Microelectronics Journal 72 (2018) 1–10

Contents lists available at ScienceDirect

Microelectronics Journal
journal homepage: www.elsevier.com/locate/mejo

A pulse oximeter system, OxiSense, with embedded signal processing using

an ultra-low power ASIC designed for testability
Devarshi Mrinal Das * , Aman Gupta, Abhishek Srivastava, Amogh Vidwans, Meraj Ahmad,
Aniruddha Shelke, Salil Kale, J. Ananthapadmanabhan, Dinesh Kumar Sharma,
Maryam Shojaei Baghini
Department of Electrical Engineering, Indian Institute of Technology (IIT) Bombay, Mumbai, 400076, India

A R T I C L E I N F O A B S T R A C T

Keywords: This paper presents a complete IoT enabled pulse oximeter (PO) system, called OxiSense, made using a custom
Medical measurement system designed low power application specific integrated circuit (ASIC) for signal conditioning. One of the novel
ASIC design for testability
aspects of the developed system is the testability of all the important modules of the ASIC (both analog and
Pulse oximetry
digital) using auxiliary circuits implemented in the ASIC. Another novel aspect is a new embedded and efficient
SpO2
Photoplethysmogram (PPG)
signal processing algorithm which is robust to small motion artifacts. This new algorithm has been implemented
on a low-power micro-controller (embedded processing mode) as well as on PC using Python (remote processing
mode). The PO ASIC is fabricated in 180 nm mixed-mode CMOS technology and it works at 1.8 V supply voltage.
The average power consumption of the analog front-end is 176 μW and that of the digital module controlling the
analog front-end is 23 μW. The prototype PO system made using the custom ASIC operates from a 3.7 V lithium
ion battery and consumes about 8 mW power. The prototype is housed in a 3D printed casing and it connects
with a display device via USB or wirelessly to laptop/smartphone/tablet. Readings on 20 subjects, both in the
lab and in hospital, show less than 2% deviation in the measured S pO2 level and worst case deviation of 2.7
beats per minute for the measured heart rate, when compared with the commercial POs.

1. Introduction Pulse oximetry is a technique of determining the arterial blood oxy-

gen saturation. This technique has gained popularity due to some of
In a developing country like India, there is a sever shortage of med- its features like non-invasiveness, continuous monitoring, fast response
ical practitioners and other medical staff in hospitals. This shortage is and no requirement of medical expertise to operate [3]. It has found
more acute in rural areas. According to an estimate [1], there are only application in diverse areas such as intensive care units, operating the-
0.7 doctors and 1.5 nurses per 1000 people in India, significantly lower aters, trauma care centers in hospitals, in ambulances and in sleep stud-
than WHO global average of 2.5 doctors and nurses per 1000 peo- ies [4–6]. It is also finding a lot of interest in areas such as wearable
ple. Thus, there is a strong case for countering this shortage through and portable medical applications [7,8] and biometric security devices
technology. With the advancement in Internet of Things (IoT) and the [9]. SpO2 is defined as the percentage of hemoglobin in the arterial
launch of 4G services in India, it is envisaged that a fewer number of blood that is in oxygenated state. If [HbO2 ] and [Hb] represent con-
medical staff will be required in hospitals to monitor the same number centration of oxygenated and de-oxygenated hemoglobin, respectively,
of hospital beds. Hence, more of the medical devices are now required then by definition,
to have internet connectivity. Further, according to [2], 50.6% of visits
[HbO2 ]
to a primary health care centre in India are due to respiratory symp- S pO2 = ⋅ 100% (1)
[HbO2 ] + [Hb]
toms. Hence, an instrument like a Pulse Oximeter which is used to mea-
sure arterial blood oxygen saturation is very important. Fig. 1 depicts A typical pulse oximeter uses 660 nm (red) and 940 nm (infra red)
a typical application scenario of IoT enabled medical device for remote wavelengths for measurements [7]. From the measurements, a param-
health diagnostics and prognostics. eter which is only dependent on SpO2 , known as ratio-of-ratios (R) is

* Corresponding author.
E-mail address: [email protected] (D.M. Das).

