272 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO.

2, APRIL 2015

Adaptive Pulse Width Control and Sampling for Low

Power Pulse Oximetry
Sagar Venkatesh Gubbi and Bharadwaj Amrutur, Senior Member, IEEE

Abstract—Remote sensing of physiological parameters could be time-to-digital converter (TDC) and perform design-space anal-
a cost effective approach to improving health care, and low-power ysis to optimize the circuit parameters. Second, we describe a
sensors are essential for remote sensing because these sensors are technique to estimate SNR of the captured photoplethysmogram
often energy constrained. This paper presents a power optimized
photoplethysmographic sensor interface to sense arterial oxygen and use this information to dynamically adjust the LED on-time.
saturation, a technique to dynamically trade off SNR for power This reduces power consumption by operating the sensor at just
during sensor operation, and a simple algorithm to choose when sufficient SNR. Third, we propose a simple algorithm that re-
to acquire samples in photoplethysmography. A prototype of duces the number of samples required, thereby reducing power
the proposed pulse oximeter built using commercial-off-the-shelf consumption. Finally, we evaluate power benefit of the pro-
(COTS) components is tested on 10 adults. The dynamic adapta-
tion techniques described reduce power consumption considerably posed dynamic adaptation techniques.
compared to our reference implementation, and our approach is There have been numerous efforts in the past to reduce power
competitive to state-of-the-art implementations. The techniques consumption in pulse oximetry. A photodiode structure that
presented in this paper may be applied to low-power sensor lowers power consumption in the photodiode interface circuit is
interface designs where acquiring samples is expensive in terms of proposed in [15] and a method to optimally size the photodiode
power as epitomized by pulse oximetry.
is described in [16]. Our work, however, is about designing a
Index Terms—Design-space analysis, low-power, pulse oximetry, low-power interface circuit for a given photodiode. The authors
SpO2, tracking-loop. in [17], [18] argue that low power analog circuits are suitable
for sensing physiological parameters. A recently proposed
photodiode interface circuit targeted to pulse oximetry uses
energy efficient transimpedance amplifiers and is a completely
I. INTRODUCTION analog design [19]. In contrast, the photodiode interface we
propose is more digital in nature and employs a time-to-digital

I NFANT mortality in rural India is substantially higher than converter (TDC). A switched integrator has been used in [20]
in the developed world [1]. The Remote Neonatal Intensive to reduce noise. Our work complements this design to further
Care Unit (RNICU) Project aims to provide affordable health reduce power consumption. Although our implementation is
care in rural areas by monitoring various physiological param- similar, it differs in subtle ways. The idea of dynamically
eters of neonates, remotely [2]. To remotely monitor physio- reducing power consumption based on the characteristics of
logical parameters, it is desirable to have sensor nodes that are the acquired plethysmogram has been claimed in [21]. We
small and discreet, and the battery powering those nodes would propose a way of accomplishing this and measure the power
therefore be of limited capacity. To reduce the frequency of bat- conserved. In [22], compressive sensing is used to minimize
tery replacement, a sensor node that consumes as little power as power consumption by reducing the number of samples ac-
possible has to be designed. quired. However, existing photodiode interface circuits use
Arterial oxygen saturation is an important physiological pa- analog band pass filters that do not easily permit sub-Nyquist
rameter in modern medicine [3]. Pulse oximeters are commonly sparse random sampling, which is necessary for compressive
used in intensive care units, sleep studies, neonatal care etc. sensing to effectively reduce power. Our work addresses this
[3]–[6]. Low-power pulse oximeters are useful for remotely issue, and the proposed architecture can acquire samples at
monitoring heart rate and oxygen saturation [7]–[14]. arbitrary time instants. A trade-off between noise performance
In this paper, we present a novel low-power pulse oximeter. required of the photodetector circuit and the duration for which
First, we propose a photocurrent detector architecture that uses a the LED is turned on for each measurement is shown in [23],
[24]. We extend this by proposing a power model for opamps
in terms of their noise performance. Using this, we optimize
Manuscript received October 04, 2013; revised January 15, 2014 and April
10, 2014; accepted May 12, 2014. Date of publication June 30, 2014; date of the trade-off between burning power in the LED versus the
current version April 29, 2015. This work was supported by DeitY, MCIT, Gov- front-end amplifier.
ernment of India and the Robert Bosch Centre for Cyber-Physical Systems. This
The rest of this paper is organized as follows. First, we briefly
paper was recommended by Associate Editor A. Demosthenous.
The authors are with the Department of Electrical and Communication review the principle behind pulse oximetry. We then present the
Engineering, Indian Institute of Science, Bangalore, Karnataka 560012, India proposed photodetector architecture. Next, we describe how it
(e-mail: [email protected]; [email protected]).
is optimized for power. Section 4 presents a technique that dy-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. namically lowers the duration for which the LED is turned on,
Digital Object Identifier 10.1109/TBCAS.2014.2326712 and proposes a simple algorithm that further decreases power

1932-4545 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 273

Fig. 1. Photodetector architecture. is the parasitic capacitance across the photodiode.

by reducing the number of samples that are acquired. We then

evaluate the performance of the proposed pulse oximeter in Sec-
tion 5. Finally, Section 6 concludes the paper.

