IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO.

3, JUNE 2013 363

A Sub-mW Fully-Integrated
Pulse Oximeter Front-End
Konstantinos N. Glaros, Student Member, IEEE, and Emmanuel M. Drakakis, Member, IEEE

Abstract—This paper presents the implementation of the first a system for recording continuously and possibly transmitting a
fully integrated pulse oximeter front-end with a power consump- multitude of physiological signals, since that further constrains
tion lower than 1 mW. This is enabled by system- and block-level the power budget. Such systems are currently under develop-
noise optimisation, also detailed in the manuscript. The proposed
design features an analogue feedback loop that enables fast and
ment in both industry and academia [2]–[10] and can signifi-
accurate regulation of the detected photocurrent level and a se- cantly improve the quality of provided care outside a clinical en-
rial-to-parallel interface allowing for extensive programmability vironment. For example, continuous regulation of oxygen pro-
of several operation parameters. The front-end was fabricated in vision to patients with chronic obstructive pulmonary disease
the AMS 0.35 technology and occupies an area of 1.35 . (COPD) has been shown to reduce time spent under hypoxia
Extensive measured results, both electrical and physiological from [11], [12] and overnight pulse oximetry can be employed for
human subjects are reported, demonstrating an estimated SNR of
39 dB and ability to detect 2% changes in , similar to com-
cheaper and more widespread diagnosis and monitoring of sleep
mercial pulse oximeters. This is despite the constrained power con- apnoea [13].
sumption which amounts to 0.31 mW for the LEDs and 0.53 mW Similar power limitations are encountered in the fabrication
for the rest of the front-end from a 3.3 V supply. Statistical results of portable multi-wavelength POs. These are attracting an in-
from 20 chips verify good matching across the Red and Infrared creasing amount of interest, as a way of obtaining both more
channels of the front-end and the accurate operation of the pro- accurate data, as well as concentration measurements of
posed analogue feedback loop.
physiologically relevant molecules other than oxyhaemoglobin
Index Terms—Biomedical electronics, biomedical monitoring, [14], [15].
pulse oximetry, . To reduce the power consumption the PO LED/photodiode
sensing probe is undoubtedly the first candidate for optimisa-
I. INTRODUCTION tion. Several works have looked into optimising its design or its
topology [16]–[19]. On the contrary, even though there are nu-

P ULSE oximetry determines non-invasively blood oxygen

saturation levels by exploiting the light absorption proper-
ties of oxygenated and deoxygenated haemoglobin. In the most
merous patents on POs ([20] provides an extensive list), there
are very few published works concentrating on optimising the
oximeter front-end for power consumption.
common setup, this is achieved by shining light from LEDs with Rhee et al. [16] attempted to minimize the power consump-
two different wavelengths onto a highly perfused tissue, typi- tion of an oximeter at system level, by optimising the clock
cally on a finger or the earlobe, and detecting the transmitted speed and the data transmission scheme employed, whereas
or backscattered light by means of a photodiode. The measured Sueppel et al. [21] opted for switching off parts of the front-end
light signals can be used to determine both heart rate and oxy- when not in use. Neither of them reports on data or
genation levels, then termed . considers the SNR of the detected signal.
The simplicity of a basic pulse oximeter (PO) coupled with Aguillar-Pelaez et al. [22] looked into the power trade-offs
increasing demand has resulted in an abundance of POs in the involved in choosing system parameters, like bandwidth and
market. Nevertheless, commercial POs use peak currents of duty cycle of the LED current. This work was later extended
tens of mA to drive their LEDs in order to obtain an adequate by a more in-depth investigation by the authors [23]. Rantala et
signal at the detector. Even though this current is modulated by al. [24] considered the effect of the system sampling frequency
a square wave with a low duty cycle (typically 5–25%), this and LED duty cycles on the signal SNR. However, neither of the
still leads to average power consumptions of a few mW in the two considers a specific PO implementation or reports measured
best case, to tens of mW in the majority of the cases [1]. results.
This is not a problem for a bedside instrument, but can sig- The most thorough and systematic investigation of PO
nificantly limit operation time in a battery-operated, portable front-end design to date has been the one by Tavakoli [25],
oximeter. The problem intensifies if the oximeter is to be part of [26]. By ingenuously redesigning the front-end transimpedance
amplifier (TIA), to allow for operation under low duty cycles,
Manuscript received October 10, 2011; revised March 07, 2012; accepted and using a full-analogue implementation, a low average power
May 03, 2012. Date of publication July 03, 2012; date of current version May consumption of 4.3 mW was achieved, without any modifica-
22, 2013. This research was supported by Toumaz Technology Ltd. This paper
was recommended by Associate Editor R. Sarpeshkar. tions to the sensing probe.
The authors are with the Department of Bioengineering, Imperial Col- However, the fully-analogue implementation of [25] comes
lege London, London SW7 2AZ, U.K. (e-mail: glarosk@gmail.com; with shortcomings. The low duty cycles and the low gain of
e.drakakis@imperial.ac.uk).
Color versions of one or more of the figures in this paper are available online
the logarithmic TIA employed limit the power of the mea-
at http://ieeexplore.ieee.org. sured signal. At the same time, the noise characteristics of the
Digital Object Identifier 10.1109/TBCAS.2012.2200677 front-end lead to an output SNR that might not be high enough

1932-4545/$31.00 © 2012 IEEE

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
364 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

to resolve small changes in [23], [27]. Furthermore, is then derived from through an empirical relation-
matching and accuracy limitations make it difficult to achieve ship determined by ad hoc calibration of each oximeter design.
errors of below 5%, required for an accurate calculation To determine a typical resolution requirement, however, the the-
[23]. oretically derived relationship based on Beer-Lambert
To overcome these problems and still achieve a signifi- law will be used [1]
cantly reduced power consumption, the proposed PO front-end
employs a sampled approach to avoid signal power degra- (2)
dation from reduced duty cycles. An integrate-and-hold
block is employed to improve noise performance by limiting Then, the relative sensitivity of with respect to can
sub-Nyquist noise sampling and to provide an easily tunable, be determined
high transimpedance on-chip, which is difficult to achieve
with conventional resistive-feedback TIAs. Furthermore, a (3)
fully-analogue control loop allows fast, frequent and accurate
tuning of the detected photocurrent level, to achieve a desired
SNR irrespective of variations at the sensing side due, for From (1) and (2) it can be shown that increases with
example, to differences between patients’ fingers. decreasing oxygen saturation values.
The result is a PO front-end which achieves a power con- The (relative) error in is related to errors in the measure-
sumption of less than 1 mW, amongst the lowest reported to ment of , as follows:
date, while still exhibiting an SNR of 39 dB or more at its
output. This consumption does not include power for clock gen-
eration, A/D conversion and signal processing, which are per- (4)
formed off-chip. However, such capabilities can be migrated
on-chip in future implementations without excessive increase in This in turn relates to errors in the measured AC component
power consumption. Moreover, the proposed design is the first
that can be fully integrated, allowing for easy integration in a
monolithic physiological monitoring platform. Various param- (5)
eters of the system, including gain, current levels, duty cycles,
operation frequency and SNR at the detector can be externally since typically .
controlled via a simple digital programming interface, allowing The error in the measurement is due to noise in the front-end
it to be tailored for varying application requirements. and is the same for both the and channels
Section II of the manuscript discusses the required appli-
cation resolution. In Section III the proposed front-end is (6)
described followed by analysis of noise considerations in
Section IV. Section V details extensive measured results and Using (4)–(6) the relative error in is
Section VI discusses comparison to state-of-the-art and con-
cludes.
(7)

