PPG Theroy
PPG Theroy
©istockphoto.com/pandpstock001
Towards a Continuous
Non-Invasive Cuffless
Blood Pressure Monitoring
System Using PPG:
Systems and Circuits Review
Guoxing Wang, Mohamed Atef, and Yong Lian
H
ypertension (or high blood pressure) is a major stantaneous BP. It applies a small force perpendicular
risk factor for dangerous health conditions such to a superficial artery [9], [10]. A pressure transducer
as heart attack, stroke, chronic heart failure, and measures the pressure value at the contact point. The
kidney disease [1]. High Blood Pressure (BP) generally external force must be small to not dramatically reduce
develops over many years without any symptoms, and the blood flow through the artery. The contact is main-
could cause damage to blood vessels and heart continu- tained, and the transducer measures the blood pres-
ously. According to WHO, worldwide high blood pressure sure wave. The tonometry method needs a continuous
causes 7.5 million deaths, about 12.8% of the total of all control for the contact pressure and the transducer,
deaths [2]. Among people with high blood pressure, only which increases the system complexity and cost. The
54% of them have their condition under control [3]. Thus, accuracy is affected by the motion artifact. The BP
high blood pressure is one of the most important chal- monitoring becomes uncomfortable as the measure-
lenges in public healthcare globally. ment time increases.
To diagnose and manage high blood pressure, it is Pulse wave velocity (PWV) is a cuffless method that
important to measure BP routinely. Although the tra- can measure beat-to-beat BP over a long period [11].
ditional clinical BP measurement provides information on PWV refers to the velocity of the pressure pulses gen-
one’s BP level, it comes with several shortfalls, e.g. incon- erated during left ventricular ejection, after opening of
venience, sparse BP readings, and white coat syndrome. the aortic valve. The device normally measures pulse
Studies show that risks associated with hypertension can transit time as the time delay between the carotid artery
be better managed through frequent BP monitoring. BP pulse wave and the femoral artery pulse wave and use
monitoring at home not only provides better prognostic that to calculate the PWV. Unfortunately, it requires a
value but also prevents from serious organ damage com- trained medical expert to locate the arteries used for
pared with clinic BP monitoring [4]. This is because in- the BP measurements. The accuracy is significantly
home BP monitoring gives more frequent BP readings than affected by the motion during BP measurement. Also,
clinic BP monitoring. The frequent measurement leads to the distance between locations to determine the arrival
more accurate BP data than sparse clinic BP measurement. time of the pulse wave must be measured precisely and
Additional benefit of home BP monitoring is that it helps to manually [12].
identify white coat and masked hypertension. Compared Ideally the best method for continuous non-invasive
with ambulatory BP monitoring, in-home BP monitoring blood pressure (cNIBP) monitoring should offer high ac-
costs less and is easy to implement. curacy, remain comfortable for long time measurements
There are two different methods for BP monitoring, (i.e. cuffless measurement), require no medical supervi-
one is cuff-based and the other is cuffless. Commercially sion, and automatically store the data for later usage. The
available BP devices based on the oscillometric method recent development in wearable wireless cuff-less contin-
use an inflatable cuff during measurement. However, the uous BP monitoring devices meet most of these require-
inflatable cuff is not feasible for long-term BP measure- ments and could become the best candidate of cNIPB.
ments due to discomfort caused by repeated inflation Due to the emerging interest in recent years in de-
and deflation. If used during sleep, it can cause sleep veloping cNIBP measuring devices, IEEE introduced a
disturbances [5], [6]. cuffless BP measurement standard, i.e. IEEE standard
Volume clamp-based BP devices can measure contin- 1708–2014, which standardizes “Wearable, Cuffless Blood
uous BP using an inflatable finger cuff with a built-in PPG Pressure Measuring Devices” [5]. Table I summarizes the
(photoplethysmography) sensor and a pressure con- recommended accuracy levels that are equivalent to the
troller unit [7]. When the volume under the finger cuff American Association for the Advancement of Medical In-
becomes constant (constant PPG signal), the BP value strumentation (AAMI) requirements [13] and the BHS pro-
can be evaluated as the pressure value in the finger cuff. tocols that are recommended for cuff-based devices [5].