II. BACKGROUND
Oxygen saturation is the percentage of haemoglobin in the
Fig. 2. The red and IR LEDs are alternately flashed once every 5 ms, and for
blood that is oxygenated. If is the concentration of de-oxy- each photocurrent measurement the LED is kept on for a duration whose
genated haemoglobin in the blood and is the concentra- precise value depends on the photocurrent at that time.
tion of oxygenated haemoglobin, then

III. PHOTODIODE INTERFACE CIRCUIT

(1)

A. Proposed Architecture
Pulse oximetry is an optical technique to measure arterial
oxygen saturation which takes advantage of volumetric pulsa- The receiver circuit we propose is shown in Fig. 1. To acquire
tions in the arteries as the heart pumps blood. Light from an a sample, the analog circuitry and the TDC are powered up, one
LED is transmitted through a body part, usually a finger or an of the LEDs is turned ON, and the switch S1 is opened. The
ear lobe, and the intensity of the light that passes through to the first stage is a current integrator which converts the photocur-
other side is measured using a photodiode. From [3], the nor- rent to a voltage ramp. The amplified ramp is then thresholded
malized absorption ratio R is calculated as against two voltage references. Finally, a time-to-digital con-
verter (TDC) measures the difference between the times when
the amplified ramp hits the two reference voltages. After the
(2) measurement is recorded, the LED, the entire analog circuitry,
and the TDC are powered down to save power. Fig. 2 shows this
where the perfusion is the ratio of AC to DC of the corre- process and Fig. 3 illustrates the output we expect from the TDC
sponding plethysmogram for red and infra-red wavelengths for one of the channels (red or IR). It is a discrete time wave-
form with samples separated by 10 ms. To acquire each sample,
the switched integrator in Fig. 1 accumulates photocurrent for
(3) a duration that typically lasts a few hundred microseconds, and
the measurement made by the TDC is recorded. As illustrated in
Fig. 3, the recorded plethysmogram has a large DC component
Here, is the received photocurrent. Substituting absorp- and a small AC component .
tion coefficient values from [3] gives Suppose that the circuit starts to acquire a sample at .
The output of the first stage is

% (4)
(6)
In practice however, is assumed to be a linear function
of as in
where is the photocurrent. This is amplified by the second
stage to give
(5)

where and are constants obtained by calibration. (7)

Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
274 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO. 2, APRIL 2015

Fig. 4. Noise in the ramp is reflected as jitter in the output of the TDC.

(but the two reference voltages are held fixed). This allows
a control algorithm to analyze the acquired photoplethysmo-
gram and adjust the gain, A, to dynamically trade-off SNR
for measurement duration [(11) and (8)]. This reduces the
power consumption since measurement duration determines
the percentage of time for which the LED and the circuitry are
Fig. 3. Illustration of the output of the TDC. The samples in this discrete powered.
time waveform are separated by 10 ms. The heart rate is exaggerated in this
illustration. In the proposed circuit, the amplified ramp is thresholded
against two references, and the difference between the times
when the ramp crosses these two references is measured. This
where A is the gain of the second stage. After thresholding, the suppresses the effect of 1/f noise in the integrator stage, which
TDC measures can be severe because of the very low signal frequency involved
(0.5–6 Hz). This is similar to correlated double sampling [26].
However, only the 1/f noise from the amplifiers (and not due
(8) to the comparator) is suppressed. Comparator 1/f noise has a
smaller effect on SNR because the signal would have under-
where . and are the instants when gone substantial amplification before reaching the comparator.
crosses and respectively. Thus, is inversely pro- Furthermore, since the power supply to the entire analog circuit
portional to the photocurrent [25]. It can be shown that (Ap- is duty cycled, 1/f noise gets further suppressed [27]. Moreover,
pendix A) the perfusion is TDC measurements at the peak and trough are subtracted to ob-
tain , which suppresses comparator 1/f noise below the
heart rate.
(9) Another aspect of the proposed circuit is that the LED cur-
rent is fixed so as to enable the LEDs to always operate at their
Even if the photocurrent is a constant, the output of the highest efficiency (luminosity per unit current). This is in con-
TDC will exhibit variance due to noise. This is illustrated in trast to implementations such as [19], [28], and [20], that con-
Fig. 4. To quantify the effect of noise, we define signal-to-noise trol the current so that the DC value of the red and IR photo-
ratio (SNR) of the acquired plethysmogram as follows [23]. plethysmogram are equal to some pre-determined value. By op-
erating each LED at its maximum efficiency point and control-
ling only the duty cycle, we not only save power but also elim-
(10)
inate one more amplifier that would have otherwise been used
to control the LED current. Finally, the proposed circuit uses a
where is the standard deviation of the jitter in the output of
time-to-digital converter (TDC) as opposed to an ADC. This is
the TDC due to noise. It can be shown (Appendix B) that
because it is relatively easier to obtain large dynamic range in
TDCs in contrast to ADCs, particularly as the supply voltage
shrinks.
To deal with low perfusion, current oximeter implementa-
tions use AC coupled amplifiers to separate the large DC level
(11) from the AC signal. Only the amplified AC signal is fed to the
ADC to make the best use of its dynamic range [20], [28]. How-
where is the noise gain of the first stage, ever, the use of AC coupled amplifiers and low pass filters in
is the input referred noise spectral density of the opamp the pipeline, as illustrated in Fig. 5, make the process of turning
used for the switched integrator, is its ON the LED and sparsely sampling the photocurrent at arbi-
noise bandwidth and its unity gain bandwidth, and trary time instants harder to implement. This is because (a) low
is the photocurrent noise spectral density. pass filters do not allow the plethysmogram to change rapidly
We employ a switched integrator front-end because it gives as would happen if the LED is turned ON at arbitrary time in-
a simple way to control the duration of photocurrent measure- stants, and (b) it is difficult to build a sample-and-hold circuit
ment. The gain of the second stage can be digitally controlled that droops voltage by no more than a few hundred microvolts
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 275

Fig. 5. Photocurrent detector based on transimpedance amplifier.

