a r t i c l e i n f o a b s t r a c t
Article history: In the emerging eld of wearable systems for remote monitoring of physiological parameters, the mea-
Received 30 September 2014 surement of bioimpedance has the potential to provide many useful information. On the other hand, in
Received in revised form 5 May 2015 this scenario, an optimization of power consumption of the circuit is crucial. A low power architecture for
Accepted 5 May 2015
the measurement of bioimpedance was identied in this work. It reduces the consumption in the most
Available online 12 May 2015
critical blocks of the system: the current driver, the signal sensing and the demodulator. The device was
Keywords:
prototyped and electrically characterized. The compromise between power consumption reduction and
Bioimpedance the increase in electrical noise was analysed and an effective signal processing technique was developed,
Wearable system showing that it is possible to achieve a signal to noise ratio good enough to enable applications like res-
Remote monitoring piration monitoring (breathing rate and amplitude) or cardiac output estimation. Preliminary tests on
Voltage to current converter healthy subjects showed a good correlation with spirometer, for breathing monitoring, and with Doppler
Impedance measurement echocardiography, for cardiac output. Thanks to the good functionality and the reduced current con-
Impedance cardiography sumption (750 A at 2.8 V power supply was obtained with a discrete-components implementation) the
module resulted suitable for the integration in wearable devices for remote monitoring of physiological
parameters, or other low power applications.
2015 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sna.2015.05.004
0924-4247/ 2015 Elsevier B.V. All rights reserved.
360 S. Rossi et al. / Sensors and Actuators A 232 (2015) 359367
2. Hardware overview
Fig. 2. Block diagram of a bioimpedance device using current sense resistor to mea-
sure the actual injected current. Fig. 4. Schematic of the VI converter.
S. Rossi et al. / Sensors and Actuators A 232 (2015) 359367 361
the impedance seen between the sensing electrode and the inter-
nal voltage reference of the circuit. The IA performs the difference
between the two baseband signals and provides an output propor-
tional to the bioimpedance Z(t). Given the operational amplier,
the use of unity gain buffers allows reaching the maximum band-
width; therefore, the gain-bandwidth product of the amplier can
be minimized with benets in power consumption. In our design
the TSV630 (STMicroelectronics) operational amplier was chosen
for the buffers (current consumption 55 A, gain-bandwidth prod-
uct 860 kHz) and the LT1789 (Linear Technology Corporation) was
selected as IA (current consumption 67 A).
The sampling signal is provided by the micro-controller. Accord-
ingly to Eq. (1), in order to measure the module of the impedance,
the system shall be able to sample the peaks of the sinusoidal
voltage drop on Z(t). This is obtained through an automatic self-
calibration performed by the system at the beginning of the
acquisition (Fig. 6). Starting from an arbitrary phase shift phi0
between the sampling pulse and the square wave from which the
quasi-sinusoidal current is generated, the value of the impedance
M is acquired, then the phase shift phi is changed of a xed quan-
Fig. 5. Output impedance of the VI converter versus frequency.
tity step and a new value of the impedance New M is acquired and
compared with M, in order to update the value of M accordingly
In this way the dominant pole of the op-amp, which in gen- to whichever is the greater. The procedure is repeated until the
eral constitutes the deviation from the ideal behaviour causing the maximum M is found. This value corresponds to the peak of the
reduction of the output impedance, is directly used for the syn- acquired voltage expressed by Eq. (1), and represents the module
thesis of the needed circuit properties. High output impedance at of the impedance. Furthermore, since the delay introduced by the
the working frequency can be obtained with a narrow bandwidth low-pass lter and the voltage to current converter in the current
amplier, with benets in power consumption. In this design a driving chain is known, the nal phase shift between the square
TSV6390 (STMicroelectronics) operational amplier is used (cur- wave and the sampling pulse obtained from the procedure allows
rent consumption 55 A), the theoretical output impedance at the estimating the impedance phase.