https://doi.org/10.1016/j.mejo.2017.12.001
Received 14 May 2017; Received in revised form 23 September 2017; Accepted 1 December 2017
Available online 7 December 2017
0026-2692/© 2017 Elsevier Ltd. All rights reserved.
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

strated that is embedded in a smart hand-held device and it is made of

COTS components. The proposed system features chopper modulation
technique to achieve high resilience to interference caused by ambi-
ent light. The authors also propose: 1) an adaptive trough detection
technique to achieve improved HR and SpO2 detection and 2) a PPG
validation scheme using correlation between PPGs.
In this paper, we present a fully integrated PO measurement system
designed and developed using custom PO ASIC in 180 nm CMOS tech-
Fig. 1. A typical application scenario of IoT enabled pulse oximeter. nology. The motivation behind our work is to demonstrate the design
of a fully integrated PO ASIC keeping in mind the testability of indi-
vidual modules of both analog and digital sections in the ASIC. This
determined. R is defined as R = (ACR ∕DCR )∕(ACIR ∕DCIR ) where ACR,IR approach can serve as a guideline, for development of such similar com-
and DCR,IR represent the ac and dc component of photodetector cur- plex bio-signal measurement systems in wearable application scenario
rent due to red/infra red LED, respectively. The AC components is only where fully integrated approach is essential from different aspects such
about 1–4% of the DC component [5]. Each pulse oximeter is calibrated as: 1) high signal integrity, 2) low power consumption, 3) small form
and an empirical relationship is determined between SpO2 and R [6]. factor, 4) robust hardware design and 5) low cost. Further, we have
In Ref. [4], a logarithmic transimpedance amplifier is proposed developed a robust and efficient signal processing algorithm for process-
which is used in the PO front-end. It is a fully analog implementa- ing PPG signals and implemented it on a low-power micro-controller.
tion and achieves a low power consumption of 4.8 mW excluding the For remote monitoring, PPG signals are communicated from the sen-
power for digital processing. In Ref. [10], the authors have demon- sor node to a nearby (∼3 m distance) local receiver, which could be
strated the possibility of integrating a reflectance PO in an electronic any hand held device like mobile phone or a PC. Short range wire-
wearable patch with wireless connectivity using commercial-off-the- less communication of PPG is achieved by using sub-mW transmitter
shelf (COTS) components and used it to perform clinical evaluation on (TX) and receiver (RX) of the Bio-WiTel system [13], which operates
three subjects. However, using ASIC helps to reduce size, weight and in the medical device radio communication (MedRadio) freq. band of
power consumption. In Ref. [11], a technique to control the baseline 401–406 MHz [14].
drift of PPGs in digital domain without using any filter is proposed. This paper is organized as follows. In Sec. 2 we present the proposed
Again, in this work COTS components were used to design the proof- PO system. In Sec. 3 the details of design for testability of the PO ASIC
of-concept prototype. In Ref. [5], the first implementation of a fully and its testing is presented. Sec. 4 presents the details of the designed
integrated PO based on mixed-signal approach is presented. In this PO systems. Sec. 5 and 6 presents the signal processing aspects and
work, the authors have demonstrated a proof-of-concept system where measurement results, respectively. Finally, we conclude in Sec. 7.
PO front end consumes only 0.85 mW excluding power for digital pro-
cessing. This mixed-signal approach has gained momentum to alleviate 2. The proposed system - OxiSense
some of the design challenges in analog circuits in scaled technology
nodes [12]. In Ref. [6], the authors have proposed a method to reduce OxiSense, is a complete medical system meant for the measurement
power consumption of the PO front end by targeting an optimum SNR of arterial blood oxygen saturation (S pO2 ) and heart rate. Fig. 2 shows
using pulse width modulation to control LED power-on time. Further, the top-level block diagram of the system. The block diagram shows
power reduction is achieved by applying predictive sampling to sample the inter-relationships amongst the main sub-systems. At the core of
the photodiode current only near the peaks of PPGs where the useful the system, lies the custom designed “IIT Bombay Pulse Ox ASIC”. This
information lies, as far as extracting arterial blood oxygen saturation PO ASIC is a standalone signal conditioning module at the front-end of
and heart rate information is concerned. In Ref. [7], a PO is demon- OxiSense. To explain the working of the system, we shall first discuss

Fig. 2. The proposed pulse oximeter - OxiSense.