over the duration when samples are not taken (on the order of
one second). Such a sample-and-hold circuit will need a large Fig. 6. Illustration of the non-ideal switched integrator output.
capacitor as well as additional circuitry to isolate it from other
amplifiers whose power supply is being duty cycled. These are
undesirable in both integrated and discrete implementations. the first stage is considered, and the second stage is assumed to
TDCs solve this problem by offering very large dynamic have unity gain. We also assume that and are fixed and
range, thereby permitting the photocurrent to be digitized explain later how we choose and .
directly without having to separate the AC signal from the DC. 1) Gain: If the desired average duration for turning on the
For example, a 22-bit TDC with a resolution of 110 ps was LED to acquire a sample is , the integration capacitor,
demonstrated in [29]. With such a large dynamic range, even , is set to be
under low perfusion, quantization noise is sufficiently small.
Furthermore, because the samples are separated by a relatively
large interval of 5 ms, some TDC architectures such as Vernier (12)
TDCs may comfortably take more time than the sampling
duration to resolve the time difference. This sets the noise gain at . At this point,
On the flip side, the noise bandwidth in Fig. 1 is the full circuit the value of at which the total power consumption will
bandwidth, whereas the noise bandwidth in Fig. 5 is limited to be minimized is not known. This will be the subject of a subse-
the signal bandwidth,1 which is usually about 6 Hz. The excess quent section.
noise bandwidth is the price for the ability to sample at arbitrary 2) Bandwidth: The output of the switched integrator, in
time instants and to power down the entire circuit between sam- Fig. 1, will not be a perfect ramp because the bandwidth of the
ples. opamp is limited (see Fig. 6). We choose the minimum band-
Having decided on the architecture, we proceed to select the width that keeps non-linearity below the acceptable level. In
parameters of the amplifiers in Fig. 1. In selecting the gain, order to do this, it can be shown that the opamp bandwidth has
bandwidth, and noise performance of the amplifier, a trade-off to satisfy
presents itself. The higher the gain-bandwidth and lower the
input referred noise level of the amplifier, the larger will be its
bias current and lower will be the necessary measurement dura- (13)
tion to achieve a target SNR leading to smaller power burnt in
the LED (because it can be shut off sooner). Conversely, using
an amplifier with smaller gain-bandwidth and poorer noise per- where is the minimum SNR necessary to make suffi-
formance will lead to lower power burnt in the amplifier. How- ciently accurate calculations [23], is
ever, the duration of measurement will now have to be longer to the noise bandwidth, and is the unity gain bandwidth. This
achieve the desired SNR, which means that more power is burnt equation has to be solved iteratively to find the minimum accept-
in powering the LED while the photocurrent is being measured. able bandwidth for a given . Equation (13) indicates that
To optimally select the parameters of the amplifier, we develop as is increased, the required bandwidth reduces.
a power model for opamps in terms of bandwidth and noise per- 3) Input Referred Noise: Referring the jitter in the output of
formance. We then perform design-space analysis. the TDC back to the input of the opamp [30], the max-
imum permissible input referred noise spectral density of the
B. Interface Circuit Design opamp (above which the SNR of the acquired plethysmogram
will become lower than ) can be shown to be
The photocurrent , perfusion , photodiode capac-
itance , and photocurrent noise spectral density are
known. We design the circuit parameters , and , and
also the gain, bandwidth, and input referred noise levels of
the opamps in Fig. 1. These parameters are chosen so that the
(14)
total power consumption of the LED and the interface circuit
is minimized. For the following analysis, only the noise from Equation (14) (derived in Appendix B) shows that as
1High frequency noise in the circuit before the sample-and-hold (S/H) stage
is increased, the input referred noise requirement gets relaxed.
aliases into the signal band. This results in higher noise than if the LED was 4) Reference Voltages: It is desirable to choose and
kept continuously ON without the S/H stage. that are as high as possible within the input common mode range
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
276 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO. 2, APRIL 2015

of the comparator so that the noise performance of the com-

parator is optimized. If and are close to each other,
becomes small, and from (14), the permissible input referred
noise of the opamp reduces, which in turn causes an increase in
the opamp power consumption. On the other hand, if they are far
apart, from (13), the bandwidth required increases, which also
causes an increase in the power consumption of the opamp. This
may be intuitively understood from Fig. 6 which illustrates the
output of the switched integrator. If is small, the output is
closer to a straight line between and . On the other hand,
if is large, a larger portion of non-linearity gets captured. As
a compromise, we choose V and V. Small
variations in the reference voltages due to comparator offset or
reference generator offset are insignificant because we use only
the ratio of readings and the voltage reference terms get can- Fig. 7. Scatterplot of supply current versus bandwidth on a log-log scale for
celed. discrete opamps. .

C. Op-Amp Power Model

We would like to estimate the power consumption of an
opamp, given its gain, bandwidth, and input referred noise
level. To do this, we assume that the opamp is constructed
by cascading a number of simple amplifiers, each having a
gain of 10. Consequently, the required number of stages is
, where is the required opamp gain.
So, the current consumption of the opamp is

(15)

where is the current per stage. The bandwidth is approxi-

mated [31] by

Fig. 8. Scatterplot of supply current versus bandwidth per input referred noise
(16) for discrete opamps. One outlier was ignored during fitting and is not shown in
this plot.

where is a model parameter.