Minimizing the error across various saturation levels, while

at the same minimizing the LED power requires knowledge of
II. OPERATION PRINCIPLE AND RESOLUTION the precise relations between SNR, and power, which
are implementation dependent. To proceed it will be assumed
In the most common PO setup light from a red and an in- that the detected photocurrents have the same DC levels, i.e.,
frared LED is shone onto the finger tip. The transmitted or , the choice adopted in the proposed design.
backscattered light is measured via a photodiode. Due to varia- Under this assumption , and (4) can
tions in the optical properties of the arterial blood between sys- be reformulated as
tole and diastole the detected light has a time-varying, pulsatile
component. Other tissue constituents, which do not vary with (8)
time (e.g., venous blood, bones, skin tissue), lead to a non-time-
varying component in the detected light. The pulsatile compo- The error in the measurement due to noise is then
nent (AC) has an amplitude typically 1%–4% of the quasi-con-
stant (DC) component [1]. The detected light signal for each (9)
channel is termed a photoplethysmographic (PPG) signal and
can be measured by determining the “ratio-of-ratios” Commercial POs commonly specify this error to be smaller
for the and PPG signals than 2 units of for an range 70–100%. Hence, the
minimum required SNR for the channel can be derived as

(1)
(10)

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 365

Fig. 1. (a) Blocks shared between forward and feedback paths and both channels: transimpedance amplifier , switched-integrator (SI) and voltage reference
[REF—Fig. 1(d)]. (b) Forward path: sample and hold (S/H), buffer, capacitive-feedback-amplifier (CFA) and multiplexed output. (c) Feedback path: low-leakage
sample-and-hold and LED driver. (d) Reference voltage generation. The encircled annotated voltages indicate connection points between the various blocks. The
photodiode capacitance, , is shown explicitly.

Relation (10) makes no assumptions as to the form of the the feedback path, employed to control the LED drive current
relationship, and thus is valid in all cases, provided in order to achieve a desired average photocurrent level at the
that . If (2) is used, the right hand side of (10) detector.
assumes a maximum value of 26.73 for , leading
to a minimum required value of 28.5 dB. Note that, A. Forward Path
since , satisfying the SNR requirement The first block is a conventional transimpedance amplifier
for one photocurrent also ensures an adequate SNR for the with resistive feedback [ —Fig. 1(a)].
measurement of the other . Its purpose is to lower the input impedance presented in parallel
In practice, it is often desirable to specify a PO’s SNR at to the large output capacitance of the photodiode, , in order
higher oxygen saturations, since it is hard to perform calibra- to allow for fast settling of the measured photocurrent and ac-
tion measurements on humans at such low saturation levels. As commodate low LED duty cycles [30]. The amplifier’s positive
the Beer-Lambert law based derivation of (2) predicts that as input is used to supply a reverse-bias voltage for the photodiode.
decreases down to 70% increases and The amplifier’s wideband thermal noise can become a
only slightly decreases, one might be tempted to simply specify limiting factor for the SNR performance of the system (see
the minimum of (10) for . However, if Section IV-A). Therefore, it is implemented as a simple
is measured, as is common, at the extremities, this is not one-stage OTA with a large tail current, which is switched
adequate. synchronously to the LEDs to reduce the power consumption.
Low saturation levels may lead to constriction of the periph- Acting complementary to the TIA is a voltage mode
eral blood vessels to maintain the provision of oxygen to the switched-integrator [SI—Fig. 1(a)]. This block serves multiple
vital central organs [28]. This leads to a significant decrease in purposes and provides the proposed PO system with many of its
the pulsatile component measured by the PO [29]. However, the appealing features. It acts as a variable-gain amplification stage
authors could not find any published data quantifying this reduc- that enables implementation of large on-chip gains exceeding
tion. As will be discussed later in detail, our measurements on 50 which can be easily tuned via the digitally-controlled
a limited number of subjects indicate a maximum reduction of integration time. Moreover, it realises a high-pass filter transfer
2–3 times for . Hence, a margin of 10 dB will be characteristic by having its positive input set to a DC level
assumed adequate in this work, leading to a requirement for a equal to that of the output of the first stage— . This
minimum SNR of 39 dB under normal oxygen saturations. is achieved by controlling the LED current via the feedback
path, as will be described in Section III-B. In this way, large AC
III. PROPOSED PO FRONT-END gains can be employed, without the large DC levels saturating
the detector. Finally, and most crucially, the low-pass filter
Schematic diagrams of the proposed PO front-end are shown characteristic stemming from the integrate-and-hold operation
in Fig. 1. Two paths can be identified in the system: the for- of the SI acts to limit the wide noise bandwidth of the .
ward path, dedicated to measuring the photodiode current and Thus, it limits the noise aliasing occurring due to the sampling

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
366 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

following the SI [Fig. 1(b)], improving the SNR. Switched-in- detection low-pass filters. At the same time, the lack of the av-
tegrators have been employed before in POs(e.g., [4], [31]), but eraging action provided by the large time constants can be a dis-
not as noise anti-aliasing filters. advantage: controlling DC level in this way is effectively equiv-
The first two blocks are common for both and alent to a point measurement of the DC level. If this varies sig-
photocurrent measurements. This eliminates errors due to mis- nificantly before the next calibration, or if an error occurs during
match. After the SI the two channels are separated [Fig. 1(b)]. the calibration (e.g., due to a motion artefact), then the measure-
A simple sample-and-hold (S/H) samples the measured voltage ments taken before the next calibration are incorrect. However,
. The stored value is buffered and fed to an amplifier with large DC level variations can be detected on-chip to trigger a
capacitive feedback and a pseudo-resistor in the feedback path new calibration. In our current implementation, the DC recali-
(CFA), similar to the one in [32]. bration interval was fixed at five seconds, which proved to be
This amplifier conditions the signal to better fit the input adequate for our measurements.
range of the subsequent A/D converter. Its closed-loop gain can The DC level control loop is essentially an integral con-
be set to 1 or 5, allowing for a further increase in the total am- troller, consisting mainly of the same two first blocks used in
plification if necessary. At the same time, the capacitive cou- the forward path [Fig. 1(c)]. During a calibration phase a LED
pling removes any slowly varying signal components and sets is switched on and the output of the SI is directly connected to
the output DC level to the value supplied at the positive input the gate of a transistor acting as a transconductance to provide
of the amplifier [ —Fig. 1(b)]. To avoid the slow set- the LED current. The loop is designed to have an adequate
tling associated with the very low frequency pole generated by phase margin for all reasonable LED current levels and tissue
the pseudoresistor-capacitor combination, a switch is used to attenuations, and still settle within a few milliseconds. The
reset the stored capacitor charge during the calibration phase, voltage value supplied to the LED driver is then sampled via a
when the input photocurrent DC level is set via the feedback very precise, low-leakage S/H and used till the next calibration
path (see Section III-B). The S/H and the amplifiers are laid out occurs. The procedure then follows for the other LED.
inter-weaved and using common-centroid techniques to mini- As expected, the S/H operation is crucial for the performance
mize mismatch between the two channels. The channel outputs of the control loop. Any charge injection errors can incur pro-
are then multiplexed to one pin to be connected to the A/D con- hibitively large offsets in the measured current DC level, as
verter, which in the current implementation is off-chip. the stored value drives directly a transistor gate. For the same
To maximize layout reuse all the amplifiers in the forward reason, the stored value needs to vary less than a few mV in the
path, except the one in the , are implemented as two-stage interval between two calibrations, which could be a few sec-
Miller OTAs with varying tail currents and compensation capac- onds. These requirements are stringent, but fortunately, can be
itors, to achieve the open-loop gains and gain-bandwidth prod- met using the differential version of the S/H reported in [33].
ucts required in the various blocks. Careful layout with common-centroid and interdigitating tech-
niques is used to improve matching between the S/H and the
LED driver stages for both channels.
B. Feedback Path