However, the vessel under the finger cuff is pressed, and The manufacturer should determine the grade level of the
the oxygen level decreases in these vessels, which can Device Under Test (DUT) according to the accuracy lev-
cause tissue hypoxia during continuous measurements. els indicated in Table I. The BP mean absolute difference
Variations in finger positioning within the cuff influence (MAD) is calculated using the DUT and a reference de-
the BP measurement, reducing its precision [8]. vice for both systolic blood pressure (SBP) and diastolic
G. Wang, M. Atef, and Y. Lian are with the School of Microelectronics, Shanghai Jiao Tong University, No. 800 Dongchuan Road, Minhang District, Shang-
hai, China, e-mail: [email protected]. M. Atef is also with the Electrical Engineering Department, Assiut University, Assiut, Egypt, e-mail:moh_atef@
aun.edu.eg.
Power Manager
Synchronized
Analog ECG and PPG Waves’s Peaks
Signal Detection/PTT
PPG Bluetooth Calculation/
Sensor Blood Pressure
Control Wireless
Module Transmission Display
Analog Digital
Input (MCU) Output
Analog
Module Module
Signal Serial Port
ECG Initialization
Sensor and Debugging
60 Secondary Peak
transmitted through thick tissue, reflectance mode can
First
be used in any part of the human body, while transmis- 40 Pulse
sion mode is limited to peripheral parts such as the fin- Foot
20
gertip, foot, and earlobe. For example, PPG signals from Dicrotic Notch
the front of the head and chest can only be detected in 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
reflectance mode. It is worth noting that fingertips and Time/s
earlobes have limited blood perfusion such that these
Figure 4. PPG waveform with basic BP features indicated.
sites are susceptible to low ambient temperatures; addi-
B. Light Wavelength PD and the LED are, the better shallow surface light can
The light penetration depth is inversely related to its ab- be detected. The practical distance between the PD and
sorption coefficient (a) . The greater the absorption, the LED is from 4 to 6 mm [18]. Green light has better immuni-
smaller the light penetration depth in tissue. To minimize ty to MAs than red and IR lights, as the IR light penetrates
the light loss in thick tissues, the tissue thickness at the more deeply and is affected by all surface and deep tissue
PPG sensor position should not be greater than the pen- movements [20]. The PPG signal from the green light is re-
etration depth of the used light wavelength. For visible duced sharply by increasing contact pressure compared
light, the penetration depth through tissue under the skin with the IR PPG signal. This is an indication of the occlu-
surface ranges from 100 nm to 1200 nm [19]. A reflectance sions of superficial skin blood vessels. The PPG second-
PPG sensor located in thin tissue regions like fingertips, ary peak and dicrotic notch are more structured with red
earlobe, and the forehead results in stronger signals than than with green light (Fig. 4) [19].
in thicker tissue regions such as the chest. Red and IR
light can penetrate deeper into the tissue than green light C. PPG Sensor Position
wavelengths can; these wavelengths can thus give better There are different possible sites for PPG sensor place-
results for deep-tissue blood flow in the muscles. In ad- ment, such as the fingertip, forehead, earlobe, chin, and
dition, the photodetector responsivity of the red and IR wrist. Fingertips give the strongest PPG signal, the fore-
wavelengths is greater than that of the green light. Green head and earlobe have good signal strength, and the
light, on the other hand, is absorbed near the surface, wrist has a low signal amplitude. Although the fingertip
so it is suitable for the shallow blood flow directly under produces the strongest PPG signal, the MAs associated
the skin. Short wavelengths like blue and green (405 to with routine activities produce a great deal of noise,
520 nm) have stronger blood absorption than red and IR which reduces the signal quality. Wrist and earlobe loca-
light (650 to 850 nm). Therefore, the signal-to-noise ratio tions also suffer from MAs due to hand and neck move-
(SNR) for a green light source is better than that of an IR ments, respectively. The forehead has a flat and large
source when used for superficial blood flow in the skin. region for sensor placement, easy attachment by using a
The received signal at a certain light penetration depth head bandage, and is a stable region with low motion ar-
can be controlled by adjusting the distance between the tifact effects. It also produces a strong PPG signal. How-
LED and the photodiode. By increasing the distance be- ever, as forehead PPG sensors work only in reflectance
tween the photodiode (PD) and the LED, the PD will de- mode, dirt and sweat may affect measurement accuracy.
tect the light rays with deeper penetration. The closer the The chin region has a fair PPG signal, but it suffers from
growing hairs, dirt and sweat [18], [21], [22]. The work in
[18] studied the different locations with respect to PPG
signal strength, attachment, artifact, and portability.