We ignore 1/f noise in this model because the proposed circuit
Parameters and were obtained from least mean
is designed to be resistant to 1/f noise and approximate input
squares fits performed in MATLAB using data from 35 discrete
referred noise [31] as
opamps from Texas Instruments with input bias currents below
50 pA [32]. Figs. 7 and 8 show scatterplots of supply current
against opamp parameters and the least mean squares fit. The
(17) mean of the absolute relative errors is 57.3%. Although this is
high, we deem this to be sufficient for reasons that will become
where is a model parameter. Given the input referred noise apparent when we discuss dynamic adaptation techniques in
and the bandwidth, we choose as the next section.

D. Trade-Off Between LED Power and Amplifier Power

(18) We assume that an SNR of 100 is necessary, perfusion is
1.5%, photodiode junction capacitance is 1 nF, and that the DC
value of the photocurrent is 100 nA. With this, we sweep the
Therefore, the current consumption of the opamp according LED on-time , compute opamp parameters needed for
to our model is each , determine power for that opamp from the previ-
ously described power model and then plot the power consump-
tion against LED on-time in Fig. 9.
When designing low-power pulse oximeters, it is tempting
to reduce LED on-time aggressively since the LEDs consume a
(19) major chunk of power. However, Fig. 9 shows that it is wiser to
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 277

err on the side of higher LED on-time since the slope is much
lower when the LED power dominates. The optimal on-time
for the probe we used corresponds to 237 s, which requires
a gain of 106, closed-loop bandwidth of 1 MHz, and an input
referred noise of . In our COTS implementation, the
TDC has a resolution of 62.5 ns, and ensuring that the SNR with
TDC quantization noise2 considered is above 100 resulted in the
on-time being 416 s. This requires a gain of 61, closed-loop
bandwidth of 0.6 MHz, and an input referred noise of 4.47
. Furthermore, the model used does not account for
noise from resistors, voltage references, photocurrent noise due
to ambient light etc. Because the trade-off curve is much less
sharper at higher on-times, it is safer to choose an on-time that Fig. 9. Trade-off between amplifier power and LED power.
is higher than optimal. Thus, the first stage is chosen to have
noise gain of 11 with unity gain bandwidth of 3 MHz, and the
second stage to have a gain adjustable in the range 1 to 10 with
a unity gain bandwidth of 3 MHz.

IV. DYNAMICALLY REDUCING POWER

We present two dynamic adaptation techniques that reduce
power by adapting the system to the user of the pulse oximeter.

A. Minimum SNR Tracking

In the previous section, while optimizing the amplifier param-
eters, we assumed that the perfusion is 1.5% and that the average
photocurrent is 100 nA. The average photocurrent depends on
the thickness of the finger, tissue perfusion etc., and can vary Fig. 10. SNR versus LED on-time for different fingers of one volunteer.
over an order of magnitude. Perfusion also exhibits consider-
able variance (0.1% to 5%) and is dependent on the ambient
temperature. Furthermore, the noise input into the system de- TABLE I
NOISE ESTIMATION PERFORMANCE
pends on ambient lighting and the extent to which ambient light
reaches the photodiode. Consequently, the pulse oximeter has
to cope with widely varying SNR levels. Because of this, the
optimal amplifier parameters determined in the previous sec-
tion should be treated as a guiding value rather than precise tar-
gets to achieve, and this is also why we accepted a rather large
relative error in estimating power consumption of amplifiers.
Fig. 10 shows how the SNR varies with measurement duration
for three fingers of a volunteer. Existing oximeters deal with
large variation in the input SNR by designing the interface cir- low-frequency content ( Hz). Assuming that the input noise
cuit with a large SNR margin. This however, causes the system is white, the filtered waveform should contain only pink noise.
to consume larger power than necessary in majority of the cases. The variance of this is calculated and scaled to obtain the noise
For instance, in Fig. 10, to achieve SNR of 100, the duration variance of the acquired photoplethysmogram [33]. Thus SNR
for which the LED needs to be turned on (and hence power) is is computed.
2.5 times higher for the middle finger than for the forefinger. To evaluate this approach, 400 samples were recorded at high
We propose measuring the SNR of the acquired photoplethys- SNR (this was ensured by keeping measurement duration above
mogram continuously and dynamically adjusting LED on-time 1 ms). This high SNR recording was deemed to have infinite
(and hence power consumption) so that the system operates at SNR. Additive White Gaussian Noise (AWGN) was then added
just sufficient SNR. We christen this technique “Minimum SNR to this recording, and an estimate of the AWGN by the above
tracking”. described method was obtained. Table I shows the performance
The signal level can be estimated by measuring the AC value of blind estimation of noise with a fourth order Butterworth
of the plethysmogram. Since the plethysmogram is known to high pass filter. The mean of the estimated noise over 1000 re-
have mostly low frequency components (most of its energy is alizations of pseudo-random noise is reported. The proposed
in the frequency range of 0.5 to 10 Hz), we filter it to remove all blind estimation method works well at low to medium SNR
2The maximum photocurrent beyond which the TDC quantization noise
and underestimates high SNRs. It is able to compute SNR read-
alone causes the SNR to be lower than acceptable is ings below 100 satisfactorily with a fourth order Butterworth
where is the resolution of the TDC. high-pass filter.
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
278 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO. 2, APRIL 2015

TABLE II
COMPONENTS USED IN THE PROTOTYPE

However, this technique is not without issues. The heart rate

can vary considerably from one cycle to the other. So, it is
possible for the loop to lock on to a fraction of the heart rate,
which results in misreporting of the heart rate. To prevent this
from happening for a prolonged duration, we reset the loop once
every 10 cycles and force it to “re-lock”. Hence, it takes much
longer to be confident of heart rate measurements. Moreover,
with this technique, a record of the complete plethysmogram is
not kept. So, the care-giver cannot visualize the plethysmogram.
Therefore, this approach is suitable when the power budget is
aggressive and the entire plethysmogram is not desired.
A flowchart describing both the “Minimum SNR tracking”
and “PLL tracking” techniques is shown in Fig. 11. To detect
the peaks, we used a time-domain window based peak detector.
The amplitude of the edge in the window is tracked once it ex-
ceeds a certain threshold. After the peak passes, the amplitude in
the window starts to fall. At this point we declare that a peak has
been found, and the detection threshold is reset to be a fraction
of the most recently observed peak amplitude. This threshold is
then gradually reduced over time so that smaller peaks, if they
occur in the future, are accepted. However, the power reduc-
tion methods discussed in this paper do not solely rely on this
Fig. 11. Flowchart describing the dynamic adaptation techniques. particular peak detection algorithm, and other methods, such as
template matching, may also be used.