In our implementation, instead of measuring both the DC C. Operation, Programmability and Biasing
and AC photocurrent components we opted for a different ap- Fig. 2 shows the main sequence of operation for the proposed
proach: the detected current DC levels are set to a desired value design. There are two main phases: calibration and measure-
(from now on called the reference current), the same for both ment [Fig. 2(a)].
signal channels, via a feedback mechanism at regular intervals In the calibration phase [Fig. 2(b)], a LED is switched on
during the PO operation. This approach has several advantages. and the calibration loop is allowed to settle. The voltage value
Firstly, the detected current level is always ensured to have a driving the LED is sampled by the low-leakage S/H. The SI is
desired value, irrespective of variations at the probing site due, subsequently bypassed and the CFA is reset. Following this, the
for example, to variations in the tissue absorption characteris- LED is switched off and the process is repeated for the other
tics among humans. Secondly, knowledge of the average de- channel.
tected current level allows for the removal of the DC compo- In a measurement cycle [Fig. 2(c)], the is switched on
nent via simple and relatively fast circuitry. This facilitates large and placed in unity gain configuration to settle. The feedback re-
AC gains without saturation problems. Finally, the DC level sistor is then switched back in the feedback path, the SI is reset
does not need to be measured, eliminating the need for extra and the system is allowed time to settle. The photocurrent is in-
circuitry for this purpose. The feedback loop does not include tegrated for a specified time, at the end of which it is sampled
large time constants usually associated with the extraction of an by the S/H at the output of the integrator. A measurement cycle
average level from the low-frequency measured signal. Hence, for the other channel begins after an appropriate delay, so that
the detected DC level can be frequently reset to a desired value, measurements are carried out at the desired programmed sample
adapting to changes in the setup, for example due to displace- rate. Where necessary, small delays of a few microseconds are
ment of the sensor probe. introduced between phases to accommodate settling of the var-
The absence of large time constants in the control loop has ious blocks.
one more potential advantage: if the PO is operated intermit- From the above, it is evident that several finely-controlled
tently, an measurement can be obtained almost directly timing signals are needed. These are provided by a custom-de-
after switching the PO on, without the need for settling of level- signed finite-state-machine operating at an 1 MHz clock, which

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 367

A. Transimpedance Amplifier

The noise behaviour of the has been thoroughly

researched before [30]. There are two main sources of noise:
thermal noise from the feedback resistor, denoted as , and
thermal noise from the OTA termed . The former pre-
vails at lower frequencies and in most applications is considered
dominant. This has also been the assumption in previous PO
designs [26]. However, at higher frequencies, the photodiode
output capacitance generates a zero in the transfer function of
the input-referred thermal noise to the output, leading
to a “noise-peaking” effect that makes the thermal noise
dominant (Fig. 4).
The large output capacitances of the photodiodes usually
employed in PO probes (typically hundreds of pF), combined
with the large gain-bandwidths needed for fast settling of the
measured photocurrent [25], lead to the noise-bandwidth
being larger than its signal bandwidth. Even worse, if the
output is sampled at the typically low frequencies used in POs,
Fig. 2. Front-end operation phases. (a) Top-level operation phases with
samples between calibrations. (b) Calibration phase sequence. (c) Measurement
most of this noise is aliased into the signal frequency band.
phase sequence. is the sampling frequency, set to 100 Hz in the proposed Such problems are well identified and are commonly addressed
implementation, and the LED duty cycle. either by reducing the gain of the “noise-peaking” via a capac-
itance in parallel to the feedback resistor, or by introducing a
low-pass filter after the to equalize the noise and signal
is supplied externally. The duration of all the phases is pro- bandwidths. The former limits the signal bandwidth, leading to
grammable through a serial-to-parallel interface. The same ap- increased settling times, while the latter is limited by chip area,
plies to other parameters of the system, including the integrating if the front-end is to be fully integrated [30]. Intriguingly, to
capacitor (1.9 or 20.7 pF), the gain of the CFA (1 or 5), the the best of the authors’ knowledge, the only publicly available
bias current (100 –3.1 mA) and the desired DC photocurrent sampled PO design employing a noise limiting filter after the
level. TIA is [24], albeit mainly aimed at limiting the external rather
The latter can be controlled with a resolution of approx. than the TIA noise.
30 nA up to almost 2 . To achieve this, a binary pro- In the proposed design two measures have been employed to
grammable 64-level current mirror scales down the current limit the amplifier thermal noise. Firstly, a noise limiting filter
generated by a standard strong-inversion CMOS current ref- is realised by the switched-integrator. It can be shown that the
erence [34]. The derived current is fed through a replica of integration operation with an integration time of leads to a
the input transimpedance amplifier , whose output is transfer function of
connected to the positive input of the SI [Fig. 1(d)]. For both
the reference generator and the attention has been paid (11)
in minimizing the flicker noise, by using large devices, and lim-
iting the noise-bandwidth, by utilizing large load capacitors (10 just before the sampling operation. This is approximately equiv-
pF) at their output stages. The current reference also provides a alent to filtering the input with a 1st order low-pass filter with a
current of 1 used to bias the rest of the system. 3 dB cut-off frequency [35]. The resulting
Finally, two testing buffers, with four multiplexed inputs noise spectrum after the sampling can then be calculated by
each, are also included on the chip to output internal signals to means of the sampling theorem, taking into account any aliasing
pins. The buffers are designed to drive large external capacitors occurring due to sub-Nyquist sampling.
and can be switched off during normal operation to save power. As an additional measure the bias current of the transistors in
the is set to hundreds of , to increase the transconduc-
tance of its input transistors and hence lower the input referred
thermal noise density. To avoid the excessive cost in power con-
IV. NOISE CONSIDERATIONS sumption the is switched on only for a slightly larger
interval than the LEDs. A reset switch [ —Fig. 1(a)] con-
necting the amplifier in unity gain configuration allows for quick
The most important consideration in the design process was transition from the OFF to the ON state. With LED duty cycles
to minimize noise, so that an acceptable SNR can be achieved at in the order of 1%, the average power consumption of the
the front-end output even for low LED duty cycles or currents. is then still only a few . This arrangement leads to significant
It is mainly this design that enables the low power operation. benefits in noise performance and to the best of our knowledge
The most important noise sources are depicted in Fig. 3 and are has never been employed before in a PO system for that pur-
discussed in this section. pose.