2
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

Fig. 3. Modules of the proposed pulse oximeter ASIC.

the modules of PO ASIC in Sec. 2.1 and 2.2. Followed by the explana- Negative input terminals of OTAs A4 and A41 are connected to pseudo-
tion of the phases of operation of the ASIC in Sec. 2.3. The hardware resistors in the negative feedback mode. This is not shown in Fig. 3 for
configuration of the PO ASIC is inspired from Ref. [5]. However, the simplicity. There are some analog buffers which are used to probe the
novel aspects of this work are: 1) design for testability considerations signal at the intermediate nodes and can be turned off during normal
of digital and analog modules, 2) IoT enabled solution by featuring a course of operations.
low power transmitter module and 3) development of a novel embed-
ded signal processing algorithm. The on-chip modules of this ASIC are 2.1.1. Low-leakage sample and hold circuit
depicted in Fig. 3 and they can be classified in two groups: 1) analog
The schematic diagram of LLSH is shown in Fig. 4. It samples the
circuits 2) digital module - Finite State Machine (FSM).
output voltage (Vout ) of the switched integrator (SI) at the end of the
calibration phase and holds it till the next calibration phase. Any devi-
2.1. Analog circuits ation in this stored value of voltage in LLSH will cause a DC current
(Δi) to flow through the transimpedance amplifier (TIAF). Assuming a
The primary role of the analog front end is to convert the weak ac gain of 10 MΩ for the combination of the TIAF and SI and a headroom
component of the photodetector (PD) current, due to both R and IR LED, of 1.5 V at the output of SI, Δi of 150 nA would saturate the SI. Fur-
to low-noise voltage signals. The ac component is usually <4 % of the ther, assuming a tissue attenuation of approx. 104 , this corresponds to
dc component [15]. The ac signal has the fundamental frequency com- a change in LED current (ΔILED = 104 × Δi) of 1.5 mA. Also, by design
ponent of 0.5–2 Hz corresponding to 30–120 beats per minute (bpm). the transconductance (Gm ) of LED driver is 100 mS. This means that a
The bandwidth of the ac component is 0.5–5 Hz [4]. The first block change in stored voltage (ΔVLED_CNTRL = ΔILED ∕Gm ) of 15 mV will sat-
following PD in Fig. 3 is a low-noise transimpedance amplifier (TIAF) urate the measurement channel. If the calibration phase is repeated
which converts the ac component of the PD current to an ac voltage sig- every 10 s, the droop rate (= ΔVLED_CNTRL ∕10s) comes out to be 1.5
nal. This enables the further processing of the signal in voltage-mode. mV∕s. Now, as the storage capacitor is 9 pF, this corresponds to a leak-
The output of the transimpedance amplifier is the input to a voltage- age current of 13.5 fA. From the foregoing analysis, we can see that this
mode switched integrator which provides programmable gain (by dig- leakage current can saturate the SI. The leakage current of a tuned off
itally controlling the integration time) and also acts as an anti-aliasing MOS transistor in 180 nm is typically of the order of few pA [16]. This
low pass filter. Using the programmable binary current reference, the is a few order of magnitude higher than the leakage current that can
reference transimpedance amplifier (TIAR), the switched integrator, the
low leakage sample and hold (LLSH) and the LED drivers, we make
in the internal calibration phase, the dc component of the PD current
equal to the current set by programmable binary current reference for
both red and infra red LEDs. Therefore, at the end of internal calibra-
tion phase, the LLSH corresponding to red and infra red channels have
two different voltage levels, which when applied to the corresponding
LED drivers, result in equal dc current in the photodetector. Hence, R
reduces to, R = ACR ∕ACIR . Now, R can be determined by just comput-
ing the ratio of peak-to-peak of red to infra red PPG signal. During the
measurement phase, to demultiplex the red and infra red channels two
sample and hold circuits made of OTAs A3 and A31 are used, respec-
tively. Finally, there is a gain stage made of capacitive feedback ampli-
fier using OTAs A4 and A41 for each of the channels. The gain is made
programmable by switching the input capacitance using 2:1 demux. Fig. 4. Low-leakage sample and hold (LLSH) circuit.