B. PLL Tracking V. EXPERIMENTAL RESULTS

A prototype was built using CC2540, a low power platform
Since acquiring samples (which involves lighting an LED)
from Texas Instruments. The Bluetooth Low Energy (BLE)
is the most power intensive part of gathering a plethysmogram,
radio was disabled, and the timer peripheral on the SoC was
one way to reduce power is to reduce the duration for which
used as the time-to-digital converter (TDC). The fingertip
the LED is turned ON for each measurement, as discussed in
probe used was a ChoiceMMed MD2000B. The other compo-
the previous sections. Another approach to reducing power is to
nents used in the prototype referenced in Fig. 1 are detailed in
take fewer samples which translates into turning ON the LED
Table II.
fewer times on average.
A representative photoplethysmogram from the prototype
The samples that are actually useful in computing are
oximeter is shown in Fig. 12. Plotted3 on the vertical axis
those at the peaks and troughs of the photoplethysmogram. So,
are measurements made by the TDC in Fig. 1. We chose the
if we can estimate when these peaks and troughs are likely to
sample rate for each channel (red or IR) to be 100 Hz because it
happen, sampling can be performed only just before the time
allows us to easily suppress the 50 Hz power line interference
when these events are expected to occur. Since this is similar to
with a 2-point FIR low-pass filter. First, the control algorithm
the way a PLL works, we call this technique “PLL tracking”.
recognizes that the SNR of the acquired plethysmogram is
We emphasize that this kind of non-uniform sampling is pos-
higher than necessary, and the step like reduction indicates the
sible only because our photodetector architecture does not have
decrease in LED on-time (increase in the gain of the second
a band pass filter in the pipeline.
stage in Fig. 1). Second, the “PLL tracking” algorithm skips
We implemented this technique, wherein after detecting the
samples between peaks. After the reduction of LED on-time,
occurrence of a peak in the photoplethysmogram, the system
skips sampling for a duration equal to 50% of the expected peak- 3The real-time plotting tool we developed is available under an open source
to-peak interval. license at https://github.com/s-gv/stream-plot.
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 279

Fig. 13. Performance of “Minimum SNR tracking” in the presence of motion

artifacts. The vertical axis is normalized to 31.25 s.

Fig. 12. A photoplethysmogram from the proposed pulse oximeter. The

vertical axis is normalized to 31.25 s. level from the past 5 peaks is averaged after removing the max-
imum and minimum signal level, and then SNR is computed.
In the event that an erroneous signal level measurement creeps
the “PLL tracker” is reset, and then it “re-locks” to the heart in through this check, one of two things can happen: either the
rate. SNR is underestimated or it is overestimated. If the SNR is un-
In our tests, the “Minimum SNR tracking” loop settled to a derestimated, the loop increases the measurement duration and
steady state within 25 seconds after the finger is inserted into burns more power for a while before reducing the measurement
the probe. This is acceptable because we expect this to be used duration. In this case, the readings are unaffected, and the only
when is remotely monitored continuously for hours at impact is that the system burns higher power than necessary for
a time. Furthermore, since each peak-trough measurement is a short while. On the other hand, if the SNR is overestimated,
tagged with SNR, can be confidently measured even if the loop reduces the measurement duration, and the SNR of
the loop is not in steady state. the acquired readings will be lower than acceptable. Because
each signal measurement is tagged with SNR, those readings
A. Calibration and Accuracy with below acceptable SNR get rejected and are not reported
to the user. After detecting that the SNR is lower than neces-
We calibrated the prototype using finger simulators from sary, the measurement duration is increased. A sample wave-
BC Biomedical group. The finger simulators had known form where the loop erroneously chooses a lower than accept-
levels of 80%, 90%, and 97%. Performing linear regression on able LED ON-time because of motion artifacts and then re-
the acquired readings gave and in (5). covers after the transient artifact has passed, is shown in Fig. 13.
Accuracy of the proposed oximeter was verified by com- Since changes relatively slowly, the occasional loss of
paring the readings from the proposed oximeter with read- readings will not affect system operation. Furthermore, it is pos-
ings from a commercial fingertip oximeter ChoiceMMed sible to avoid even this if we are willing to burn slightly more
MD300C2D. The accuracy test was performed on 10 adults. power by setting the target SNR of the loop to be slightly higher
Among the 10 volunteers, the highest measured was 99% than the minimum required SNR. Then, even if the loop occa-
and the lowest was 93%. In 9 cases, the difference between sionally acquires readings at SNR levels below its target, the
reported readings from the prototype oximeter and the SNR still remains higher than what is necessary, and the read-
commercial oximeter was no more than 1%. In one case, the ings continue to be valid. The fact that motion artifacts are tran-
difference was 2%. sient, coupled with the constant step approach to changing the
measurement duration, reduces the impact of such artifacts.
B. Impact of Motion Artifacts
The “PLL tracking” algorithm is also affected by motion ar-
One of the issues with pulse oximetry is that it is difficult to tifacts, but in a more benign manner. The rapid occurrence of
obtain readings when there is relative motion between the edge-like features in the plethysmogram during motion artifacts
probe and the body. Techniques to alleviate this have been pro- can cause the “PLL tracking” algorithm to underestimate the
posed in [34], [35]. While those techniques cannot be directly peak to peak interval. Once the transient artifact has passed, the
applied to our system, we demonstrate that the power reduc- tracking algorithm, which now expects a short peak to peak in-
tion techniques we have described in this paper are themselves terval, waits until the next peak due to a heartbeat and “relock”
resistant to occasional motion artifacts. However, similar to a to the heart rate. A recording that demonstrates this is shown
conventional oximeter, the proposed power optimized oximeter in Fig. 14. Even in the extraordinary event that the motion arti-
is still vulnerable to motion artifacts. fact causes the algorithm to lock onto an integral fraction of the
We argue that motion artifacts will not deteriorate the per- heart rate, the duration for which the heart rate is misreported
formance of “Minimum SNR tracking”. To reduce chances of is controlled. This is because the tracking loop is forced to re-
motion artifacts affecting signal level measurement, the signal lock once every 10 cycles. The output of the heart rate from this
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
280 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO. 2, APRIL 2015