working frequency results higher than 1 M when the circuit is Since the phase shift is quantized, it is possible that the sampling
driven by a pure sine-wave. When the circuit is driven by a ltered pulse, instead of sampling exactly the peak of the voltage, samples
square wave, as in this work, the output impedance is very high a point close to it, causing an error in the reconstruction of the
only for the fundamental harmonic, while it is lower for the other module of the impedance:
harmonics. This determines an increase in the current depend-
ence on the load, which can be modelled with a theoretical output z = R Mz ej(z z )
M z
= Mz cos z
impedance of about 200 k, fully compliant with the application,
as discussed in Section 5. in which M z is the measured module, which results an approxima-
tion of the actual module Mz because of the wrong estimation z of
2.2. Demodulator the actual phase z . It is worth to highlight that, since the error on
the module depends on the cosine of the error on the phase, very
The demodulator is realized through AC coupled buffers, with good accuracy on the module is still obtained even with a quite
high input impedance and high-pass transfer function, and sam- rough quantization on the phase shift steps: with a phase error of
ple and holds. Each buffer is used to sense the voltage of one of 5 degrees, the error on the module is only 0.4%. Therefore, it is not
the sensing electrodes; this voltage is then sampled at the car- needed to continuously run the calibration: since the time depend-
rier frequency, in order to obtain a baseband signal proportional to ent changes of the impedance phase z (t) are small, z (t) can be
approximated with a constant with negligible error on the module, valid breath only if its peak respects specic requirements based
and once the right phase shift for the sampling pulse is calculated, it on its amplitude and on its distance with the adjacent troughs.
can be used to continuously sample the module of the impedance. Finally, three different techniques are used to detect noise
With this approach minimum consumption is required, since affecting the signal:
only one chain with a sample and hold is used and the accuracy
of the module measurement is good. This is important because the Electrode detachment detection
module is the main information used in many important applica- Z0-based identication of saturation or no current injection
tions, like breathing monitoring [16] or impedance cardiography Very small breath amplitude due to poor contact quality
both for cardiac output estimation [17] or for uid status assess-
ment [2]. For other applications, like body composition, also the
If at least one of these techniques marks the signal as unreliable, the
phase is needed [1], but this kind of measurements are less interest-
resulting breathing rate is not computed. If instead the signal qual-
ing in terms of a continuous monitoring through a wearable system,
ity is good, the respiratory rate is easily estimated as the number
because only a few measurements per day are needed. In any case,
of breaths detected in the last 60 s.
a phase estimation of the impedance is obtained with the proposed
method and, furthermore, this architecture is very exible and it is
also possible to use a synchronous sampling approach [14] to obtain 3.2. Cardiac output extraction
the real or imaginary part of the impedance.
Cardiac output estimation is based on the analysis of the Z
waveform and of its rst derivative dZ/dt [17,19]. The typical range
3. Data acquisition system and processing overview for Z amplitude is 0.10.2 [17]. With so small variations noise
becomes an important issue to deal with, especially when the
The hardware architecture described in Section 2 is compliant excitation current is decreased to reduce the power consumption:
with many bioimpedance applications. Our target, in this work, is lower current means lower input voltage in the read-back circuit,
to demonstrate its suitability in trans-thoracic applications, in par- and this results in a poorer signal to noise ratio (SNR). Typical val-
ticular for breathing monitoring and cardiac output extraction in ues reported in literature [19] and used by commercial device like
wearable devices. Niccomo (medis. Medizinische Messtechnik GmbH) for the exci-
The bioimpedance module was integrated in a device also tation current are in the range 14 mA. As mentioned in Section
including an analog front-end for electrocardiogram (ECG), a 3-axis 2.1, a current of 100 A is used in this work; therefore, with such
accelerometer for activity and body posture estimation, and a Blue- a decrease in current level, a low SNR is unavoidable. Furthermore,
tooth transceiver. Bioimpedance and ECG signals were acquired as discussed in Section 5, since the overall architecture is optimized
through the analog to digital converter embedded into the micro- to minimize the current consumption, an increased level of noise
controller (STM32F1, STMicroelectronics) and transmitted to a PC is unavoidable.