3
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

saturate the SI. The main contributor to leakage in the present design
is the leakage through switch S1. Implementation details of the switch
S1 is described in Ref. [5]. Even if the leakage current does not satu-
rate the SI and it affects both the channels in the same way, it would
introduce error in the determined value of the ratio of AC component
to the DC component of PPG. This parameter (AC/DC) may be used
to determine tissue perfusion. Following is the operation sequence for
the LLSH circuit shown in Fig. 4. First S1 and S4 are closed simulta-
neously. After some delay S2 is closed and S3 is opened. Voltage gets
sampled on storage capacitor. This is the sample mode. Now, S1 and
S4 are opened simultaneously. After some delay S2 is opened and S3 is
closed simultaneously, which switches LLSH in the hold mode.

2.2. Digital FSM

The analog part of the PO circuit requires 24 static logic signals

(0 V or 1.8 V) to control the currents, resistances and switches and 22
dynamic logic signals to control the operations of sampling and control
related switches during various phases of calibration and measurement,
as shown in Fig. 3. These signals are generated by an on-chip digital
block. The digital block has 33-state FSM which generates the static
and the dynamic logic signals based on the input. The FSM requires a
162 bit input which is provided through a serial input parallel output
162 bit shift register. The programming is controlled by the PROG_EN
input signal. When the PROG_EN goes low the programming stops and
the data in the 162 bit shift register gets loaded as input to the FSM. The
Fig. 5. Set-up for functionality testing of digital FSM.
FSM starts upon receiving the START input. There is one clock for the
entire digital block, which is shared by both the 162 bit shift register
and the FSM. The digital block occupies a die area of 300 μm × 300 μm.
3.1. The digital FSM
2.3. Phases of operation
Fig. 5 shows the block diagram of the setup used for testing the
The operation of the PO can be divided into two phases, namely: digital block. The digital FSM consists of the core FSM block and a 162
1) calibration phase and 2) measurement phase. During the calibration bit programmable shift register to initialize the FSM. The block diagram
phase the aim is to make DCR same as DCIR . To achieve this, first the also depicts the connection of the external circuits and instruments to
red LED is turned on and the calibration loop consisting of TIAR, TIAF the chip as well as the connection of the blocks inside the chip.
and SI is allowed to settle. The LLSH circuit samples and holds the The digital block receives 5 input signals which are CLK_IN, RESET,
output value of SI. Then the red LED is turned off and the same cycle is PROG_EN (logic high implies shift register is being programmed), SIN
then repeated for the infra red LED. Thus, at the end of the calibration (serial data input) and START (to start the FSM). These signals come
phase, we have two different voltage signals in the two LLSH such that from Spartan 3E FPGA Starter Kit [17], which provide low logic of 0 V
DCR is same as DCIR , which is equal to the reference current set by and high logic of 3.3 V. For logic level translation, we used level trans-
the programmable binary current reference. In the measurement phase, lator MAX13014 which converts 3.3 V–1.8 V. 162 bit programmable
first the red LED is turned on. The transimpedance amplifier (TIAF) shift register receives clock of 1 MHz. This shift register will be in reset
converts the photodetector current to a voltage signal. The following mode when signal RESET is logic high. To program the shift register,
stage, SI acts like a programmable gain amplifier whose gain can be RESET is made logic low and serial input, SIN will be shifted in at the
adjusted by adjusting the integration time. It also acts like a low pass rising edges of clock, CLK_IN, when program enable signal PROG_EN
filter, filtering out any high frequency noise present. This voltage signal is logic high. The PROG_EN is kept at logic high for 162 clock cycles
is sampled by the sample and hold circuit corresponding to the red during which the required bits are programmed in to the shift register.
channel. Once the measurement cycle of the red channel is completed, The MSB (i.e 161st bit) is loaded first and LSB (i.e 0th bit) is loaded at
the measurement cycle of the infra red channel begins after a delay. The last. After 162 rising edges of clock, CLK_IN, program enable PROG_EN
sampling frequency is programmable and it is set to 100 Hz. Red and is made low.
infra red LEDs are not kept turned on during the entire measurement The FSM takes data from this 162 bit shift register in parallel and
cycles, but duty cycling is done to conserve power. the FSM starts working when signal START is made logic 1. The FSM
switches states on rising edge of the clock, CLK_IN, and generates 24
static logic signals for programming the analog circuits, which does not
3. Design for testability and PO ASIC testing change with time, and 22 dynamic logic signals for the switching cir-
cuits in the analog block. All the outputs (22 dynamic logic signals +24
To ensure that the functionality of the analog modules and the per- static bits =46 output signals) generated by FSM go to 46 2:1 MUXs as
formance of digital FSM can be tested, design for testability features are shown in Fig. 5. Each of these MUXs receives inputs from FSM and FSM
incorporated in the design of the PO ASIC. In Ref. [4], all the modules alternative circuit (shift registers). To work with digital block (FSM) we
were individually tested in previous chips before the fully integrated make select line of these 46 2:1 MUXs high using DIGI_CNTRL_SWITCH
chip was fabricated. By ensuring design for testability, time and cost of pin. Now, the 24 static bits pass through 24 2:1 MUXs and 22 dynamic
development has been reduced in this work. Additional circuitry, which logic signals pass through 22 2:1 MUXs and go to the analog front-end
include 46 2:1 MUX, one 22:1 MUX, one 25 bit shift register and an FSM modules. To test these 46 outputs from FSM, we have 25-bit shift regis-
alternative circuit, has been included. ter to check the 24 static bits and 22:1 MUX to check 22 dynamic logic

