Building IoT-Enabled Wearable Medical Devices: an Application to aWearable, Multiparametric, Cardiorespiratory Sensor

Arthur Gatouillat1, Bertrand Massot2, Youakim Badr1, Ervin Sejdic3 and Claudine Gehin2

1Univ Lyon, INSA-Lyon, LIRIS, UMR5205, F-69621, France.2Univ Lyon, INSA Lyon, INL, UMR5270, F-69621, France.

3 Department of Electrical and Computer Engineering, Swanson School of Engineering, University of Pittsburgh, Pittsburgh,PA, USA.

arthur.gatouillat@insa-lyon.fr

Keywords: Wearable Sensor, Medical IoT, Heart Rate Sensor

Abstract: Recent developments in personal and mobile healthcare have shown promising results in term of patients’ qualityof life and quality of care improvements. This can be achieved through continuous monitoring of patients’physiological function using wearable non-invasive biomedical sensors. The remote collection and processingof such data can then be used to provide rapid medical response if a problem is detected or to offer preventivemeasures. However, the integration of wearable sensors into wider-scale framework is still a major challenges,as real-time data collection and remote configuration capabilities must be integrated to strongly constraineddevices. Here, we show how such requirements can be integrated into a multiparameter, cardiorespiraotywearable sensor and how this sensor can be integrated into wide-scale Internet-based frameworks. We thusmanufactured a biomedical-grade heart rate, instantaneous heart-rate variability and respiratory sensor. Thesensor tested in real life ambulatory condition, and we showed an Internet-based proof of concept exhibitingthe integration of our sensor into wide-scale healthcare frameworks. Finally, we anticipate that wearablehealthcare will greatly improve patients’ quality-of-life by using IoT-based wearable devices similar to thesensor developed in this paper.

1 INTRODUCTION

Personalized and mobile healthcare are growing fieldsof interest in the biomedical community. The ideabehind such concepts is to provide patients with healthrecommendations and diagnostic tailored to their indi-vidual needs. This can be achieved using continuousremote monitoring of the patients’ physiological func-tions associated with data analysis to offer preventivemeasures, or rapid medical response if a physiologi-cal malfunction is detected. Consequently, personal-ized healthcare is a promising approach to improvethe financial and therapeutic efficiency of healthcare(Van Hoof and Penders, 2013), by avoiding unneces-sary hospitalization while preserving the safety of thepatients.

Personalized healthcare can be achieved usingwearable sensors networks (Massot et al., 2013), whichare used to accurately monitor vital signs. Cardiachealth estimation devices are of particular interest be-cause of the crucial nature of the cardiac function andbecause it has been well explored by the biomedi-

cal sensor community, for instance in (Massot et al.,2015), (Magno et al., 2014), (Altini et al., 2011),(Khayatzadeh et al., 2013) and (Tuominen et al., 2017).Moreover, cardiac activity parameters can be used asan indicator of several pathologies. In particular, heartrate variability (HRV) provides insights about the au-tonomic nervous system parameters which reflects thepatients’ emotional state (Task Force of the EuropeanSociety of Cardiology the North American Societyof Pacing Electrophysiology, 1996). However, oneof the major research problem at the moment is theintegration of developed wearable sensors into biggerframeworks. Indeed, personalized healthcare assumeshealth data to be collected in real-time and that suchdata is analyzed by powerful algorithms on distantservers. Personalized healthcare also makes remotesensor reconfiguration necessary in order to better suitboth patients and healthcare professional expectations.

All the identified wearable electrocardiogram-based cardiac activity sensors, despite accurately mea-suring cardiac activity parameters, do no featureenough connectivity and functionality to be able to

use them in a personalized healthcare context. Onlywearable ECG or ECG-based cardiac parameters sen-sors were studied because of their accuracy in terms ofcardiac parameters estimation. The accuracy of otherpopular methods such as plethysmography is ques-tioned by the community (Schafer and Vagedes, 2013)and suffer from heavy movement artifact. Authors of(Magno et al., 2014) designed a wearable heart rate andrespiratory rate sensor with Bluetooth low energy andproprietary 802.15.4 connectivity by combining sev-eral development kits. This solution lacks remote con-figuration capabilities and proper integration, whichboth are critical aspects of personalized healthcare. In(Tuominen et al., 2017), a remotely configurable ECGsensor was designed, but data is stored locally on a SDcard, which makes this solution unsuitable for remotehealthcare applications. Another wearable ECG sensorwas developed by (Izumi et al., 2015), is remotely con-figurable, and data are sent using wireless connectivity.However, this device use near field communication(NFC) for all of its connectivity. The short range ofNFC application makes it unpractical for real life datacollection, as data can only be collected when a NFCtag is brought in proximity of the sensor. A heart rate(HR) and HRV sensor is described in (Massot et al.,2015). This device offers on-board cardiac parameterscalculation, and computed parameters are sent to asmartphone using wireless connectivity. However, thissensor is not remotely configurable, and data is onlystored locally on the smartphone, which makes someadaptation necessary in order for this device to be usedin a personalized healthcare concept.