TABLE III
COMPARISON OF POWER CONSUMPTION

Fig. 14. Performance of “PLL tracking” in the presence of motion artifacts.

The vertical axis is normalized to 31.25 s.

of the radio (which is turned OFF), is reported. All the elements

of the platform including analog circuitry, TDC, processor, and
the system clock generator, with the exception of the 32 kHz
sleep timer, are duty cycled.
A constraint posed by the CC2540 platform is that the pro-
cessor core has to be active when the timer, which we employ
as a TDC, is enabled. So, when “Minimum SNR” tracking is
enabled, and the LED ON-time reduces, the fraction of the time
for which the entire platform is active also reduces. Hence, the
processor power reduces. Similarly, when “PLL” tracking is
enabled, the entire platform, again with the exception of the
32 kHz sleep timer, sleeps for about 0.5 seconds when samples
Fig. 15. Power divided into LED Power, Analog power, and Processor power are skipped, and thus the average power goes down. Using “PLL
for fingers of 3 volunteers. “Plain” refers to our implementation with both
dynamic adaptation techniques disabled.
tracking” cuts the number of samples acquired by 35%, and as
we expected, the power consumption gets cut by the same per-
centage.
scheme has to be inspected and cleaned (for example, using a On average, “Minimum SNR tracking” reduces power con-
median filter) before reporting the heart rate to the user. sumption by a factor of 3.9. Furthermore, with this technique,
as we expected, the power consumption varies from one finger
C. Power Consumption to the other while the SNR is held constant. “PLL tracking” and
Power consumption was measured using Agilent source “Minimum SNR tracking” in tandem reduce power by on
measurement unit U2722A. The microcontroller and the analog average compared to when both techniques are disabled.
components were powered by separate channels, and the power Table III compares the power consumption of the proposed
for each was thus measured. The LEDs were powered from design with other implementations. The power reported is the
the microcontroller IO pins. Since we measured the duration total power consumption of the platform which includes LED
for which the LEDs were powered, and the LED current was power, analog power, and the power consumed by the pro-
kept at a fixed level (about 5 mA for the red LED and 7 mA for cessor. Data for commercial oximeters was obtained from [19].
the IR LED), we could compute the LED power and subtract This comparison only serves to show that our implementation
it from the total power consumed by the microcontroller to is competitive because the absolute numbers are dependent on
separate LED power from processing power. the probe characteristics. Finally, assuming that “PLL tracking”
The plot in Fig. 15 shows the impact of each of the proposed is completely ineffective and that the LED ON-time is at its
dynamic adaptation techniques and also how they work in maximum value of 2 ms (the TDC in our implementation over-
tandem to reduce power. Analog power is the total power flows at 2 ms), the estimated worst-case power consumption of
consumed by the analog parts of the system, which includes our prototype is about 16 mW.
opamps, comparators, and reference generators. Processor
power refers to the power consumed by the digital parts of the VI. CONCLUSION
oximeter, which includes the power consumed by the TDC, the Remotely sensing physiological parameters could reduce
oscillator for the system clock, and the processor (which runs mortality while lowering the cost of health care, and low-power
control algorithms and calculates SpO2, heart rate from the sensors are essential for remote sensing. We have presented a
samples). The power for the entire platform, with the exception novel low-power pulse oximeter. In pulse oximetry, acquiring
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 281

samples is expensive in terms of power because of the need where is the photodiode noise current and is the voltage
to light an LED during sampling. Design-space analysis was noise due to the opamp voltage noise. Substituting and
performed to explore the trade-off between spending power in and subtracting
the LED and burning power in a higher performance amplifier.
However, designing for the worst-case results in excessive
power consumption in most situations. We showed that dy-
namic adaptation techniques lower power considerably by
operating the sensor at the edge of the SNR requirement and (26)
acquiring samples only when necessary.
Substituting from (8)

APPENDIX A
PERFUSION IN TERMS OF TDC MEASUREMENTS (27)

The peak and trough of the TDC measurements in Fig. 1 may

Re-arranging (26), squaring, and taking expectation
be written as

(20)

(28)
(21)
Since the photodiode noise and opamp noise are uncorrelated,
where and correspond to the peak and the chain rule for expectations can be used to obtain
trough of the photocurrent (see Fig. 3). Thus

(22)
(29)

Substituting for from (27)