Circuit
Board (CPU, Outer Ring They introduced a decision matrix to determine the pre-
RF Transmitter, (Housing) ferred sensor locations, and the forehead received the
etc.) highest score.
wrist radial and ulnar arteries. The wristwatch is com- III. Blood Pressure Modeling and Algorithms
fortable to wear, but the PPG signal coming from the There are several methods to estimate the BP depending
wrist area is weak. An array of phototransistors and on the sensing and calculation techniques. One method
LEDs were placed to improve the PPG sensor sensitiv- is to calculate BP based on the PTT. The Another meth-
ity and its accuracy. A conductive fiber wristband was od uses PPG signal only to calculate the BP in order to
employed to reduce external noise. An armband-based reduce the hardware complexity.
wearable monitoring device was developed in [25] us- There are two main directions for BP estimation
ing a near-infrared LED and a phototransistor in reflec- using PPG only. The first one involves the parametric
tive mode. models, which try to extract specific parameters such
A Nellcor forehead reflectance-mode sensor [26] se- as instantaneous heart rate, systolic, and diastolic pe-
cures the sensor by means of a headband. Integrating riods from individual PPGs. Then, the BP can be pre-
the PPG sensor into preexisting equipment like a sol- dicted using these extracted parameters and a para-
dier’s helmet provides a comfortable and acceptable at- metric model (Section III B). The second direction is
tachment for the user [18]. the non-parametric models. Non-parametric BP meth-
In [22], an ear-worn reflective PPG sensor with im- ods try to extract specific features (time domain or
proved surface coupling was proposed. The sensor has frequency domain features) with strong relations to BP
been proven to be robust against pressure-induced sen- from the PPG waveform.
sitivity degradation [22]. A wireless earpiece PPG sensor The main steps to calculate the BP from PPG signals
with a magnetic earring was introduced in [27]. The sensor are listed in [32]; some steps can be merged from differ-
can be comfortably worn on the earlobe, and it contains an ent algorithms into the following steps:
embedded accelerometer to provide MA noise reference 1) Preprocessing: deals with noises caused by power
for adaptive noise cancellation. However, the earlobe be- line interface, baseline wandering [33], motion arti-
comes susceptible to inaccuracies due to peripheral vaso- facts [34]–[37], and signal quality assessment [38],
constriction and perfusion. To overcome these limitations, [39]. Classification is performed to discriminate be-
the external auditory canal has been proposed as an alter- tween motion artifacts and clean PPG. A band pass
native monitoring site, on the hypothesis that this central filter is used to remove the high-frequency noise
site will be better perfused [28]. above 5 Hz and reject the DC and low-frequency
For low cost and high integration ability, blood pres- noise component below 0.5 Hz.
sure can be estimated using a smartphone camera.
Initial BP screening with a low-cost solution using a
simple and portable device is the aim for such smart- Smartphone
phone health-care applications [29]–[31]. The PPG sig- LED PPG ANN
nal is processed by an artificial neural network (ANN)
built in a smartphone to evaluate the BP, as shown in Display
Fig.7 [31]. The on-line evaluation of the experimental Camera Acquired
Frames Signal
BP results of the proposed method shows a maximum
Processing
error of 12 mmHg [31]. This reported BP determina-
tion accuracy using a smartphone is still worse than
Figure 7. Capturing and processing PPG data using smart-
the 8 mmHg required for it to be used as a commercial
phone [31].
healthcare device.
Td
T 90 of Main Peak
vessel elasticity and the wave speed of a pressure pulse
0.8 Maximum First Derivative: in a thin-walled vessel can be described by the Moens-
Tsmaxdrv
T maxdrv1
Ts 1
Kortweg equation ([48]):
0.6 Valley of
Dicrotic Notch L = hE 0 exp (gP)
0.4 Peak of PWV = , (1)
Dicrotic Notch PTT tR
Tdn Valley
0.2
Tdn Peak
where PWV is the pulse wave velocity, L is the length of
0 the vessel, PTT is the pulse transit time, h is the thick-
0 0.2 0.4 0.6 0.8 1 ness of the vessel wall, R is the radius of the vessel, t
Time (s)
is the blood density, E 0 is the zero-pressure modulus in
Figure 8. PPG waveform with important BP features indicated. mmHg, and g is a constant that depends on the particu-
lar vessel (typically 0.016 mmHg - 1 to 0.018 mmHg - 1 ).