Recently, the study of the interconnection of mul-tiple devices featuring real-time data collection andremote configuration capabilities was studied underthe scope of the Internet-of-Things (IoT) (Gubbi et al.,2013). In the IoT, a variety of heterogeneous com-municating devices are interconnected, but also com-municate with external services implemented on re-mote servers. Consequently, the IoT seems like anapproach of particular interest when trying to solve thechallenges of wide-scale personalized healthcare (Fer-nandez and Pallis, 2014), and this paper will describethe integration of IoT characteristics on a biomedicalwearable cardiac and respiratory sensor to enable itsintegration into a wide scale personalized healthcareframework.

Our solution focuses on the integration of remotereconfiguration capabilities to our sensor, the imple-mentation of different functional modes to enableadaptive capabilities, and the real-time streaming ofbiomedical data. All these functionalities must pre-serve measurements accuracy. The impact of integrat-ing IoT characteristics to the sensor on both hardware

and software choices, but also on the overall consump-tion of the device was carefully studied in order toensure our sensor validity from a real-life perspective.

The following paper is organized as follows. Sec-tion 2 introduces the materials and methods enablingthe integration on IoT properties to a biomedical gradesensor, but also describes the design of analog/digitalelectronics to achieve the desired functional and non-functional goals. In section 3, we present a compre-hensive caracterization of the sensor from a powerconsumption and data quality evaluation perspective,and finally section 4 discusses the potential applicationof such sensor for patients and medical practitioners.

2 MATERIALS AND METHODS

The manufactured IoT-enabled wearable cardiac andrespiratory activity sensor is displayed in Figure 1.a.This circuit board is small (40 mm×20 mm×6 mm),and it is packaged along with a 300 mAh battery in aplastic enclosure. The complete sensor (i.e., PCB, bat-tery, plastic enclosure and cables) weighs only 26.7 g,making it light enough for wearable applications.Thesensor is designed to be used with three electrodes:left arm, right arm and a common mode rejection elec-trode, as illustrated in Figure 1.b. The left arm andright arm electrodes are connected to the sensor us-ing wires, and the sensor is directly attached to thecommon mode rejection electrode. Consequently, thedevice can be comfortably worn by the patients forextended periods of time. The battery of the sensoris charged using the micro-USB port, and the sensorcan be reset to its default state by a push on the singlebutton of the device. This sensor is able to measureboth the heart rate, heart rate variability parametersand the respiration waveform (RWF).

40 mm

20 mm

(a) (b)

Figure 1: Manufactured sensor (a) and body placement (b)

2.1 Sensor Hardware Description

The global sensor architecture is given in Figure 2.Because of the potential overhead due to the addition

of IoT-related characteristic to our device, componentswere selected in order to maximize their computationaland power efficiency.