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
368 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

Fig. 3. (a) The most important noise sources. (b) The corresponding transfer functions. The noise sources model the thermal noise of the photodiode , the
voltage source used to reverse-bias the photodiode , the resistors of the and the SI , the OTAs of the and
the SI and the voltage reference generator .

noiseless, while the thermal noise of the OTA does not have a
significant effect; no significant “noise peaking” occurs in this
case, as the input capacitance of the integrator is much smaller
than the output capacitance of the photodiode. The noise PSDs
at the output before sampling are

(14)

Fig. 4. Closed-loop and thermal noise transfer functions for the resistive-feed- (15)
back transimpedance amplifier . is the amplifier open-loop gain and
the photodiode output capacitance.
with , the excess noise factor and input transconduc-
tance of the switched-integrator OTA.
The output noise PSD due to the noise sources just Observe that for , as in our proposed design, the
before sampling can be easily calculated1 SI leads to at least doubling the thermal noise introduced by the
’s feedback resistor. However, the gain from reducing the
(12) under-sampling of the wideband thermal noise outweighs
this cost.
(13)
C. Reference
In (13) is the input transconductance of the , Noise from the reference generation circuits, used to program
which in strong inversion is proportional to the square root the desired reference current level, is presented to the positive
of the bias current. The quantity is an excess noise factor, input of the SI and thus contributes to the output noise. If the
lumping all thermal noise sources from the various transistors PSD of this noise is , then the noise PSD at the output of
of the OTA. and are the closed the SI, before sampling, is
loop transfer function and the noise transfer function of the
input-referred noise to the output respectively
(Fig. 4). (16)
From (12) and (13) it can be seen that it is desirable to max-
imize both the integrator’s input resistance and the ’s It is important to notice that, similar to the SI thermal noise
input transconductance. However, choosing also at- (15), even if a very large is employed, the reference noise
tenuates the measured signal power by and is thus is always present at the output.
not beneficial. A sensible choice is . The maximum In the proposed implementation this noise can be attributed
value for the chosen resistance is then limited by chip area con- to a multitude of sources. The most significant are thermal noise
straints, as well as the power supply, since the voltage drop on from the CMOS current reference transistors, the OTA and
due to the DC photocurrent should not saturate the the resistor employed in the TIA of the reference generation
output [30]. [ —Fig. 1(d)]. The flicker noise of the current reference
also contributes to . To reduce this noise, large capac-
B. Switched-Integrator itors (10 pF) are employed at the outputs of both the current
The SI introduces itself some noise in the signal path, mainly reference and the to limit the noise bandwidth. Despite
due to the resistor at its input. The feedback capacitor is this choice, the reference noise still constitutes a large part of
1The OTA employed in
the total output noise in the final design. Further reductions of
leads to a zero in the transfer function for the
thermal noise of . Due to the large bias current of in the proposed this noise would require a low-noise programmable current ref-
design, this resides at high frequencies and is not considered in (12). erence design or a differential form of the proposed front-end.

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 369

D. External Noise Sources

There are two sources of external noise: the voltage source
providing the photodiode reverse bias voltage and the pho-
todiode. The former is applied at the positive inputs of the
and (Fig. 1). Its output noise PSD before sampling can be
calculated to be

(17)

with the noise PSD of the external voltage source,

whereas and are the noise transfer functions
of the , from their positive input to their output.
is identical to the transfer function for the
thermal noise (Fig. 4), whereas is approximately that
of a unity-gain buffer, due to the large output resistance of the
CMOS current reference. For low frequencies the two effec-
tively cancel out. For high frequencies is significantly
attenuated, due to low-pass filtering of the DC voltage . Fig. 5. Photograph of the fabricated PO front-end.
Eventually, is mainly transferred directly to the output
with unity gain. Therefore, a low-noise voltage regulator is TABLE I
used to provide in the current design. Again, elimination MEASURED RESULTS
of this noise would require a differential implementation.
The photodiode noise can be modelled as shot noise due to
the photocurrent and shot noise due to dark current

(18)

V. MEASURED RESULTS

A. Fabrication and Experimental Setup

The proposed design was fabricated in the AMS 0.35
mixed-signal 3.3 V technology. Custom layout was carried out
for all analogue cells (Cadence Virtuoso 5.1.41), whereas the
digital blocks were generated using automatic synthesis and
routing tools (Cadence RC compiler and First Encounter) and
the cells provided by AMS. A microphotograph of the fabri-
cated design is shown in Fig. 5. The total design area, excluding
testing structures is 1.265 (1.149 excluding the bias
generation).
A dedicated PCB has been designed to interface to the in-
tegrated circuit (IC). The board is powered by 4 AA batteries,
and voltage regulators are employed to derive any voltages not
B. Measurement
directly available from the batteries (TPS73150 and LP3878-
ADJ). The ADS1259 14.4 Ksps 24-bit A/D converter is used The parameters used for the measurements are sum-
to digitize the IC outputs, with its internal reference of 2.5 V. marized in the first column of Table I. The output signal is de-
A ZestSC1 FPGA board serves to program the IC and the A/D multiplexed into the and photoplethysmographic (PPG)
via their serial-to-parallel interfaces, provides the IC clock and signals. Each of these is filtered via a 2nd order IIR Butterworth
handles the data transfer between the PCB and a personal com- filter with a cut-off frequency of 10 Hz, to eliminate out-of-band
puter via a USB interface. On the PC side a mixture of C and noise. For each interval between two subsequent calibrations
Matlab code controls the communications. Matlab is also used (5 s duration), the filtered PPG signal is correlated with a pat-
for signal processing and visualization. Transient measurements tern (Fig. 6) and possible peak locations are located by thresh-
with sampling rates faster than 2 kHz have been carried out with olding. The maxima and minima in the vicinity of each location
a 7300A LeCroy oscilloscope. are detected, ignoring peaks close to the calibration times. The
A commercial SPOmedical 6106 reflectance finger probe was AC value of the PPG signal is calculated by subtracting neigh-
used for sensing. bouring maxima and minima, corrected for ramp-like baseline

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
370 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

Fig. 6. Pattern used to detect peaks in PPG signal via correlation.