Under the assumption that the overall percentage
change in the arterial wall thickness (h) and arterial ra-
dius (R) are small over the physiological pressure range,
R
equation (1) can be simplified to relate BP to PTT [49], i.e.:
ECG 2
2RtL
P = 1 c ln c m - 2 ln (PTT) m = k 1 ln (PTT) + k 2, (2)
p hE 0
Z
] L tb
] 1 + exp (bPint) , h = 0
]
]
P = [ 2L · ln (6)
] gh tb
] exp [b (Pint - Phydro)] + exp (- bPhydro) - exp (- bPhydro)
]] ,h ! 0
exp [b (Pint - Phydro)] + 1 - 1
\
0.6 grade B for both DBP and mean blood pressure (MBP)
0.5 estimations and reaches grade C for SBP according to the
0.4 BHS protocol. In [42], an ANN-based learning process is
0.3
0.2 Trought Depth Difference used to train the features from the PPG signal only. The
Pulse Length
0.1 results can fulfill the AAMI requirement. Using only the
0 2 4 6 8 10 12 14
PPG signal, [43] implements a random forest (RF)-based
Time/s
prediction model and achieves grade B for both SBP and
Clean PPG Artifact
DBP by BHS. Even though the non-parametric BP models
exhibit better accuracy than the parametric models, they
Figure 10. Time domain features of the PPG signal and mo-
suffer from heavy calculations.
tion artifact effects [32].
Table III compares the different PPG-based BP models
and algorithms. The first five methods in Table III [25],
xtracted features. Non-parametric models try to solve
e [34], [37], [58], [59] use both ECG and PPG signals with lin-
the low accuracy of parametric models. ear or nonlinear regression. These methods have higher
Examples of non-parametric learning methods can be accuracy, but they require both ECG and PPG signals and
found in [30], [41]–[43]. Reference [30] tries to classify the frequent calibration. The methods in [30], [41]–[43] do
SBP and DBP within different prediction ranges with 19 not need frequent calibration, and only PPG information
features extracted from the PPG signal using only a sup- is needed. However, they require excessive training and
"
ing both ECG and PPG signals. The x PPG y=x+m + X
(Body Volume
last three methods [45], [46], [60] Changes) Sensor Restored
m – Signal
are only based on the PPG signal. Motion
"
Artiface m
They use linear regression and
Estimated
non-parametric models that do Motion
not need as much training and Artifact
calculation as the SVM methods. Noise Source Accelerometer a Adaptive
(Body Motion) (Noise Reference) Filter
However, they do require calibra-
tion and have the highest error.
VPD VDD
+
–
MAGC
AGC
RF M2
Low-Pass Filter Vbias2 M3
– Peak
+ AGC Circuitry M7 Detector
Vout
Vbias3 Vbias1
M8
TIAF M4 M5
SI
STIA
Cint
Iph Rf M10
– VTIA Rint
A – Log Amp
+0 Vint Vcnt
Cd A1
OTAF +
Vph
REF Vref ≈ avg (VTIA)
Figure 15. Logarithmic TIA with ambient light rejection and
Figure 14. TIA and variable-gain switched-integrator (SI) [70]. AGC loops [67].
Combiner
Register
Accumulator OUT [k]
a voltage mode switched-integrator Vd (t ) LCD
Iphoto(t ) Kaccum
(SI). This block serves multiple pur 45 –1
1–z
poses, acting as a variable-gain am- Rpar
plification stage that enables imple-
CLK(32 kHz)
mentation of large on-chip gains Off-Chip
exceeding 50 MΩ, which can be eas- Photodiode
ily tuned via the digitally controlled LIQAF1 LIQAF2 LIQAF3
integration time.
A logarithmic TIA depicted in Figure 16. Light-to-Digital Converter architecture [71].