Analog front-endSignal amplification

Respiration measurementsSignal filtering

Analog-to-digital conversion

Bluetooth moduleAdvertising

Connection/Disconnection handlingData communication

ADS1292R

BLE113

Disposable ECGelectrodes

Digital bus

Programmable system-on-chipHardware ECG processing

So�ware parameters computationIoT state machine

Battery level monitoringPSoC5LPDigital bus

Figure 2: Hardware architecture of the sensor

In order to minimize CPU load and CPU wake-uptime caused by real-time signal processing, it mustbe performed using dedicated hardware. This led tothe choice of a PSoC 5LP (Cypress Semiconductor,San Jose, CA) for the device microcontroller and ded-icated signal processor. Indeed, this integrated cir-cuit (IC) offers both an ARM Cortex Core M3 CPUcore and a programmable digital filter block in a sin-gle package, leading to smaller PCB, thus improvingthe comfort of the patients. Because Bluetooth LowEnergy (BLE) is the main protocol used in resourceconstrained devices for the IoT and because of itswidespread adoption (Harris III et al., 2016), it wasselected to provide wireless connectivity to our sen-sor. The BLE113 integrated chip (Silicon Labs, Austin,TX) was selected as the BLE dedicated microprocessorbecause it integrates both a full hardware and softwareBLE stack with integrated antenna design in a small15.75 mm× 9.15 mm× 1.9 mm package. This inte-grated circuit is fully programmable, and all commu-nication dedicated firmware can be deployed on thisIC, thus reducing the microprocessor computing load.In order to be able to simultaneously measure both theECG and the respiration waveform, the ADS1292R(Texas Instruments, Dallas, TX) was selected. Thislow-power analog front-end (AFE) integrates two dif-ferential amplifiers and two 24 bits analog to digitalconverters. It also features a right-leg drive (RLD) am-plifier to implement common mode rejection (Winterand Webster, 1983) along with lead-off and respirationsignal modulation and demodulation circuitry. Themeasurements are sent using a SPI bus at a config-urable sampling rate (from 125 to 8000 samples perseconds).

In the default state, the analog front end is con-

figured to measure both the respiration waveform onchannel 1 and the ECG signal on channel 2 at a sam-pling rate of a 1000 SPS using the internal clock of theIC. This sampling rate enables a 1 ms resolution for RRinterval measurements. The RLD amplifier and lead-off detection circuitry are both enabled, along with therespiration modulation and demodulation modules.

The RR interval computation is implemented us-ing hardware filtering and peak detection. The digitalfilter block of the PSoC5LP is used to compute thesmoothed moving average derivative of the ECG sig-nal, which is then used in combination with adaptivethresholding in order to compute the RR interval. De-tails about the implementation are given in Figure 3and in (Massot et al., 2016).

>

<

count

capture

reset

comp

value

Up CounterSampling CLK

1 kHz

DMA1Trig. Int.

Counter valuetranfered to RAMwith DMA1

DFB

DelayedDFB

Output

AdaptativeThreshold

ECG signal from AFEtransfered withexternal DMA

Figure 3: Hardware detection of RR intervals

Using this hardware-centric approach to performcomputationally intensive real-time signal processingreduces the CPU load and increases the CPU sleeptime, thus reducing global power consumption. All theother sensor functional tasks are implemented using asoftware approach, and will thus be further describedin the following section.

2.2 Sensor Software Description

The firmware of the sensor was developed to achieveseveral functional and non-functional goals. The firstgoal is focused on the measurements and their trans-mission: the real-time communication of the instanta-neous heart rate, the computation and communicationof HRV parameters every 5 minutes, the real-time com-munication of 6 respiration samples every second. Thesecond goal is to transmit non-functional informationabout the sensor, typically the battery level or if thesensor is unattached. Finally, the last goal is the im-plementation of IoT dedicated software, which is usedto integrate the sensor into a wider IoT-based frame-work. Such software is used to transmit informationabout the sensor state in order to trigger remote sensorconfiguration.

2.2.1 Measurements Dedicated Software

This firmware is implemented on the PSoC5LP, andit can be divided into three sections: the computationof the instantaneous heart rate when a heart beat isdetected (i.e., when an interruption is triggered by thepeak detection hardware). The heart rate can be easilycomputed using the following equation:

HRBPM =60×SRRRint.

=60×1000

RRint.(1)

where HRBPM is the instantaneous heart rate in beatsper minutes (BPM), SR is the sampling rate in samplesper seconds and RRint. is the RR interval in millisec-onds. This computation is implemented directly on themicrocontroller and the RR interval values are bufferedto be used in the HRV parameters computation.

Every 5 minutes, the HRV parameters are com-puted from the buffered RR interval values in accor-dance with the recommendations of the joint TaskForce of the European Society of Cardiology (TaskForce of the European Society of Cardiology the NorthAmerican Society of Pacing Electrophysiology, 1996),and these parameters are summarized in Table 1.

Table 1: Computed HRV parameters

Variable Unit Domain DescriptionSSDN ms Time Standard deviation of buffered RR

intervalsRMSSD ms Time Quadratic mean of differences be-

tween consecutive RR intervalsLF/HF n.u. Freq. Low-frequency (0.04 to 0.15 Hz)

to high-frequency (0.15 to 0.4 Hz)components ratio of the PSD of thebuffered intervals

Norm. LF % Freq. Normalized low-frequency compo-nents to sum of low- and high-frequency components of the PSDration, i.e., LF/(LF+HF)

The computation of the frequency domain pa-rameters are based on the power spectral density(PSD) estimation. Because the RR intervals signalis unevenly sampled, the Lomb-Scargle periodogram(Lomb, 1976) was determined using the fast computa-tion algorithm developed by (Press and Rybicki, 1989).We used the same optimal algorithmic parameters than(Massot et al., 2016). Indeed, authors of (Massot et al.,2016) conducted a systematic evaluation of the impactof the parameters on the trade-off between computa-tion time and periodogram precision, resulting in theproposition of optimal parameters with respect to boththe periodogram precision and the computation time.