drifts (see [1], p.135–143). A value is then derived by dividing

the closely spaced AC values for the two PPG signals.
To derive an accurate value a calibration of the system
on human subjects is necessary. This is time consuming and,
more importantly, requires inducing desaturation to human sub-
jects. However, the intention of this work was not to produce a
fully-functional, ready for commercial use PO. Instead our pri-
mary goal was to demonstrate that the proposed front-end can
achieve the desired resolution, that is, that it can detect small
enough changes in , despite the reduced power consump-
tion. Therefore, the relationship from [25],
, has been used instead. Finally, a median
filter using two samples preceding and following each
value was applied to remove any large variations in the derived
signal, that can occur, for example, if a single peak fails
to be detected.
Measurements were performed on 13 subjects, 10 males and
3 females, with ages between 20 and 30 years old. The subjects
were seated with the probe on their left index finger and their
arm resting on a desk. During a 5 min period, the subjects were
asked twice to forcibly exhale and hold their breath for as long
as possible. Relaxation time was allowed before and between
breathholds to allow the measured signal to return to the
level measured under normal breathing. The start and stop times
of each breath holding were noted. Fig. 7 shows typical signals
recorded from one of the subjects.
The levels measured while the subjects were breathing
normally were .A
desaturation event was defined as a drop higher than 2 units in a
subject’s level after breath holding. Under this definition, Fig. 7. Measured signals for one of the subjects. (a) PPG signals for
channels just before and after the start of a breath hold. (b) Calculated
21 desaturation events were detected in 26 breath holding ef- values for channel—A mean value is calculated for every 5 s interval
forts. This number increased to 21/21 events, when breath hold- between calibrations. (c) Calculated signal. The vertical lines annotate
ings with durations smaller than 20 s were excluded. The de- instances where the subject stopped (StopB) or started (StartB) breathing.
tected maximum changes in level were
. More than 50% of these changes were below 5%. Other
characteristics of the signal during induced hypoxia, like These results demonstrate that the proposed front-end is able
a delay between breath holding and the measured change in the to resolve changes in to levels similar to those typically
signal on a finger probe have also been observed (mea- specified for commercial POs, despite the reduced power con-
sured delays between breath holding and sumption. Since no calibration was performed on the prototype
falling by 2%). system, there is little value in comparing the absolute measured

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 371

values to those of a commercial PO. Nevertheless, the measured

values are within the expected range.
The measurements described above have also been used to
estimate the reduction in the detected AC signal during breath-
holding (see Section II). The detected peak-to-peak AC signal
was measured for time intervals between the establishing
a normal level and the initiation of a breath holding, consid-
ered to be intervals of normoxia. The AC signal amplitude was
then also extracted for an interval starting 10 s after the start of
a breath holding and extending up to the end of the desatura-
tion event.2 For the channel the AC signal was measured to
be reduced by times, with the largest
reduction corresponding to a measured value of 84%. A
further point measurement with one subject showed a 2.3 times
reduction for an of 70%.
Finally, a 2 min measurement was carried out with the sub-
Fig. 8. Estimated SNR versus integration time, sampling frequency and refer-
jects breathing normally to estimate the signal levels ence current. The non-varying parameters are set to values of ,
under normoxia. These were measured to be and .
and for the and PPG
signals respectively.
A representative case of how SNR varies with the integration
C. Noise Measurements time is illustrated in Fig. 8. Increasing increases propor-
In order to measure the system noise, a constant photocurrent tionally the low frequency gain while at the same time
input is desirable. For this purpose, a piece of white paper was limits the 3-dB cut-off of the switched-integrator
placed in the oximeter probe to function as a time-invariant transfer function, . The former leads to a proportional in-
reflecting surface. For each desired value of average detected crease in the amplification of the signal and some noise sources.
photocurrent (reference current), it was tested that the mea- The latter reduces the sub-sampling factor for
sured photocurrent was equal to that resulting when a finger was any noise PSD with bandwidth larger than that is
placed in the probe—although the LED current was different filtered by the SI.
due to the different attenuation in the two cases. The system Therefore, the noise sources can be distinguished in:
was placed in a closed box to avoid ambient light interference. a) those not affected by the SI transfer function ( -
Different settings for the reference current , integration (17), -(16)), remaining constant as changes;
time and sampling frequency were specified. For b) those having a bandwidth smaller than , whose
each case, 2 min records of the front-end’s output were obtained output noise after sampling increases proportionally to
via the A/D and the FPGA interface. The modified periodogram ( -(16)), and;
was calculated using the Hanning window. The in-band noise c) those that have a bandwidth larger than
was then derived by integrating the PSD estimate for and are shaped by the SI ( , , ,
frequencies 0.1–10 Hz, the frequency band where the —(12) to (15)). As increases, their
signal resides [1]. PSD increases due to a higher SI low frequency gain
To estimate an indicative SNR, the signal power was cal- , but is sub-sampled by a smaller factor
culated by multiplying the detected DC level with the mean , leading to an increase proportional to
contrast value estimated for the channel in .
Section V-B, 2.8%. This was then multiplied by the gain For low values, the output noise of the bias source
, with . Only the SNR for the was measured to be dominant. Therefore the output noise is
channel was considered, since for , it is lower constant. On the other hand, the measured increases lin-
than that of the under normal values (see Section II). early with , leading to SNR increasing approximately by
Since is calculated via subtraction of two detected signal 20 dB/decade.
values, 3 dB were subtracted to account for noise in both For moderate values ( ,
measured values (the output noise is expected to be white). ), the noise from “wideband” noise sources [category (c)]
Summarizing, the SNR was calculated as starts dominating, resulting in an increase in noise proportional
to . As keeps increasing linearly with , the SNR
(19) increases by 10 dB/decade.
For higher , the “narrowband”
output noise generated by the on-chip reference, , dom-
2The reduced AC signal measurement interval was associated with the start inates. As this increases proportionally with [category(b)],
of the breath holding, rather than the desaturation event, because of noticeable the measured SNR saturates to a constant value.
reduction in the AC signal in these periods, which might be explained by the
vasoconstriction starting before the actual drop in oxygen saturation is detected The variation of the SNR with the sampling frequency fol-
at the finger probe. lows a different trend (Fig. 8). As the plotted curve corresponds

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
372 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

Fig. 9. Estimated iso-SNR variation with and . The

SNR is measured in dB with a contour plotted every 2 dB. For large ,
the on-chip current reference noise leads to slower increase in SNR. Points A
and B exhibit the same SNR but different average detected photocurrent.