(23)

where is and is
(30)
. From (22) and (23), the perfusion
P is Since and ,

(24)
(31)

Thus, the perfusion computed from TDC measurements is

equal that computed from the photocurrent measurements. Substituting this in (10)

APPENDIX B
NOISE ANALYSIS OF THE PHOTODIODE INTERFACE CIRCUIT (32)
In this section, we analyze the noise performance of the
switched integrator and derive (14). The ramp that is fed to the Re-arranging terms in (31), using (10), and setting
comparator may be written as

(25)
(33)
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
282 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 9, NO. 2, APRIL 2015

From (8) and (12) [12] N. Watthanawisuth, T. Lomas, A. Wisitsoraat, and A. Tuantranont,
“Wireless wearable pulse oximeter for health monitoring using
zigbee wireless sensor network,” in Proc. Int. Conf. Electrical Engi-
(34) neering/Electronics Computer Telecommunications and Information
Technology, 2010, pp. 575–579.
[13] T. Torfs, V. Leonov, C. Van Hoof, and B. Gyselinckx, “Body-heat pow-
Substituting this into (33) and using (9), the maximum per- ered autonomous pulse oximeter,” in Proc. 5th IEEE Conf. Sensors,
missible input referred noise spectral density is 2006, pp. 427–430.
[14] Y. Mendelson, R. J. Duckworth, and G. Comtois, “A wearable re-
flectance pulse oximeter for remote physiological monitoring,” in Proc.
28th Annu. Int. Conf. IEEE Engineering in Medicine and Biology Soc.,
2006, pp. 912–915.
[15] R. Haahr, S. Duun, K. Birkelund, P. Raahauge, P. Petersen, H. Dam, L.
Norgaard, and E. Thomsen, “A novel photodiode for reflectance pulse
oximetry in low-power applications,” in Proc. 29th Annu. Int. Conf.
(35) IEEE Engineering in Medicine and Biology Soc., 2007, pp. 2350–2353.
[16] C. Pujary, M. Savage, and Y. Mendelson, “Photodetector size consid-
erations in the design of a noninvasive reflectance pulse oximeter for
telemedicine applications,” in Proc. 29th IEEE Annu. Bioengineering
ACKNOWLEDGMENT Conf., 2003, pp. 148–149.
[17] R. Sarpeshkar, “Ultra low power electronics for medicine,” in Proc. Int.
The authors would like to thank Manikandan, Viveka, and Workshop Wearable and Implantable Body Sensor Networks, 2006, p.
1, pp. -37.
Pushkar for their suggestions and Hitesh for assisting with the [18] L. Turicchia, S. Mandal, M. Tavakoli, L. Fay, V. Misra, J. Bohorquez,
CC2540 platform. They also thank the anonymous reviewers for W. Sanchez, and R. Sarpeshkar, “Ultra-low-power electronics for non-
their comments, and acknowledge the volunteers who partici- invasive medical monitoring,” in Proc. IEEE Custom Integrated Cir-
cuits Conf., 2009, pp. 85–92.
pated in this research. Signed consent was obtained from these [19] M. Tavakoli, L. Turicchia, and R. Sarpeshkar, “An ultra-low-power
volunteers on a less-than-minimum-risk consent form which de- pulse oximeter implemented with an energy-efficient transimpedance
amplifier,” IEEE Trans. Biomed. Circuits Syst., vol. 4, no. 1, pp. 27–38,
scribed the experiment and informed participants that we would Feb. 2010.
not disclose any personally identifiable data and that their par- [20] K. Glaros and E. Drakakis, “A sub-mW fully-integrated pulse oximeter
ticipation or their refusal to participate in the research at any front-end,” IEEE Trans. Biomed. Circuits Syst., vol. 7, no. 3, pp.
363–375, Jun. 2013.
point in time would in no way affect their privileges at the in- [21] A. Al-Ali, “Low power pulse oximeter,” U.S. Patent Application 8 457
stitute or their professional relationship with the investigators. 703 B2, 2007.
[22] P. Baheti and H. Garudadri, “An ultra low power pulse oximeter sensor
based on compressed sensing,” in Proc. 6th Int. Workshop on Wearable
REFERENCES and Implantable Body Sensor Networks, Jun. 2009, pp. 144–148.
[23] E. Pelaez and E. Villegas, “Led power reduction trade-offs for ambula-
[1] D. G. Bassani et al., “Causes of neonatal and child mortality in India: tory pulse oximetry,” in Proc. 29th Annu. Int. Conf. IEEE Engineering
A nationally representative mortality survey.,” Lancet, vol. 376, no. in Medicine and Biology Soc., 2007, pp. 2296–2299.
9755, p. 1853, 2010. [24] K. Glaros and E. Drakakis, “Trade-offs for low power integrated pulse
[2] H. Rao, D. Saxena, S. Kumar, S. V. Gubbi, B. Amrutur, P. Mony, oximeters,” in Proc. IEEE Biomedical Circuits and Systems Conf.,
P. Thankachan, K. Shankar, S. Rao, and S. R. Bhat, “Low power re- 2009, pp. 245–248.
mote neonatal temperature monitoring device,” in Proc. 7th Int. Conf. [25] Precision Switched Integrator Transimpedance Amplifier 2013 [On-
Biomedical Electronics and Systems, 2014. line]. Available: http://www.ti.com/lit/ds/symlink/ivc102.pdf
[3] J. Webster, Design of Pulse Oximeters. Boca Raton, FL, USA: CRC [26] C. Enz and G. Temes, “Circuit techniques for reducing the effects of