The respiration waveform is buffered in the RAMof the PSoC5LP, and is transmitted wirelessly everysecond. The first step of the respiration signal pro-

cessing is the downsampling of the respiration sig-nal from 1000 SPS to 6 SPS. A sampling frequencyof 6 Hz, resulting in a Nyquist rate of 3 Hz, is wellwithin the bandwidth of respiratory signals, which isat most 1.5 Hz (Zhao et al., 1994), and decreasing thesampling rate reduces energy consumption because ofthe smaller amount of transmitted information. Af-ter signal downsampling, the respiration waveform issmoothed using an exponential filter:

y[n] = (1−α)y[n−1]−αx[n] (2)

, with α a configurable filter parameter. For the pre-liminary results, this parameter was set to 0.5. Oncethe signal is filtered, the value are communicated tothe Bluetooth IC using a signed 24 bits integer format.

2.2.2 Communication Dedicated Software

In this section, the firmware of the BLE113 IC will bedetailed. The scripting language provided by SiliconLabs, BGScript, was used to program this firmware. Acustom GATT server was implemented with respectto the heart rate profile detailed in the Bluetooth 4specification: the mandatory generic access profile, de-vice information, heart rate and battery services wereimplemented. The heart-rate service was augmentedwith a customized heart rate variability characteris-tic. This characteristic is of length 5, and the SSDN,RMSSD, LF/HF ratio and normalized LF componentsratio are encoded as single-byte unsigned integers. Anadditional service was implemented for the respirationalong with another service dedicated to IoT-relatedinformation (i.e., information about the sensor non-functional states).

The BLE113 also implements device advertise-ment when the sensor is not connected and if a discon-nection occurs according to the heart rate profile: theadvertising interval is between 20 and 30 ms for thefirst 30 seconds, and it is increased to be between 1and 2.5 s for the next 30 seconds. If no device connectto the sensor after this advertising period, the sensor isput in a very low-energy deep sleep mode.

2.2.3 IoT Dedicated Software

To enable our respiratory and cardiac activity sensorwith IoT-based characteristics in order to include theobject in wider IoT-based frameworks, the require-ments are three-fold:• The data must be streamed in real time, meaning as

soon as a data is acquired, it must be transmitted.• The sensor must be able to provide information

about its functional and non-functional state inorder to be able to build smart-scenarios based onthe IoT-enabled devices.

• The sensor must be remotely configurable in or-der to be integrated to wider-scale auto-adaptationscenarios.

In order to fulfill this set of requirements, the sen-sor was modeled using a labeled transition system(LTS). LTSs are widely used models of computation(MoC) because of their relative simplicity and highexpressiveness.

The LTS of our sensor is given in Figure 4. Thetransitions between the states are either controllable(i.e., a remote tier can force the transition) or non-controllable (i.e., the sensor automatically determinesits state from environmental measurements). As aconvention, a transition written as “a \ b” means thetransition is triggered on detection of internal event aor is remotely triggered using event b. When a transi-tion is only labeled with a single event, it is assumedthat this event is internal.

As displayed in Figure 4, the sensor has 5 non-functional states:

• When turned on or after a reset operation, the sen-sor is in the initialization state. In this state, theAFE is configured with the default parameters andthe BLE113 is software reset. This state automat-ically calls the normal operation state when theinitialization process is over.

• The default state is the normal state. In this state,both ECG and respiration signal are acquired. Theinstantaneous heart-rate is sent in real time (i.e.,the new value is transmitted as soon as it is avail-able), the HRV parameters are computed every 5minutes, and the respiration signal is packaged andtransmitted every second.

• The failsoft state is an energy saving state. It istriggered either externally or when the battery levelof the sensor reaches a low level (i.e., 20% in ourcase).

• When the sensor detects a disconnected lead, itis placed in the unattached state. In this state, atimer is launched, and the sensor is turned off ifthe timer overflows. If the sensor is reattached, itgoes back to the normal operation state.

• The stop state denotes a very low-energy deepsleep mode. The sensor can be placed in this stateat any time.