4
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

signals.
The 24 static bits through 24 2:1 MUX also go to 25 bit shift regis-
ter. The 25 bit shift register receives clock of 10 kHz and a shift control.
This shift register will shift the signal serially at the falling edge of
the clock when shift control is high. When shift control is low then,
parallel data will be loaded and at the serial output hardwired logic 1
will appear. The 24 constant bits which come from FSM are the fol-
lowing contents of 162 bit shift register: bits 118–121, bits 138–147,
bits 148–151, bits 152–153, bits 154–155, bits 156–157. Measurements
show that the static output bits of digital FSM are correct.
Further, there are 22 waveforms generated by the FSM, which go
to 22 2:1 MUXs and then to analog front end. From these 2:1 MUXs
these 22 waveforms, also go to 22:1 MUX. From the output of this 22:1
MUX we can observe each waveform one at a time by controlling the
select lines. To verify the waveforms generated by FSM on-chip are
correct, we compare the pulse width and period of all these waveforms Fig. 7. Output of transimpedance amplifier (TIAF) for an input current of 80 nAp−p .
with the VHDL code level simulation results. We confirmed that all the
waveforms generated are correct. Hence, we verified the functionality
of the digital module. age through on-chip analog buffer (VTIAF_BUF ) is observed on oscillo-
scope. Input to off-chip V-to-I converter is provided from function gen-
3.2. Analog modules erator (F.G.). The transfer function of the off-chip V-to-I convert is
iin = −vin /1MΩ where iin is the output of off-chip V-to-I converter and
By design there is provision to characterize the working of the indi- input to the on-chip TIAF. Output voltage of on-chip TIAF will be given
vidual analog blocks i.e. the transimpedance amplifier, switched inte- by vTIAF_BUF = −iin (500 kΩ). Hence, theoretically vTIAF_BUF = 0.5 v2 .
grator, sample and hold circuit (for R and IR channels) and finally, the Fig. 7 shows one of the test cases when a 800 nA peak-peak input cur-
capacitive feedback amplifiers (for R and IR channels). Analog buffers rent is provided as input. For this particular test case the output voltage
are included, so that we can measure the output of the individual mod- swing is 39.6 mV which gives a transimpedance of 495 kΩ. From differ-
ules without loading them (see Fig. 3). These buffers are powered from ent measurements, we observed that the transimpedance gain is same
a separate power supply pin and are turned-off during normal oper- for different input current swing as expected, thereby confirming its
ation. Fig. 5 also shows the FSM alternative circuit. The role of this functionality.
block is to hold the analog modules in a particular state indefinitely
during testing of the analog modules. Additionally, they have been so 3.2.2. Testing of switched integrator
designed that in case the digital FSM fails to function correctly, we can The test setup for testing the switched integrator (SI) is shown in
by-pass the on-chip FSM and generate the FSM externally on FPGA and Fig. 8. The integration period (𝜏int ) of the SI is programmable. For test-
apply to the analog front end via the FSM alternative module. ing the SI module, a DC input (VSI ) is applied to the SI and the difference
between this voltage and the voltage at the non-inverting input is inte-
3.2.1. Testing of transimpedance amplifier grated for 𝜏int = 260 μs. The non-inverting input of SI is set to 998 mV
Fig. 6 shows the block diagram of the test setup for testing the by the TIAR. As the magnitude of the difference between the applied DC
transimpedance amplifier (TIAF). Transimpedance amplifier receives input voltage and the voltage at the non-inverting terminal decreases,
current from the photodetector and converts it into voltage provid- the slope of the output of SI (VSI_O_BUF ) decreases. Fig. 9 shows one of
ing transimpedance gain of 500 kΩ. For standalone testing, input cur- the test cases in which a DC input voltage of 985 mV is applied to SI.
rent is provided from an off-chip V-to-I converter and the output volt- As the input is less than the inverting input, output voltage increases