Fig. 15 uses compression tech-
niques to reduce the signal level at
high input current [67]. Additionally, an AGC loop is intro- rejection is needed to prevent PPG receiver saturation.
duced to further extend the dynamic range. M1 acts as the An optical preamplifier with ambient photocurrent re-
logarithmic amplifier transistor. M6, M7 and M9 provide jection is shown in Fig. 17 [72]. The transimpedance
the feedback resistors, their resistances are controlled amplifier converts the current from the photodiode into
by the AGC signal. The transimpedance gain is specified a differential voltage. Surrounding the amplifier is an
by the current mirror ratio ^ g m5 /g m4 h and the feedback re- outer feedback loop comprising an error amplifier and
sistance. M9 acts as a feedback transistor placed directly a transistor. The rejection circuit operates as follows:
across the input and output terminals of the current mir- The DC component of the photocurrent produces an
ror (M4, M5). Using this direct feedback topology, M9 re- offset in the average levels of the differential outputs.
duces the input impedance seen at the photodiode, im- This offset is integrated over time by the error amplifier.
proving TIA speed by the cost of the lower output swing. Transistor M ctl acts as a variable current sink that, at
In [71] the traditional TIA is replaced by a digital-to- steady-state, eliminates the average photocurrent from
resistance converter (DRC) consisting of a logarithmic the signal path. The average photocurrent consists of
resistor network and a digital multi-bit RD modulator as the ambient photocurrent and the DC component of the
shown in Fig. 16. A parallel resistor, R par, improves sen- signal. The transistor M 10 in Fig. 15 [67] has the same
sitivity and dynamic range by extending the minimum function as M ctl . For the pulsed PPG signal used to re-
value of I bias that is supported. The combination of the duce sensor power consumption, the traditional method
DRC and R par enable continuous operation of the Light-
to-Digital Converter (LDC) over approximately a three-
order-of-magnitude range of I bias . The DRC is controlled
Vref
through feedback, which consists of a digital accumula-
tor and a nonuniform quantizer based on a laddered in-
verter quantizer /amplifier /filter (LIQAF) [71]. The vari- Transimpedance
Cd ′ Rf Preamplifier
able resistance R DRC (t) is achieved with a set of PMOS
switches connected to taps of a polysilicon resistor +
+ –
string ladder, with the PMOS switches being controlled is + Idc is Vout
by the output of a RD digital modulator that provides – +
Idc 0
A
high-resolution resistance (32 steps from 64 K to 10 MΩ)
Rf –
control through dithering. In contrast to the previous
Mctl
AGC approaches, the LDC system offers a high digital in- +
Error
tegration that minimizes analog complexity.
Amplifier
[68], [72] is not accurate. A DC photocurrent rejection as adding an on-chip tuning capacitor may be required
TIA with sample-and-hold (S/H) in the feedback loop to improve matching.
was introduced in [73]. The additional S/H block helps A block diagram of the implemented PPG SoC system
to detect the pulsed signal average accurately. Using a introduced in [74] is shown in Fig. 18. It consists of an in-
differential signal path helps maximize the preamplifier tegrate-and-hold amplifier, a dynamic range enhancement
immunity to noise from the power supply and substrate. (DRE) current source, a 9-bit successive approximation
The photodiode, however, is connected to only one ter- (SAR) ADC, and a digital core to supply feedback and con-
minal. This creates an asymmetry at the input of the dif- trol signals to the analog blocks (not shown for simplic-
ferential structure. As a result, an additional capacitor is ity), such as the sampling switch at the negative terminal
required at the other input of the transimpedance ampli- of the op-amp and the reset switch across the CF [74].
fier in order to match the photodiode capacitance and to The DRE circuit generates a current (Ist), which subtracts
rebalance the differential circuit. Perfect matching en- the static component of the photocurrent (Iph), and the
sures that noise injected at the bias voltage appears as residue (Iph-Ist) is integrated by the integrate-and-hold
a common-mode signal that is effectively rejected by the amplifier. Clean PPG data can be accessed from the SoC
differential structure. In practice, however, some mis- by using a standard SPI protocol, allowing for easy inter-
match can be expected, and additional measures such facing with low-power microcontrollers. However, this
DRE technique is very complex and requires more area
for the digital part compared to the analog solutions.
DRE Digital
C. PPG Low Frequency Contents
Feedback
VCMI The PPG signal has a low frequency of about 0.5 Hz and
a high frequency up to 5 Hz. Therefore, the AGC and DC
Ist
+ Vout photocurrent cancellation circuits need to be designed
ADC
– with large time constants ^T = 1/RC h to have less than
IRes Rf
a 0.5 Hz low cutoff frequency. In addition, the bandwidth
Iph of the PPG amplification stage should be around 5 Hz to
Cf
reduce the out of band noise. This implies that large ca-
pacitances and large resistances must be implemented.