The BLE113 can generate “no connection” (abbre-viated as “no conn” in the LTS) events if the sensordoes not receive any connection request during theadvertising interval. This causes the sensor to go intothe very low power stop state.

Practically, this LTS is implemented on the ARMcore of the PSoC5LP. Events are either generated in-

Init.

Normal Failso�Unatt.

Stop

init_done

batt < 20% \ setMode(failso�)

batt ≥ 20% \ setMode(normal)

(batt < 5% ∨ no_conn) \ setMode(stop)

no_conn \ setMode(stop)

(timeout ∨ no_conn) \ setMode(stop)

hardware_reset

unattached

unattached

Figure 4

ternally or transmitted by the BLE113, and the sensorstate is exposed using a custom service.

2.3 Gateway Software and IoT-BasedFramework

Android phones were chosen as the main gatewaysto our sensor. This choice was motivated principallybecause phones can be carried by the subjects, result-ing in a mobile gateway with perpetual connection tothe Internet. Consequently, an Android applicationwas developed to connect to the sensor, but also toplot the data, store it locally on the phone or trans-mit them over the Internet. The application is basedon the nRF Toolbox open-source application1 devel-oped by Nordic Semiconductor (Oslo, Norway), andscreenshots of the application are given in Figure 5.

Figure 5: Companion Android application

Because data is collected in real time and sensorconfiguration can be triggered at any moment, a pub-

1https://github.com/NordicSemiconductor/Android-nRF-Toolbox

lish/subscribe type protocol was chosen. Practically,the Eclipse Paho MQTT client was integrated in theapplication. MQTT is a lightweight publish/subscribemessaging Internet-based protocol. It was designed tobe used in memory and bandwidth constrained devices.This protocol is built around the notion of topics. Atopic is an URI-like string used to describe transitingdata. In the MQTT terminology, a publisher sends datain a topic. If a subscriber subscribes to the same topic,it will receive the data instantaneously. Formally, alldata is sent through a central broker that can be seen asa hub distributing data between the relevant publishersand subscribers. This publish/subscribe mechanismthus implements real-time data streaming over the In-ternet, and enables connection to various external Webservices or other connected devices.

3 RESULTS

3.1 Power ConsumptionCharacterization

In this section, a comprehensive evaluation of the sen-sor power consumption with respect to connectionparameters and non-functional IoT-state defined in Fig-ure 4 is performed.

Current consumption measurements were per-formed using a Keithley 2400 sourcemeter (Beaverton,OR, USA). The first characterization to be performedwas a static power consumption evaluation of the nor-mal, failsoft, and stop states. The initial state was notmeasured because it is only a transitional state with ashort duration, and the unattached state was not mea-sured because the sensor is in the same hardware con-figuration than in the normal state, resulting in equalpower consumption. This static characterization wasperformed using default Android 7.0 BLE connectionparameters (i.e., connection interval of 48.75 ms, time-out of 20 s and latency of 0). The consumption resultsare given in Table 2. With a power consumption of10.53 mW in the normal state, the battery life of thesensor if of about 75 hours using a 300 mAh battery.This battery life is extended to about 85 hours if thefailsoft mode is used. In the stop mode, the sensor canlast more than two months on a fully charged battery.

Table 2: Static power characterization of IoT states

IoT state Power consumption (mW)Normal 10.53Failsoft 9.18

Stop 0.505

In order to further optimize power consumption

of the sensor, the influence of the BLE connectionparameters were also studied. There are three BLEconnection parameters: interval, latency and timeout.The connection interval designates the period betweentwo master request to the slave, while the latency de-fines the number of connection intervals that can beignored safely by the slave. Finally, the connectiontimeout designates the period after which the BLEmaster will consider that the connection with the slaveis lost. It is only after the timeout period that the BLEmaster can attempt with reconnection. For the sensor,we fixed both the connection interval and connectiontimeout and the impact of the latency was studied onthe overall power consumption. The results of thischaracterization are given in Figure 6, and the con-sumption of both the normal IoT state and the failsoftIoT state are plotted against the latency for variousconnection interval and connection timeout. For allstates and connection interval and timeout parame-ters, and increase in the latency results in decreasingoverall power consumption, with a stronger decreasefor smaller latency values. It is worth noting that fora latency of 8, connection instability was observed.In order to enable smaller power consumption, whilekeeping acceptable connectivity quality, the sensorrequests a latency of 2 when a new connection is es-tablished. However, masters (such as Android phones)can reject connection parameters update and force anew set of less constrained connection parameters (i.e.,smaller connection intervals, zero latency and smallerconnection timeout), causing an increase in the overallpower consumption. This problem was solved with aslight oversizing the battery capacity in order to alwaysachieve at least 48 hours of battery life.