Fig. 6. Block diagram of test setup for testing TIAF. Fig. 8. Block diagram of test setup for testing SI.

5
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

Fig. 10. Micrograph of the pulse oximeter ASIC.

Fig. 9. Output of SI with 985 mV input and integration time of 260 μs. fully before proceeding with the designing of the prototype, OxiSense.

4. The designed PO system - OxiSense

Table 1
Characterization result of SI.
Using the PO ASIC described, two PO systems are developed.
SI input SI output SI output SI output The main difference between them is the wireless transmitter mod-
voltage (DC) voltage (DC) voltage (DC) voltage signal
ule. OxiSense-1 uses a standard bluetooth module (HC-06), whereas in
𝐕SI 𝐕SI_O_BUF before 𝐕SI_O_BUF after 𝐕SI_O_BUF slope
integration integration (×103 V/s) OxiSense-2 we have integrated our custom designed biotelemetry trans-
mitter [13]. The generalized block diagram of OxiSense was presented
970 mV 992 mV 1.67 V 2.607
earlier in Fig. 2. In Fig. 10 the chip micrograph of the PO ASIC is shown.
985 mV 992 mV 1.508 V 1.984
995 mV 992 mV 1.256 V 1.015 OxiSense-1 consists of a custom designed PCB which contains the
1V 992 mV 1.132 V 0.538 PO ASIC, a commercial micro-controller board (MSP430F5529 Launch-
1.01 V 992 mV 880 mV −0.431 pad [18]) from TI and a commercial PO probe (Nellcor DS100-A). The
1.02 V 992 mV 632 mV −1.385
micro-controller board provides the digital input signals to the FSM of
1.03 V 992 mV 384 mV −2.338
1.04 V 992 mV 136 mV −3.292 the PO ASIC and receives the analog output voltage signals from the PO
ASIC. The output signals (PPGs) corresponding to the red and infra red
channels are acquired by the micro-controller. The on-chip 12-bit ADC
of the micro-controller converts the analog signal to digital signal. The
linearly with time. Table 1 shows the measurement results for differ- signals acquired can be transmitted to a PC via bluetooth or USB. On-
ent values of VSI . From this it can be seen that as the input voltage VSI board buck boost circuits (TPS63000) convert the variable supply of the
moves closer to 998 mV, the slope of the output of SI decreases. Li-ion battery to fixed 1.8 V and 3.3 V. The battery can be charged using
Similarly, the testing of other analog module i.e. sample and hold the on-board battery charging circuit when OxiSense-1 is connected to
circuit and capacitive feed-back amplifier has been carried out success- a USB port. Fig. 11 shows OxiSense-1 housed in 3D printed casing, in

Fig. 11. PO system with USB and Bluetooth connectivity - OxiSense-1: (a) Housed in a 3D printed casing (b) Disassembled prototype (c) Python GUI on laptop (d) Assembled prototype
and PO probe (e) GUI on smartphone.