for signal processing and parameters extraction. A detailed descrip- Taking into account these considerations, the digital signal
tion of the complete device is outside the scope of this paper, processing for cardiac output estimation was designed. The goal
which focuses on the bioimpedance front-end. It is mentioned was to recover a SNR good enough to allow cardiac output estima-
here because it was used as the data acquisition system for a tion, according to methodologies widely presented in literature. In
deeper evaluation of the bioimpedance module, allowing both elec- particular, after a rst bandpass ltering stage (3 dB bandwidth:
trical characterization and data acquisition on healthy subjects. 0.76 Hz) which reduces the noise outside the bandwidth of the
In terms of safety, it was tested accordingly to the IEC 60601 signal (both due to breathing and electronics), ensemble averag-
by a notied body and it obtained the CE marking before the ing [20] is used to improve the signal quality by reducing the noise
execution of the tests reported in Section 4. All the algorithms inside the signal bandwidth. The signal periods corresponding to
were developed for an easy integration into the on board micro- cardiac cycles are identied through the ECG R-peaks position,
controller; their computational cost was controlled, avoiding, for then, being pi (n) the period i of the signal, the sample k of the
example, frequency-domain analysis or other resource consuming ensemble averaged period p is computed as:
approaches.
1
N
p(k)
= pi (k)
3.1. Breathing extraction N
i=1
The rst step in the processing of the raw Z signal for breath- where N is the number of periods used for the average.
ing analysis is a down-sampling from the original 256 Hz sampling The strong effectiveness of this approach compared with sim-
frequency to 32 Hz. This is possible because the typical spectral pler solutions, like, for example, averaging of consecutive samples,
components of the respiratory activity have an extremely low is related to the nature of the noise. It is well known that averag-
frequency (well below 1 Hz), and by decreasing the number of oper- ing N random samples xi results in the strongest reduction of the
ations required for each sample, we can achieve a longer battery variance when the samples are uncorrelated, since it is:
life. A tight bandpass ltering is then performed to isolate the spec-
1 1 1
tral components of interest. This is accomplished by combining a N N
low-pass IIR lter with a 3 dB cut-off frequency of 0.27 Hz with a Var xi = Var (xi ) + 2 Cov xi xj .
computationally efcient high-pass FIR lter with a 3 dB cut-off N N2 N
i=1 i=1 i=
/ j
frequency of 0.06 Hz. This attenuates possible noise components
and completely removes any type of baseline wandering. After- As described in Section 4.1 and shown in Figs. 8 and 9, the noise
wards the resulting signal is smoothed through a moving window is not uncorrelated: its power spectral density is not white and
average lter to attenuate any possible high-frequency component the autocorrelation function shows that low value of correlation,
still present. below 0.2, can be obtained only between samples with a distance in
The ToddAndrews peak detection algorithm [18] is then time of at least 1 s. In this scenario, simple averaging of consecutive
applied to identify the occurrences of peaks and troughs in the sig- samples results in a small decrease in noise power. On the other
nal. Each sequence of trough-peak-trough detected is considered a hand, ensemble averaging returns the average between samples
S. Rossi et al. / Sensors and Actuators A 232 (2015) 359367 363
which are far enough to present a weak correlation, resulting in a of the current source between 164 k and 257 k (mean 213 k,
stronger decrease in noise power. standard deviation 43 k).
The averaged period is then processed by applying the deriva- To verify the gain behaviour a variable load was connected to
tive and identifying its maximum as well as the points commonly the sensing terminals of the circuit. The current was forced into the
referred as B and X [17,19] which corresponds to the rst deriva- variable load through two 330 resistors (i.e. each sensing termi-
tive zero crossing and the rst local minimum after the derivative nal was also connected to one injecting terminal through a 330
maximum. The stroke volume (SV) is calculated using the formula resistor). The variable load was changed between 0 and 120
described in [21,19] and cardiac output (CO) is computed as and was measured, the procedure was repeated on ve samples.
The gain resulted between 81.2 and 85.6 (mean value 83.1, standard
CO = SV HR deviation 1.83), with a linearity error of 0.1 % or lower on a full scale
where HR is the heart rate. of 120 .
Noise is another important parameter, as remarked in Section
3.2. Since electrical noise is a stochastic process, it is interesting
4. Experimental results
to characterize it through representative statistics, like autocorre-
lation function and power spectral density. For this purpose, the
For the evaluation of the designed bioimpedance module two
noise was measured on ten acquisitions, each ten minutes long.
complementary paths were followed. First of all, to evaluate
The sensing input terminals were both short-circuited to the ref-
the general functioning and to quantitatively asses the per-
erence, which is available on one of the injection terminals (Fig. 3).
formances, the module was characterized measuring relevant
The input referred power spectral density is reported in Fig. 8,
electrical parameters. Then, in order to demonstrate its suitabil-
while the time autocorrelation of one of the noise acquisitions is
ity in real applications, it was used to measure breathing rate and
shown in Fig. 9. The input referred noise resulted about 5 Vrms
amplitude and cardiac output. Below the procedures followed for
corresponding to 0.05 rms .
each one of these tests are reported, as well as the obtained results.