0 1 2 3 4 5 6 7 8Latency

6.5

7.0

7.5

8.0

8.5

Sens

or p

ower

con

sum

ptio

n (m

W)

Normal mode, interval = 67.5 ms, timeout = 1.0 sFailsoft mode, interval = 67.5 ms, timeout = 1.0 sFailsoft mode, interval = 742.5 ms, timeout = 10.0 s

Figure 6: Characterization of the BLE connection latencyparameter effect on the overall power consumption

3.2 Data Acquisition

In this section, various data acquisition are presented.This first experiment to be performed was the valida-

tion of RR-interval acquisition using synthetic ECGsignals generated using an Agilent 33220A (SantaClara, CA, USA) arbitrary waveform generator. Theobjective of this test is twofold: it validates the com-putation of RR-interval and the reliability of the BLEconnection. In order to simulate heart rate variability,triangular frequency modulation was applied to thesynthetic ECG signal:

base frequency = 1.25 Hzfrequency deviation = 200 mHz

(3)

which theoretically results in RR intervals rangingbetween:

RRint <1000

1.25−0.2≈ 952.4 ms

and, RRint >1000

1.25+0.2≈ 689.7 ms

(4)

The sensor was then connected to a Raspberry Piused as a BLE data logger used to record RR-intervalsand HRV parameters, and an excerpt of the collectedRR interval values is given in Figure 7. The exper-iment was performed during 1 day, 17 hours and 9minutes. A total of 185153 were logged. Data was an-alyzed for records violating the theoretical limits (witha tolerance of 1% over the limits), and no violationswere found.

0 250 500 750 1000 1250 1500 1750time (s)

700

750

800

850

900

950

RR In

terv

al (m

s)

Figure 7: RR interval determined from simulated ECG signal

The next experiment was data acquisition duringshort respiratory exercises. During this exercise, ayoung healthy patient was asked to perform a seriesof forced inspiration and expiration during 1 minutein order to trigger sinus arrhythmia. Results fromthis experiment are given in Figure 8, where both theinstantaneous heart rate and the normalized respiratorywaveform are plotted. This figure clearly illustratessinus arrhythmia, with increasing heart rate duringinspiration because of a decreasing vagal tone anddecreasing during expiration.

The final experiment consisted on a short-term am-bulatory testing of the sensor. For this experiment, apatient was asked to wear the sensor during one hour,while performing normal daily activities. Cardiorespi-ratory data was recorded using the companion Android

600

700

800

RR In

terv

al (m

s)

60 70 80 90 100 110 120time (s)

0.0

0.2

0.4

0.6

0.8

1.0

Nor

mal

ized

RW

F

Figure 8: RR-Interval and respiratory waveform during con-trolled respiratory exercise

application, and results from this experiment are givenin Figures 9 and 10. Figure 9 displays both the instan-taneous RR interval and normalized respiration signals.No measurement artifact were observed for both thesignals. ECG artifacts, resulting in RR interval arti-facts, are typically caused by electrical muscular activ-ity interfering with ECG signals. Such adverse effectswere minimized by placing on the lower part of the ribcage. Indeed, this location features minimal musclethickness, thus reducing interactions between ECG sig-nal and electrical muscular activity. It is worth notingthat the respiration signal features a slowly evolvingcomponent. This is because the respiratory activity isin fact an impedance measurement, and because theimpedance of the electrode-skin interface can slowlyevolve over time because of various factors (such ashumidity, the presence of sweat, etc.).

400

600

800

1000

RR In

terv

al (m

s)

0 500 1000 1500 2000 2500 3000 3500time (s)

0.0

0.2

0.4

0.6

0.8

1.0

Nor

mal

ized

RW

F

Figure 9: Real-life RR-Interval and respiratory waveform

Figure 10 displays the 4 heart rate variability pa-rameters plotted as a function of time. These HRVparameters correspond to the HR measurements ex-

hibited in Figure 9, and fall within the typical rangesdefined by the literature (Nunan et al., 2010).

40

50

60

70

80

SDN

N (m

s)

20

30

40

50

RMSS

D (m

s)

0 1000 2000 3000 4000 5000time (s)

50

60

70

80

LF n

.u. (

%)

0 1000 2000 3000 4000 5000time (s)

1

2

3

4

5

6

LF/H

F

Figure 10: Real-life HRV parameters

The preliminary results presented herein-aboveproves that our wearable sensor provides high qualitydata. The next experiment will deal with the integra-tion of our wearable sensor in wider-scale IoT-basedhealthcare framework.