to a value of where the sub-sampled noise sources are domi-

nant, the effect of noise aliasing is evident. Increasing reduces
the sub-sampling factor , by which the noise PSD
is multiplied, and hence the integrated output noise increases as
, leading to an SNR increase of approx. 10 dB/decade.
Finally, Fig. 8 also depicts how SNR changes with . The
only noise sources affected by changes in are the photo-
diode and the reference noise. Using (18)–(19) the former can
be shown to be negligible. is derived by down-scaling the Fig. 10. Gain variations across 20 chips. (a) channel gain. (b) Ratio of
current generated by the CMOS current reference. As in- channel gains.
creases, the scaling factor increases, and so does the output noise
. For small this is not dominant and the SNR in- would allow higher SNR to be achieved for lower , and thus
creases by 20 dB/decade. For larger currents, it dominates, lim- reduce further the LED power consumption.
iting the SNR to a constant value. Similarly, a differential implementation or a redesign of the
From the above it can be observed that there are optimal op- current reference would allow exchanging LED power for even
erating regions for the proposed PO. An advantage of the dig- higher SNR by increasing .
ital programmability is that multiparametric measurements can
be performed to determine these conditions. Fig. 9 illustrates D. Matching
graphically measured iso-SNR variations with and .
Determining requires measuring two amplitudes
For example, note that even though both points A and B cor-
[(1)–(2)]. Therefore, matching between channels is crucial for
respond to an estimated SNR of 40 dB, the latter corresponds to
the accurate measurement of the ratio .
a smaller average detected current, which could lead to a smaller
To measure gain variation in the two channels the PO was
LED power consumption.3
programmed for a photocurrent of 110 nA and excited via a
It should also be remarked that the SI employed in the pro-
current source with a 2 Hz sinusoidal output of 5 nA amplitude
posed architecture is effective for moderate integration times
with 110 nA DC offset. Fig. 10 shows the gain variation across
and photodiode current values. For these operation regions, it
20 chips for the gain of the channel and the gain ratio between
limits the noise bandwidth allowing an increase of 10 dB/decade
the two channels.
in SNR with . If the SI was omitted, the output noise would
The matching is excellent, leading to gain ratio variations of
be more heavily sub-sampled leading to smaller SNR. Unfor-
smaller than 0.2%, which is much smaller than the 1% error
tunately, the dominance of at low values, where
required for a 40 dB SNR.
sub-sampling is even more severe, does not allow the same ef-
fect to manifest itself in the measured results. Reducing this
E. Calibration Loop
noise, or adopting a differential implementation to cancel it,
3This
The accuracy of the calibration scheme employed is crucial,
assumes a constant light attenuation by the tissue for all parameter
values. To estimate the LED power consumption, the change in attenuation for as it replaces the measurement of the DC components of the
different parameters should be considered as well [27]. PPG signals (Section III-B). To evaluate this, a white paper was

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 373

for 5 min, without performing any more calibrations. The mea-

sured difference in the stored value after 5 min across 10
chips was corresponding to a drift
rate of . The variation in the LED
current will then be , with the transcon-
ductance of the transistor driving the LED [Fig. 1(c)]. Even for
the maximum possible value of , the
maximum variation in the LED current is less than 1.6% over a
period of 1 min. In practice, will be much smaller
than 38.61, since the transistor driving the LED is not operating
in weak inversion and source degeneration was employed to re-
duce its transconductance [Fig. 1(c)]. Moreover, variations in
the PPG signal will usually mandate a calibration interval much
smaller than 1 min. In the measurements of Section V-B
a calibration interval of 5 s was used.
Fig. 11. Difference in detected voltages at the output between the The above results suggest that the proposed calibration
channels at the first measured sample after a calibration is performed. scheme works well as an alternative to the traditional DC level
measurements in a PO front-end.

placed in the finger probe to provide a constant reflected light F. Power

input to the photodiode. For every chip the calibration was ac- The front-end can be separated into three sections in terms of
tivated 8 times, every 2 seconds, for each of the chan- power consumption: the analogue part, the digital control cir-
nels. The total calibration time for each channel was 17 ms, with cuitry and the LEDs. To measure the power, a separate power
the channel calibration following directly after that of the supply line was used for each section. A bench-top power supply
channel. with a “sense” output enables introducing a resistor in series
Between calibrations the front-end was operating in measure- with the power supply, while still supplying the desired voltage
ment mode [Fig. 2(c)], performing a measurement every 10 ms to the tested circuit. The average voltage difference across the
. The output was measured during the first resistor can then be used to estimate the average supply current.
time each LED was turned on following a calibration. The 8 The power consumption depends on the front-end configura-
measurements obtained for each chip were averaged to provide tion, and mostly the LED duty cycle (directly related to ),
one measurement per chip. The output was chosen for the reference current value and the sampling frequency
these measurements, as its earlier location in the signal acqui- . The power consumption for three configurations is reported
sition chain provides a more accurate estimate of the detected (Table I).
photocurrent than the front-end output. For each of them the desired is specified. The calibra-
The measured voltage level across 20 wafers was tion loop then ensures a suitable LED current to achieve this
for the channel photocurrent at the detector. As a result the LED power con-
and for the channel, with sumption varies with the characteristics of the probed tissue and
the reverse photodiode voltage set at . More sig- will be different for each individual. To obtain an estimate the
nificant is, however, the matching between the detected levels, light attenuation characteristics were measured for the experi-
depicted in Fig. 11. The results show excellent matching be- mental setup described in Section V-B, for each of the 13 vol-
tween channels, with the maximum measured variation corre- unteers and the three configurations. The average measured at-
sponding to of the av- tenuations are shown in Table I. The LED peak current was
erage detected DC signal level. then calculated as and the power consumption as
Another issue that may degrade the performance of the cali- , where is the
bration scheme is the accurate and variation-free storage of the LED duty cycle. The reliability of the average LED current cal-
voltage level supplied to the LED driving circuitry between cal- culation was also verified by comparing the calculated values
ibrations (Section III-B). The performance of the low-leakage to the average LED power supply line current when testing the
differential S/H employed for that purpose is crucial [Fig. 1(c)]. device on one of the authors’ fingers.
The two main limitations are errors in the sampled value due to To estimate the SNR at each of the three configurations, the
charge injection and drift in the stored value due to leakage. To contrast values of the PPG signals were also mea-
alleviate charge injection errors a differential sample and hold sured in the same setup for all subjects and under normoxia con-
topology was used [33], in conjunction with a well matched ditions. The SNR was then calculated as in (19) by substituting
layout. The measurements presented in Fig. 11 confirm that, 0.028 with the mean measured contrast value for each configu-
though there are charge injection errors, these are well matched ration (Table I).
between the two channels. The proposed configuration (A—Table I) can achieve the re-
The drift in the stored voltage of the low-leakage S/H cell was quired SNR of 39 dB with an average power consumption of
measured by carrying out a calibration for one reference current less than 1 mW. If a higher SNR is necessary, this is easily
(350 nA) and then operating the front-end in measurement mode achieved by increasing and/or (Section V-C). An SNR