An example of an acquired raw signal is reported in Fig. 7.
Fig. 7. Raw Z impedance signal. Fig. 9. Autocorrelation function of the input referred voltage noise.
364 S. Rossi et al. / Sensors and Actuators A 232 (2015) 359367
5. Discussion
Acknowledgements
to develop smart solution in biorobotics, telecomunication, managment, smart city and bio-application. Paolo Dario has been and is working on theoretical models,
eld. design methods, precision technologies (including micro, nano and bio) for fabri-
cation of novel components and integrated systems, machines and robots, and on
Dr. Monica Vatteroni is PostDoc at the Biorobotics Institute (BRI) since 2008. She the application of such systems to different elds, including (micro) endoscopic and
is responsible for the research and development of image sensors, vision systems endoluminal surgery, rehabilitation and assistance to disabled and elderly people,
and sensorized platforms mainly focused on biomedical applications. She manages and humanoid robotics. Paolo Dario has contributed to the foundation of such new
a team for the design, production and testing of miniaturized vision systems, CMOS elds as micro and biomechatronics, and to the elds of biorobotics and neuro-
imagers and system level platforms for biomedical parameters monitoring within robotics, frontier research areas whose ultimate goals are to integrate biology and
several European and Regional founded projects. Within the BRI Monica Vatteroni neuroscience with robotics, and to educate the kernel of a new generation of engi-
has been involved in many research projects as operative responsible of technical neers, able to combine solid engineering capabilities, with curiosity and imagination,
activities with an ofcial role in the technical management board in one case. She has and with the ability to incorporate in their engineering skills interdisciplinary and
been scientic responsible for 3 regional founded projects. Her research interests transdisciplinary knowledge.
are in the eld of robotic sensitized platforms and imaging technologies, with a main
focus on vision systems and VLSI technology for CMOS camera design. She is author
and co-author of over 30 peer reviewed journal and conference publications and Luigi Della Torre was born in Monza, Italy in 1963 and started its activities in 1986
holder of 6 patents. She is a regular invited reviewer for high impact journals and working in SGS (now STMicroelectronics) within the group that dealt Video, specif-
conferences in the elds of biomedical engineering, sensors, electronics, imaging ically focusing on components for signal processing and audio/video interfaces. In
and vision systems. 1994 he started to work on Set Top Box architectures, covering all aspects of the
product from the development to the production, in particular the denition of the
Prof. Paolo Dario is professor of Biomedical Robotics at SSSA and the Director of working platform for STB ITALTEL. In 1998 he became a member of the system
The BioRobotics Institute, leading a large team of more than 100 researchers. Today, research team in STMicroelectronics, Advanced System Technology (AST), dealing
Paolo Dario is the Coordinator of the PhD Program in BioRobotics at SSSA, currently with multistandard decoder and digital receivers for analog TV. In 2000 he col-
enrolling 92 PhD students. Paolo Dario is the Founding Coordinator of the Center of laborated on the development of on chip architectures for a receiver for UMTS
Micro-BioRobotics@SSSA of the Italian Institute of Technology (IIT), and he is cur- Basestations and in 2002 he started to work on channel coding and source for video
rently Senior Scientist of IIT. He has been and is the coordinator of more than 60 streaming. In 2004 he took the responsibility of the group that deals with Con-
European projects, of more than 50 national projects, and of more than 50 indus- nectivity, for the development of communications solutions for High-Throughput
trial projects. Paolo Dario is the editor of two books on the subject of robotics, and WLAN-based MIMO and UWB MBOA. Since 2008 he has been responsible of the AST
the author of more than 500 scientic papers (250 on ISI journals). Paolo Darios main Remote Monitoring Group, leading the research activity about wearable systems for
eld of interest is biorobotics, a frontier research area that combines bio-inspiration remote monitoring of physiological parameters.