3.3 Integration to an IoT-BasedFramework

As specified in Section 2, the integration of our wear-able sensor to wider-scale healthcare frameworks oc-curs through the Android companion application. In-deed, upon sensor connection, the application attemptsto connect to a local MQTT broker. If the connectionis successful, the application will publish heart rateand HRV values as soon as they are available on twotopics:

• interface/hr

• interface/hrv

As a result, cardiac parameters are streamed in realtime using these two topics, and this is the basis tobuild wider-scale healthcare framework. As a proofof concept, a monitoring Web-based graphical userinterface was implemented. It simulates the kind ofGUI that can be made available to physicians in orderto have real time health information on their patients.A sample screenshot of the Web-based GUI is givenin Figure 11.

4 DISCUSSION

In this paper, we introduced a wearable cardiores-piratory sensor that can be easily integrated to wider-

Figure 11: Web application visual

scale healthcare framework. This sensor satisfies cru-cial healthcare-related requirements such as data reali-ability, remote configuration capabilities or scalability.Wearable healthcare and the Internet-of-Things havesimilar objectives: the use of a mass of connected ob-jects (which also are worn by patients in the contextof wearable healthcare) that monitor physical param-eters in order to trigger based on data analysis. Forinstance, our wearable sensor can be integrated in aliving-lab, where the living environment of the patientis continually monitored. This integration relies on theuse of widely adopted technologies such as BLE, butalso on Internet-based technologies such as MQTT inorder to enable wide-scale connectivity. Internet-basedconnectivity implies that our wearable sensor is ableto connected to Internet healthcare services, which canperform advanced data analysis and detect potentialhealth crisis of the patient and trigger urgent medicalresponse.

Wearable healthcare have a strong potential on theimprovement of the patients’ quality of life and qualityof care. Indeed, the wearable nature of devices en-able patients to be continuously remotely monitored.Combined with recent advances in home-automation,wearable healthcare will drastically improve in-homecare for a variety of patients. Collected data processedusing big-data techniques or advanced signal process-ing could also have great predictive values for theevolution of chronic diseases and could be used toprovide better and earlier care for patients.

However, the personal nature of collected healthdata mandates strong security mechanism, which werenot explored in our contribution. Indeed, at the mo-ment our system only features basic data encryptionin compliance with the BLE standard, and the MQTT-based solution does not use any security mechanism.Security of our overall system needs to be improved,more particularly in terms of access control (i.e., thepatients of physicians must be able to know and con-trol who accesses their medical data) and identity

management, and recently developed decentralizedblockchain-based solutions (Zhu et al., 2017) can beexplored in order to provide comprehensive werablehealthcare system security.

5 CONCLUSION

In this paper, we presented a multiparametric, car-diorespiratory wearable sensor. In order to answerhealthcare requirements, more particularly in terms ofthe ability to integrate wearable sensors into wider-scale frameworks, we considered an IoT-based ap-proach. Indeed, we equipped our sensor with remoteconfiguration capabilities while preserving quality-of-data and real-time streaming capabilities, which arekey requirements of wearable healthcare systems. Thissensor was implemented using carefully selected hard-ware, and it was comprehensively characterized interms of energy consumption, which is another majorconcern of wearable healthcare devices. Indeed, be-cause battery charging usually implies the sensor isnot collecting physiological data, potentially relevantdata can be lost, and the charging time to battery liferation must thus be as big as possible. The data col-lection capabilities of our sensor were also extensivelytested on both synthetic ECG signals and in real-lifeambulatory conditions. Successful testing and inte-gration to Internet-based framework proved that oursensor can be used in a wide-scale wearable healthcareframework.

ACKNOWLEDGMENT

The authors would like to thank the COOPERAfunding program of Region Auvergne Rhone-Alpesfor their generous financial support.

REFERENCES

Altini, M., Polito, S., Penders, J., Kim, H., Van Helleputte,N., Kim, S., and Yazicioglu, F. (2011). An ECG patchcombining a customized ultra-low-power ECG SoCwith bluetooth low energy for long term ambulatorymonitoring. In Proceedings of the 2nd Conferenceon Wireless Health, pages 15:1–15:2, New York, NY,USA. ACM.