The total area of the resistor and capacitor will be quite
Figure 18. A block diagram of the system on chip (SoC) with
large. Some special techniques have been explored to
DRE for static photocurrent rejection [74].
solve the above problems.
1) MOS pseudo-resistor
A resistance of hundreds of mega ohms can be achieved
Vctrl1 Vctrl2
Vctrl when a MOSFET is biased in the subthreshold region. This
Ma Mb Ma Mb type of MOS resistor is commonly used together with a ca-
Va Vbulk Vb Va Vc Vb pacitor in biomedical applications for low cutoff frequency
[75]. Since tuning a resistor by trimming is inconvenient
(a) (b)
and costly, a tunable MOS pseudo-resistor was proposed
Ma Mb [76], as shown in Fig. 19(a). The tolerable DC input voltage
Va Vb range was measured to be 0.5 V. Considering that this MOS
(c) resistor exhibits asymmetric and nonlinear resistance as
the voltage varies across it, [77] proposed a balanced tun-
Figure 19. MOS pseudo-resistors: (a) common tunable MOS able MOS pseudo-resistor in their programmable biomedi-
pseudo resistor [76], (b) balanced tunable MOS pseudo re- cal sensor interface chip, as illustrated in Fig. 19(b). The
sistor [76], and (c) MOS-bipolar pseudo resistor [78].
two transistors can be turned on in an alternating m anner
2) Switched capacitors
Switched capacitors can be used to implement on-chip r01
Vin Vout
PVT-insensitive high resistance as shown in Fig. 20(a) –Av1 –Av2
[79]. In this topology, the switching frequency fs and the Y
capacitor in the middle determine the resistance precisely
as 1/fs C . The switched-capacitor resistor in Fig. 20(b) miti- VDD
gates manufacturability and interference issues by realiz- Two-Stage MX4
MY4 Vb3
ing a tenfold resistance increase by charge sharing in the Miller
Cm1
switched-capacitor circuits [80]. Unfortunately, the switch- Capacitance
Multiplier MX3
ing produces additional noise along with the weak analog MY3 Vb2
signal in the front end, decreasing the receiver sensitivity. Rm2 Cm2 X Vb
r0x
r0y
MX2
3) Miller capacitance multiplier MY2 Vb1
A Miller capacitance multiplier is used to implement a Rb
large effective capacitance value using a small value ca-
MCN2 MCN1
pacitor [81], [82]. A two-stage Miller capacitance multipli-
er for compensation is proposed to implement the high GND
MBN MBN
5) Small transconductance OTA
The LPF can be designed using a
VSS
Gm-C technique. A reduced trans-
conductance (g m) OTA suitable
Figure 23. Low g m OTA for low-frequency applications [86].
for low-frequency applications is
required. Several techniques have
been proposed to implement small
g m for designing these filters with low cutoff frequencies
VDD [75]. The current division technique has been proposed
M3A
M3B to reduce the small signal transconductances in voltage-
IBias to-current converters [85]. The currents of two MOSFETs
M2B
are partially canceled at the output. The small signal
M2D +
–
Vin Vin transconductance of the OTA can be reduced to 10 - 8 A/V.
The current division method is insufficient to implement
M1A M1B transconductances at the nV/A level. To achieve an even
Mu
1: N2 M2C M2A N 2 : 1 smaller transconductance, the above methods can be
Mu combined [86], refer to Fig. 23. The transconductance of
the OTA can be reduced to 10 - 9 A/V. The drain currents
N of M 1 and M N are partially canceled at the output of the
OTA, leading to very large transconductance reductions.
Figure 24. PMOS-input symmetrical OTA with series-paral-
lel current division to reduce transconductance without loss
It can be shown that the resulting small signal transcon-
in a linear range [87]. ductance becomes the following:
i0 N- 1 g
Gm = = mMR (12)
v1 - v2 M+ N+ 1
dotted line [73]. Here, the virtual ground at the negative
input of the opamp has been extended to node X since In equation (12), M is defined as the ratio of transcon-
the a -block is enclosed within a loop. For PPG sensing ductances of M M and M 1, while N is the ratio of trans-
application, a single-ended version is sufficient. The conductances of M N and M 1 .
new transfer function is given by the following: Another technique to reduce the OTA g m is the se-
ries-parallel (SP) current division structure, which re-
- R F /R i
A (s) = , (10) duces the transconductance based on a SP current mir-
1 + SR F C ror. The transconductance of the OTA can be reduced to
a
a11 10 - 12 A/V [87], indicated in Fig. 24. For the NMOS current
mirrors in Fig. 24, N unit transistors are placed in series
a
f= , (11) and in parallel to achieve an effective output transcon-
2 rR f C
ductance G m = g m1 /N 2, where g m1 is the gate transcon-
With this technique, CS-LPF is implemented using a ductance of the transistors M 1 .