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
374 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 7, NO. 3, JUNE 2013

of 48.5 dB can be obtained by increasing the LED power by 3.5 TABLE II

times (C—Table I). Note that this configuration still has a total PULSE OXIMETER POWER CONSUMPTION
power consumption below 2 mW. Finally, in some applications
a low peak supply current may be desirable, for example due to
constrains imposed by the battery. In the proposed front-end,
peak current can be easily traded for integration time, main-
taining similar power and SNR levels (B—Table I).
The analogue part of the front-end was measured to draw an
average supply current of 110 while the digital part 50 .
This is estimated to be practically the same for all three configu-
rations, as the clock and sampling frequencies are identical, and
the only front-end current that changes with the varying param-
eters is the average current supplied to the OTA of the
(see Section III-A). This is equal to and is esti-
mated to vary from 2 to 7 .
A breakdown of the power consumption in a simulation
shows that 30% of the static power is dissipated on the CMOS
current reference. While directly reducing its supply current
power consumption, as demonstrated by measured data. It is dif-
would arguably increase its output noise, this could be reduced
ficult to carry out a fair comparison of the reported power with
by employing more noise-efficient designs or even a differential
the multitude of other available implementations. Industrial de-
topology. A further 24% of the static supply current is drawn
signs rarely quote detailed technical specifications and few aca-
by the SI and another 21% by internal unity-gain buffers. These
demic efforts have investigated the power consumption of a PO
can be limited by careful optimisation of the respective blocks,
[16], [22], [25]. Even when power consumption is reported or
without any significant impact on performance. The authors
can be derived from marketing documentation, it is meaningless
believe that a static power consumption of under 100 is a
without the provision of an associated SNR figure, as this paper
realistic target for a future design.
has demonstrated as well. To the best of our knowledge, this
Concerning the digital part, the power can be easily reduced
work is the first to report measured SNR estimations related to
further if a smaller clock frequency is used. This determines the
signals. Finally, the use of different types of sensor probes
minimum time step that can be used in the front-end’s operation.
makes the value of a comparison between just power consump-
That could be increased to 10 , from the current 1 , without
tions questionable.
any performance penalty.
Nevertheless, for the reference of the reader, an estimated
Concluding this section, it should be stressed that all the
power consumption for a few of the lowest power POs has been
above measurements have been carried out under controlled
compiled in Table II. For all the commercial POs the power con-
conditions. In a less controlled environment external interfer-
sumption has been estimated by multiplying the supply current
ences, such as ambient light or motion artefacts, can degrade
extracted from the quoted operation time and indicative battery
the SNR of the measurement further. This problem can be
capacities by 3.3 V. The proposed front-end achieves a LED
mitigated to a certain extent via better sensor design (e.g., [14],
power more than 4 times smaller than all implementations, but
[16]) or via signal processing (e.g., [36], [37]). Such methods
those in [16], [21], [39], where no SNR considerations were
are likely to result in an increase in the power consumption of
reported. Most commercial POs feature a power consumption
the front-end.
more than an order of magnitude higher [25].
Moreover, the reported power is that required only by the
This low power consumption has been achieved in our case
front-end and the LEDs and does not include power for clock
by carefully examining the SNR requirements and trade-offs
generation, A/D conversion or any signal processing following
and proposing ways to tackle the noise bottleneck in the PO
the front-end. If the front-end is used as part of a larger sensor
front-end: a switched-integrator was employed as a noise-lim-
interfacing system (e.g., [2]), such blocks will typically be al-
iting and a high-pass filter, and a switched large bias current
ready available. If not, the design of a 100 kHz clock generator,
was used to lower the TIA thermal noise. The noise is limited
a 200 Sps 12-bit A/D converter and an embedded processor with
to the desired levels despite the fact that the system is fully in-
a low clock frequency or a low-power radio to transmit the data
tegrated, the first ever reported, something which poses further
to a processing node is possible with a power budget of few
restrictions to noise requirements due to the limited available
hundreds of [2], [16], [38]. Alternatively, more power-ef-
capacitor area [23].
ficient, yet imprecise, analogue signal processing solutions like
the one proposed in [26] could be employed. ACKNOWLEDGMENT
The authors would like to thank the reviewers for their com-
VI. CONCLUSIONS ments and suggestions.
In this work a pulse oximeter front-end was presented that REFERENCES
can achieve an adequate SNR to detect changes in levels [1] J. G. Webster, Design of Pulse Oximeters, ser. Medical Science Se-
similar to those specified for a commercial PO with very low ries. Bristol, PA: Philadelphia: Institute of Physics Pub., 1997.

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.
GLAROS AND DRAKAKIS: A SUB-mW FULLY-INTEGRATED PULSE OXIMETER FRONT-END 375