Press, 1997. op-amp imperfections: Autozeroing, correlated double sampling, and
[4] N. Netzer, A. H. Eliasson, C. Netzer, and D. A. Kristo, “Overnight chopper stabilization,” Proc. IEEE, vol. 84, no. 11, pp. 1584–1614,
pulse oximetry for sleep-disordered breathing in adults: A review,” 1996.
CHEST J., vol. 120, no. 2, pp. 625–633, 2001. [27] E. Klumperink, S. L. J. Gierkink, A. Van Der Wel, and B. Nauta, “Re-
[5] Z. Zhu, R. K. Barnette, K. M. Fussell, R. M. Rodriguez, A. Canonico, ducing mosfet 1/f noise and power consumption by switched biasing,”
and R. W. Light, “Continuous oxygen monitoring: A better way to IEEE J. Solid-State Circuits, vol. 35, no. 7, pp. 994–1001, 2000.
prescribe long-term oxygen therapy,” Resp. Med., vol. 99, no. 11, pp. [28] S. Lopez, Application Note 4327, Pulse Oximeter Fundamentals and
1386–1392, 2005. Design 2012 [Online]. Available: http://www.freescale.com/files/
[6] A. K. Simonds, “Oximetry feedback flow control simulation for 32bit/doc/app_note/AN4327.pdf
oxygen therapy,” J. Clin. Monit. Comput., vol. 21, no. 2, pp. 115–123, [29] J. Guo and S. Sonkusale, “A 22-bit 110 ps time-interpolated time-to-
2007. digital converter,” in Proc. IEEE Int. Symp. Circuits and Systems, 2012,
[7] R. Haahr, S. Duun, M. Toft, B. Belhage, J. Larsen, K. Birkelund, and pp. 3166–3169.
E. Thomsen, “An electronic patch for wearable health monitoring by [30] T. Sepke, P. Holloway, C. Sodini, and H.-S. Lee, “Noise analysis for
reflectance pulse oximetry,” IEEE Trans. Biomed. Circuits Syst., vol. comparator-based circuits,” IEEE Trans. Circuits and Syst. I, Reg. Pa-
6, no. 1, pp. 45–53, 2012. pers, vol. 56, no. 3, pp. 541–553, 2009.
[8] F. Adochiei, C. Rotariu, R. Ciobotariu, and H. Costin, “A wireless low- [31] B. Razavi, Design of Analog CMOS Integrated Circuits. Noida,
power pulse oximetry system for patient telemonitoring,” in Proc. 7th India: Tata McGraw-Hill, 2002.
Int. Symp. Advanced Topics in Electrical Engineering, 2011, pp. 1–4. [32] Data Sheets From Texas Instruments 2013 [Online]. Available:
[9] W.-Y. Chung, Y.-D. Lee, and S.-J. Jung, “A wireless sensor network http://www.ti.com/lsds/ti/amplifiers-linear/operational-ampli-
compatible wearable u-healthcare monitoring system using integrated fier-op-am p-products.page
ecg, accelerometer and spo2,” in Proc. 30th Annu. Int. Conf. IEEE En- [33] J. Immerkar, “Fast noise variance estimation,” Comput. Vis. Image Un-
gineering in Medicine and Biology Soc., 2008, pp. 1529–1532. derstand., vol. 64, no. 2, pp. 300–302, 1996.
[10] P. Shaltis, S. Rhee, and H. Asada, “Implementation and validation of [34] J. Patterson and G.-Z. Yang, “Ratiometric artifact reduction in low
a power-efficient, high-speed modulation design for wireless oxygen power reflective photoplethysmography,” IEEE Trans. Biomed. Cir-
saturation measurement systems,” in Proc. 28th IEEE Annu. Northeast cuits Syst., vol. 5, no. 4, pp. 330–338, 2011.
Bioengineering Conf., 2002, pp. 193–194. [35] S. Rhee, B.-H. Yang, and H. Asada, “Artifact-resistant power-efficient
[11] E. Hughes, M. Masilela, P. Eddings, A. Raflq, C. Boanca, and R. Mer- design of finger-ring plethysmographic sensors,” IEEE Trans. Biomed.
rell, “Vmote: A wearable wireless health monitoring system,” in Proc. Eng., vol. 48, no. 7, pp. 795–805, 2001.
9th Int. Conf. E-Health Networking, Application and Services, 2007, [36] B. E. Sueppel and D. W. Mortara, “Low power pulse oximeter,” U. S.
pp. 330–331. Patent 6 697 655, Feb. 24, 2004.
Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.
GUBBI AND AMRUTUR: ADAPTIVE PULSE WIDTH CONTROL AND SAMPLING FOR LOW POWER PULSE OXIMETRY 283

Sagar Venkatesh Gubbi received the B.E. degree Bharadwaj Amrutur (M'94–SM'13) received the
in electronics and communication from the Sri B.Tech. degree in computer science and engineering
Jayachamarajendra College of Engineering, Mysore, from the Indian Institute of Technology, Bombay,
India, and the M.E. degree in electrical and com- India, in 1990, and the M.S. and Ph.D. degrees in
munication engineering from the Indian Institute electrical engineering from Stanford University,
of Science, Bangalore, India, in 2011 and 2013, Stanford, CA, USA, in 1994 and 1999, respectively.
respectively. He worked at Bell Labs, Agilent Labs, and
His research interests include ultra low-power cir- Greenfield Networks. Currently, he is an Associate
cuits and systems. Professor in the Department of Electrical and Com-
munication Engineering, Indian Institute of Science,
Bangalore, India, working in the areas of VLSI
circuits and systems.

Authorized licensed use limited to: University of Kerala. Downloaded on April 22,2024 at 12:20:43 UTC from IEEE Xplore. Restrictions apply.