38 pF capacitor to achieve a cutoff frequency from 62 to
1.8 Hz (and even lower). The effective capacitive reduc- D. Modulation
tion is from 500 to 12500 times [73]. Sources of additive noise to the PPG signal include system-
The a -block consists of two transistors, with M 1 op- atic circuit offsets, flicker noise, external light sources inter-
erating at the triode region and acting as a linear resistor ference, and 50 Hz power line noise. As PPG signals have a
low bandwidth (5 Hz), higher frequency additive noise can any in-band transmitted noise reducing the SNR. There-
be removed with low-pass filtering. However, static (DC) fore, low-noise and low-ripple LED driver is required.
offsets and low frequency noise (also known as fixed pat- For wearable/mobile devices, the battery lifetime is a
tern noise), which occupy the same spectrum as the PPG critical issue. The battery lifetime can be extended by
signal, require different rejection techniques. One solution reducing the LED ON time; also, the LED current should
is to modulate the light source to shift it to a higher spec- be reduced as much as possible, as shown in Fig. 27. The
trum range where there is less noise content, as shown sampling rate needs to be high enough not to miss the
in Fig. 25. With a modulated light
source, the receiver can effectively
suppress the interference from am-
bient and 1/f noise. Demodulation Modulation Demodulation
envelope detection is attractive
due to the simplicity of its imple-
mentation and low power use, but + + + +
it is prone to distortion by addi- Vin Va Vb LPF Vout
tive noise. The system illustrated – – – –
+
in Fig. 25 uses decimation by 4 FIR
filters for the demodulation. Down-
Aggressors
sampling the signal by a factor M re- (a)
sults in placing a copy of the signal
Vin Signal Va Vb Vout
back in the baseband, completing Aggressores
the demodulation [74]. For demodu-
lation using multiplications inside
the MCU [88], as shown in Fig. 26, 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
the MCU spent much time in full- f/fch f/fch f/fch f/fch
(b)
power processing mode, which led
to an average high power consump- Figure 25. (a) Block diagram and (b) frequency-domain illustration of the chopper [74].
tion. To reduce power consumption
in the MCU, the percentage of time
that the processor core is in a
sleep or power-down state should
be increased. Processing time can Analogue Domain Digital Domain
be reduced by choosing an alter-
native MCU core that supports (1 + m [n])/2 Local Oscillator
LED
hardware multiplication [88]. Driver PWM
m [n]
Even though light modulation Homodyne
will save the MCU and LED trans- Bias Midrange
LPF
Output
mitter power dissipation, the detec- Stabilisation Setpoint
Fifth Order
+
tion with the pre-amplifier requires LPF Product
DAC Detector Baseband
a higher bandwidth and hence First Order –
LPF Output
higher bias current than if modula- ADC
Vout[n] Fifth Order
tion was not employed. Photodide
Amplifier
E. LED Driver
A key characteristic for the PPG
LED driver circuit is its low noise; Figure 26. A block diagram illustration of a single channel PPG homodyne detection
system with current offset compensation [88].
the optical receiver will amplify
Table IV.
Recently published integrated PPG sensors.
Process Biasing Noise Rejection Integrated LED Driver
[84] 0.35 μm 2.5 V/58 μA 40.8 nA 15 μA, 165 pF on chip No
[73] 0.35 μm 2.5 V/240 μA 3.53 nA 43.9 μA, 2.2 nF Off chip No
[71] 180 nm 0.5 V/8 μA 4 nA 3.5 μA logarithmic DRC No
[74] 180 nm 2.3 V/425 μW 2.5 nA 100 μA, DRE current source LSB 100 μA maximum current
of 25.6 mA
[70] 0.35 μm 3.3 V/1.88 mW 445 μV No 1.3 mW, Feedback path to
control the LED current