[2] A. C. W. Wong et al., “A 1 V, micropower system-on-chip [26] M. Tavakoli, L. Turicchia, and R. Sarpeshkar, “An ultra-low-power
for vital-sign monitoring in wireless body sensor networks,” in pulse oximeter implemented with an energy-efficient transimpedance
Proc. IEEE Int. Solid-State Circuits Conf., Dig. Tech. Papers, amplifier,” IEEE Trans. Biomed. Circuits Syst., vol. 4, no. 1, pp. 27–38,
2008, pp. 138–602. 2010.
[3] R. Matthews, N. McDonald, P. Hervieux, P. Turner, and M. Steindorf, [27] K. N. Glaros, “Low-Power Pulse Oximetry and Transimpedance Am-
“A wearable physiological sensor suite for unobtrusive monitoring of plifiers,” Ph.D. dissertation, Dept. Bioengineering, Imperial College
physiological and cognitive state,” in Proc. Annu. Int. Conf. IEEE Eng. London, London, U.K., 2012.
Med.-Biol. Soc., Aug. 2007, pp. 5276–5281. [28] U. Leuenberger, J. Hardy, M. Herr, K. Gray, and L. Sinoway, “Hypoxia
[4] E. Jovanov, A. Milenkovic, S. Basham, D. Clark, and D. Kelley, “Re- augments apnea-induced peripheral vasoconstriction in humans,” J.
configurable intelligent sensors for health monitoring: A case study of Appl. Phys., vol. 90, no. 4, pp. 1516–1516, 2001.
pulse oximeter sensor,” in Proc. Annu. Int. Conf. IEEE Eng. Med.-Biol. [29] A. Lima and J. Bakker, “Noninvasive monitoring of peripheral perfu-
Soc., 2004, vol. 2, pp. 4759–4762. sion,” J. Clin. Monit. Comput., vol. 31, no. 10, pp. 1316–1326, 2005.
[5] E. Jovanov, C. Poon, G. Yang, and Y. Zhang, “Guest editorial body [30] J. G. Graeme, Photodiode Amplifiers: Op Amp Solutions. New York:
sensor networks: From theory to emerging applications,” IEEE Trans. McGraw-Hill, 1995.
Inf. Technol. Biomed., vol. 13, no. 6, pp. 859–863, 2009. [31] J. Patterson and G. Yang, “Ratiometric artifact reduction in low power
[6] K. Wac et al., “Mobile patient monitoring: The mobihealth system,” reflective photoplethysmography,” IEEE Trans. Biomed. Circuits Syst.,
in Proc. Annu. Int. Conf. IEEE Eng. Med.-Biol. Soc., Sep. 2009, pp. vol. 5, no. 4, pp. 330–338, 2011.
1238–1241. [32] R. R. Harrison and C. Charles, “A low-power low-noise CMOS am-
[7] J. Yoo, L. Yan, S. Lee, Y. Kim, and H.-J. Yoo, “A 5.2 mW self-config- plifier for neural recording applications,” IEEE J. Solid-State Circuits,
ured wearable body sensor network controller and a 12 wirelessly vol. 38, no. 6, pp. 958–965, 2003.
powered sensor for a continuous health monitoring system,” IEEE J. [33] M. O’Halloran and R. Sarpeshkar, “A 10-nW 12-bit accurate analog
Solid-State Circuits, vol. 45, no. 1, pp. 178–188, 2010. storage cell with 10-aA leakage,” IEEE J. Solid-State Circuits, vol. 39,
[8] P. Pandian et al., “Smart vest: Wearable multi-parameter remote no. 11, pp. 1985–1996, 2004.
physiological monitoring system,” Med. Eng. Phys., vol. 30, no. 4, pp. [34] E. Vittoz and J. Fellrath, “CMOS analog integrated circuits based on
466–477, 2008. weak inversion operations,” IEEE J. Solid-State Circuits, vol. 12, no.
[9] R. Haahr et al., “An electronic patch for wearable health monitoring by 3, pp. 224–231, 1977.
reflectance pulse oximetry,” IEEE Trans. Biomed. Circuits Syst., vol. [35] G. Xu and J. Yuan, “Performance analysis of general charge sampling,”
6, no. 1, pp. 45–53, 2012. IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 52, no. 2, pp. 107–111,
[10] K. Li, S. Warren, and B. Natarajan, “Onboard tagging for real-time 2005.
quality assessment of photoplethysmograms acquired by a wireless re- [36] M. Hayes and P. Smith, “A new method for pulse oximetry possessing
flectance pulse oximeter,” IEEE Trans. Biomed. Circuits Syst., vol. 6, inherent insensitivity to artifact,” IEEE Trans. Biomed. Eng., vol. 48,
no. 1, pp. 54–63, 2012. no. 4, pp. 452–461, 2001.
[11] M. Iobbi, A. Simonds, and R. Dickinson, “Oximetry feedback flow [37] J. Goldman, M. Petterson, R. Kopotic, and S. Barker, “Masimo signal
control simulation for oxygen therapy,” J. Clin. Monitor. Comp., vol. extraction pulse oximetry,” J. Clin. Monit. Comput., vol. 16, no. 7, pp.
21, no. 2, pp. 115–123, 2007. 475–483, 2000.
[12] Z. Zhu, R. K. Barnette, K. M. Fussell, R. M. Rodriguez, A. Canonico, [38] R. Yazicioglu et al., “Ultra-low-power wearable biopotential sensor
and R. W. Light, “Continuous oxygen monitoring-a better way to pre- nodes,” in Proc. Annu. Int. Conf. IEEE Eng. Med.-Biol. Soc., Sep. 2009,
scribe long-term oxygen therapy,” Intens. Care Med., vol. 99, no. 11, pp. 3205–3208.
pp. 1386–1392, 2005. [39] A. Wong, P. Kong-Pang, Z. Yuan-Ting, and K. Hung, “A near-infrared
[13] N. Netzer, A. Eliasson, C. Netzer, and D. Kristo, “Overnight pulse heart rate measurement IC with very low cutoff frequency using current
oximetry for sleep-disordered breathing in adults,” Chest, vol. 120, no. steering technique,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 52,
2, pp. 625–625, 2001. no. 12, pp. 2642–2647, 2005.
[14] M. Aoyagi, T. Fuse, N. Kobayashi, K. Machida, and K. Miyasaka,
“Multiwavelength pulse oximetry: Theory for the future,” Anesth. Konstantinos N. Glaros (S’05) received the
Analg., vol. 105, no. 6, pp. S53–S58, 2007. Diploma in electrical and computer engineering
[15] K. Yamakoshi and Y. Yamakoshi, “Pulse glucometry: A new approach from Aristotle University of Thessaloniki, Greece,
for noninvasive blood glucose measurement using instantaneous differ- in 2006, the M.Sc. degree in analogue and digital
ential near-infrared spectrophotometry,” J. Biomed. Opt., vol. 11, pp. integrated circuit design from Imperial College
054028–054028, 2006. London, U.K., in 2007, both with distinction, and
[16] S. Rhee, B.-H. Yang, and H. H. Asada, “Artifact-resistant power-ef- the Ph.D. degree in analogue IC design from the
ficient design of finger-ring plethysmographic sensors,” IEEE Trans. Department of Bioengineering, Imperial College
Biomed. Eng., vol. 48, no. 7, pp. 795–805, 2001. London, in 2012.
[17] M. Savage, C. Pujary, and Y. Mendelson, “Optimizing power con- His research interests include sensor interfacing
sumption in the design of a wearable wireless telesensor: Comparison and instrumentation circuits, continuous-time filters,
of pulse oximeter modes,” in Proc. Annu. IEEE Bioeng. Conf., 2003, and low power integrated circuit design for biological and biomedical appli-
pp. 150–151. cations.
[18] P. Branche and Y. Mendelson, “Signal quality and power consumption Dr. Glaros has been awarded various performance scholarships by the Foun-
of a new prototype reflectance pulse oximeter sensor,” in Proc. Annu. dations of State Scholarships (IKY). He has also been a scholar of the Bodossaki
IEEE Bioeng. Conf., 2005, pp. 42–43. Foundation, Greece.
[19] R. G. Haahr et al., “A novel photodiode for reflectance pulse oximetry
in low-power applications,” in Proc. Annu. Int. Conf. IEEE Eng. Med.- Emmanuel M. Drakakis (M’05) received the
Biol. Soc., 2007, pp. 2350–2353. B.Sc. degree in physics and the M.Phil. degree
[20] B. Nordstrom, W. Shea, and E. Petersen, “Techniques for detecting in electronic physics and radioelectrology from
heart pulses and reducing power consumption in sensors,” U.S. Patent Aristotle University of Thessaloniki Macedonia,
7 499 740, 2009. Greece, and the Ph.D. degree in analogue IC design
[21] B. Sueppel and D. Mortara, “Low power pulse oximeter,” U.S. Patent from the Department of Electrical and Electronic
6 697 655, 2004. Engineering, Imperial College London, U.K., in
[22] E. A. Pelaez and E. R. Villegas, “Led power reduction trade-offs for 2000 under the supervision of Dr. A. Payne.
ambulatory pulse oximetry,” in Proc. Annu. Int. Conf. IEEE Eng. Med.- He is an academic member of staff in the Depart-
Biol. Soc., 2007, pp. 2296–2299. ment of Bioengineering at Imperial College London
[23] K. N. Glaros and E. M. Drakakis, “Trade-offs for low power integrated where he joined in October 2001. In the Department
pulse oximeters,” in Proc. IEEE Biomed. Circuits Syst. Conf., 2009, of Bioengineering he has founded the Bioinspired VLSI Circuits and Systems
pp. 245–248. Group. He has authored or coauthored a large number of peer-reviewed journal
[24] B. Rantala and A. Backman, “Pulse oximeter,” U.S. Patent 6 912 413, and conference papers and several book chapters.
2005. Dr. Drakakis has received many prizes for research excellence and is involved
[25] M. D. Tavakoli, “An analog VLSI front end for pulse oximetry,” Ph.D. in numerous cross-disciplinary research projects. He has also served as an As-
dissertation, Mass. Inst. Technol., Cambridge, 2006. sociate Editor in several IEEE publications.

Authorized licensed use limited to: CP Chen. Downloaded on August 11,2020 at 06:06:59 UTC from IEEE Xplore. Restrictions apply.