Fernandez, F. and Pallis, G. C. (2014). Opportunities andchallenges of the internet of things for healthcare: Sys-tems engineering perspective. In Proceedings of theEAI International Conference on Wireless Mobile Com-munication and Healthcare, pages 263–266.

Gubbi, J., Buyya, R., Marusic, S., and Palaniswami, M.(2013). Internet of things (IoT): A vision, architecturalelements, and future directions. Future GenerationComputer Systems, 29(7):1645 – 1660.

Harris III, A. F., Khanna, V., Tuncay, G., Want, R., andKravets, R. (2016). Bluetooth low energy in denseIoT environments. IEEE Communications Magazine,54(12):30–36.

Izumi, S., Yamashita, K., Nakano, M., Kawaguchi, H.,Kimura, H., Marumoto, K., Fuchikami, T., Fujimori,Y., Nakajima, H., Shiga, T., and Yoshimoto, M. (2015).A wearable healthcare system with a 13.7 µA noise tol-erant ecg processor. IEEE Transactions on BiomedicalCircuits and Systems, 9(5):733–742.

Khayatzadeh, M., Zhang, X., Tan, J., Liew, W. S., and Lian,Y. (2013). A 0.7-V 17.4-µW 3-lead wireless ECGSoC. IEEE Transactions on Biomedical Circuits andSystems, 7(5):583–592.

Lomb, N. R. (1976). Least-squares frequency analysis ofunequally spaced data. Astrophysics and Space Science,39(2):447–462.

Magno, M., Spagnol, C., Benini, L., and Popovici, E.(2014). A low power wireless node for contact andcontactless heart monitoring. Microelectronics Journal,45(12):1656–1664.

Massot, B., Noury, N., Gehin, C., and McAdams, E. (2013).On designing an ubiquitous sensor network for healthmonitoring. In Proceedings of the International Con-ference on e-Health Networking, Applications and Ser-vices, pages 310–314.

Massot, B., Risset, T., Michelet, G., and McAdams, E.(2015). A wireless, low-power, smart sensor of cardiacactivity for clinical remote monitoring. In Proceedingsof the International Conference on E-health Network-ing, Application Services, pages 488–494.

Massot, B., Risset, T., Michelet, G., and McAdams, E.(2016). Mixed hardware and software embedded sig-nal processing methods for in-situ analysis of cardiacactivity. In Proceedings of the 9th International JointConference on Biomedical Engineering Systems andTechnologies, pages 303–310.

Nunan, D., Sandercock, G. R., and Brodie, D. A. (2010).A quantitative systematic review of normal valuesfor short-term heart rate variability in healthy adults.Pacing and Clinical Electrophysiology, 33(11):1407–1417.

Press, W. H. and Rybicki, G. B. (1989). Fast algorithmfor spectral analysis of unevenly sampled data. TheAstrophysical Journal, 338:277–280.

Schafer, A. and Vagedes, J. (2013). How accurate is pulserate variability as an estimate of heart rate variability?:A review on studies comparing photoplethysmographictechnology with an electrocardiogram. InternationalJournal of Cardiology, 166(1):15 – 29.

Task Force of the European Society of Cardiology the NorthAmerican Society of Pacing Electrophysiology (1996).Heart rate variability: Standards of measurement, phys-iological interpretation, and clinical use. Circulation,93(5):1043–1065.

Tuominen, J., Lehtonen, E., Tadi, M. J., Koskinen, J., Pnkl,M., and Koivisto, T. (2017). A miniaturized low powerbiomedical sensor node for clinical research and longterm monitoring of cardiovascular signals. In Proceed-ings of the International Symposium on Circuits andSystems, pages 1–4.

Van Hoof, C. and Penders, J. (2013). Addressing the health-care cost dilemma by managing health instead of man-aging illness: An opportunity for wearable wirelesssensors. In Proceedings of the Conference on De-sign, Automation and Test in Europe, pages 1537–1539.EDA Consortium.

Winter, B. B. and Webster, J. G. (1983). Driven-right-legcircuit design. IEEE Transactions on Biomedical Engi-neering, 30(1):62–66.

Zhao, L., Reisman, S., and Findley, T. (1994). Respirationderived from the electrocardiogram during heart ratevariability studies. In Proceedings of 16th InternationalConference of the IEEE Engineering in Medicine andBiology Society, pages 123–124.

Zhu, X., Badr, Y., Pacheco, J., and Hariri, S. (2017). Auto-nomic identity framework for the internet of things. InProceedings of the International Conference on Cloudand Autonomic Computing, pages 69–79.