6
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

On the PC, the GUI has been developed using Python. Front-end of
the GUI displays the PPG along with S pO2 and HR measured, while
the back-end program stores the PPG signals in the database for further
processing, if required. The Python implementation of signal acquisi-
tion, processing and GUI was also replicated in form of an App on
Android platform. The Buildozer [19] tool chain within Python-for-
Android project [20] was used for this purpose, which ensures that
the entire software remains open-source. Since the tool chain is still
in development phase, not all Python libraries have been implemented
in Java and some parts of the Python code had to be re-written with-
out using more advanced Python libraries. However, all functionality
was retained as is between the PC and Android platforms. The Android
App communicates with the prototype using a Bluetooth link, imple-
mented through the Pyjnius module [21] which makes some Android
APIs accessible to Python.

5. Signal processing in OxiSense

In OxiSense, a novel PPG signal processing algorithm is implemented

Fig. 12. PO system with Bio-WiTel and USB connectivity - OxiSense-2.
on a low-power MCU (MSP430F5529). By virtue of this processing,
S pO2 and heart-rate are determined. Additionally an enhanced PPG
signal can be obtained if the recordings are made in the presence of
unavoidable motion artifacts such as measurements in ambulance or
dissembled condition along with the dimensions of the main compo- due to medical conditions like Parkinson’s disease. A typical PPG signal
nents and in working condition (not housed in casing to demonstrate in the absence and presence of motion artifacts is shown in Fig. 13a and
fully assembled working system). It also shows the graphical user inter- b, respectively.
face (GUI) on a PC and on an Android based smartphone connected to There are two modes of operation of the proposed signal process-
OxiSense via bluetooth. ing algorithm and they are: 1) embedded processing mode and 2) remote
OxiSense-2 integrated with the custom designed TX ASIC is shown processing mode. In the embedded processing mode, calculation of met-
in Fig. 12. TX of the Bio-WiTel system is based upon the principle of rics (S pO2 and HR) and signal enhancement if required, are performed
injection locking for reduced phase noise and edge combining for fre- on the MCU itself. The MCU transmits only the extracted features of
quency multiplication to achieve significantly low power consumption the PPGs (global and dicrotic notch maxima and minima correspond-
at sufficiently high data rate while meeting the spectral mask require- ing to each beat) and calculated S pO2 and HR to a suitable display
ment by Federal communication commission (FCC). Binary frequency device. In addition to low power processing, this also significantly
shift keying (BFSK) modulation is employed in the TX by pulling the reduces the required transmission power by 10–30x through lower bit-
anti-resonant frequency of a crystal oscillator before frequency multi- rate. In remote processing mode, all the processing is done on a lap-
plication. While radiating −18 dBm power from a 50 Ω antenna, TX top/smartphone/tablet and the MCU only transmits raw data. The pro-
consumes 250 μW power from 1 V supply in 100% on state. Moreover, cessing in this mode of operation is same as in embedded processing
it follows the spectral mask requirement at the maximum data rate of mode.
72 kb/s. RX of the Bio-WiTel system is based upon low intermediate In either of the modes of operation, the overall signal processing
frequency (IF) architecture. RX is designed by considering all system sequence is as follows: 1) digital filtering of the raw PPG signal of R and
level issues at 400 MHz as shown in Ref. [13]. For example, due to the IR channels, 2) feature extraction and signal enhancement (if required)
inevitable usage of off-chip inductors and capacitors in the tank circuit for R and IR channels and 3) metric computation (S pO2 and heart-
of the local oscillator (LO) of the RX, it is difficult to tune the LO fre- rate).
quency at a particular value. A slight deviation in LO frequency from
1. Digital filtering – Two sources of noise are commonly observed in
the desirable value leads to a change in IF value and hence bit error biomedical signals, low frequency baseline drift and high frequency
rate (BER) of the system degrades. To address this issue, RX is designed ambient noise (50 Hz interference and its harmonics [11]). A high
to have a programmable demodulator to achieve minimum BER for a sampling frequency of 303 Hz is used to sample the analog signal
particular IF. RX consumes 990 μW from 1.8 V supply and operates at providing SR , so that digital filtering can be used to remove high
max. data rate of 200 kb/s with a BER ≤ 10−4 .
frequency noise yielding the pre-processed signal S P .

Fig. 13. A typical recorded PPG signal (a) showing various waveform features extracted (b) in the presence of motion artifact.

