wat afgelei is vanaf golfvorm (wavelet) entropie en Shannon energie, asook sonder toesig gegenereerde kenmerke wat deur middel van gestapelde geraas kansellasie outokodeerders gegenereer word. Dis klassifikasie model wat getoets is in die kliniese toets het ’n akkuraatheid, sensitiwiteit, spesifisiteit en area onder kurwe van 85%, 83%, 87% en 0.85 afsonderlik gehad. Die EKG probe terminaal herkonstruksie het gelei tot aanvaarbare wortel van die gemiddelde kwadraat foute met waardes vanaf 181 tot 266 µV, en uitstekende Pearson r korrelasie waardes van 0.91 - 0.95, vir die geherkostruksieerde voorhartversterkingsprobe terminale. Alle korrelasie waardes was statisties van belang met p « 0.01. Die resultate wat verkry was in hierdie studie vergelyk gunstig met die oorspronklike retrospektiewe studie asook voorafgaande studies wat in die navorsingsveld onderneem is. Hierdie bewyse ondersteun die moontlikheid van ’n draagbare, lae koste toestel wat in staat is om pasiënte vanuit afgeleë nedersettings met potensiële hart abnormaliteite te verwys na ’n hospitaal vir verdere ondersoek.
Description
Thesis (MEng)--Stellenbosch University, 2018.
Keywords
Cardiovascular disease, Machine learning, Telemedicine, Screening, Medical -- Apparatus, UCTD
Citation