7
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

Fig. 14. Measured PPG signals corresponding to R & IR channels of a subject (a) in resting condition: Raw signal (output of PO ASIC), Post-filtering signal (after digital filtering) (b) in
presence of intentionally induced motion artifacts by moving the finger vertically at a frequency of about 2–3 Hz. Also showing the enhanced signal reproduced using the 4 features
extracted per beat from filtered signal.

(a) Baseline drift removal through mean cancellation: AT (W) = max{S P [n], S P [n − 1], … , S P [n − W]}
(3)
S1 [n] = SR [n] − f (SR [n − 1], SR [n − 2], … , SR [n − l1 ]) − min{S P [n], S P [n − 1], … , S P [n − W]}
f → Simple average for embedded processing mode
(b) Secondly, from the buffered data in W, we extract and store
f → Median filter for remote processing mode
values of start and end points of all negative contiguous slope
(b) High frequency noise cancellation by FIR filtering: segments, as in Ref. [22]. We store these values in set, F, where
⎧ ∑
n
[] { }
ai .S1 i if n ≥ l2 , F = {ki , S p [ki ]}, {li , S p [li ]} i≥0 (4)
⎪
S p [n] = ⎨ i=n−l2 +1 (2)
⎪ In the negative slope segments Ni (i ≥ 0) the end points are as
⎩ S1 [n] if n < l2 . shown in Fig. 13(a) {ki , S p [ki ]} & {li , S p [li ]}. ki is a peak can-
ai →31-tap FIR filter coefficients with cutoff frequency of 10 Hz. didate, while li is a trough candidate. Following equations are
Except in cases of extreme motion artifacts, the above pre- used to remove spurious candidates from set F.
processing is sufficient to produce a clean PPG signal. Fig. 14(a) • Retain Ni if |ki − li | < 0.1 s and |S p [ki ] − S p [li ]| > 0.5 ∗ AT (W)
shows a typical raw and post digitally filtered PPG signal • Remove Ni+1 if |ki − ki+1 | < 0.3 s and |S p [ki ] − S p [li ]|
recorded using OxiSense. Fig. 14(b) shows the measurement of > |S p [ki+1 ] − S p [li+1 ]|
a typical PPG signal in the presence of motion artifact. As it is We search and store notch features as shown in Fig. 13(a), in
observed from Fig. 14(b), only digital filtering is not sufficient to the next step (c) using the retained trough candidates, li .
suppress motion artifacts. Hence, next step that is feature extrac- (c) Thirdly, a notch maximum is searched within a time-duration of
tion is performed. 0.1–0.3 s following each trough candidate, li . To keep the algo-
2. Feature extraction for signal enhancement – A simple algorithm rithm simple, only derivative is used and the point where the
is implemented in a low-power MCU to extract critical features i.e. derivative goes from positive to negative is selected as a poten-
beat and notch peaks and troughs (Fig. 13 (a)) from the processed tial notch maxima candidate. Assume this happens at m times
PPG signal, S p . The key principle of our feature extraction algo- during the duration of 0.1–0.3 s after the trough candidate, li .
rithm is to use contiguous positive and negative slope segments [22] We use a search algorithm to find maximum of the potential
to identify peak and trough candidates which can then be used to notch maxima candidates within the time duration of 0.1–0.3 s
estimate S pO2 and heart-rate metrics. Depending on nature of the and discard other points.
figure movement high frequency noise and/or distortion can get • Retain maximum of notch maxima points, that is
introduced in the PPG signal. An example of such a corruption is max{S p [kli1 ], S p [kli2 ], … , S p [klim ]} as notch maxima.
depicted in Fig. 13 (b). The algorithm utilizes a combination of fixed (d) Finally, a notch minimum is searched within a time duration
and adaptive thresholds to identify such motion artifacts. The algo- of 0.15 s following the previously identified notch maximum.
rithm is made agnostic to the polarity of hardware gain by identify- All local minima are identified in the duration of 0.15 s after
ing the direction of steepest slope segments. The signal processing notch maximum using negative derivative, and the location with
for feature extraction is as follows. lowest signal value is adjudged as notch minimum. Assuming
(a) Firstly, raw data is buffered in a window W, selected to be 2 s there are q local minimas. We search for the minimum of all
to incorporate at least 1 heartbeat. An adaptive threshold (AT ) the local minimas within the duration of 0.15 s after identified
is calculated dynamically over the window W as follows. notch maximum.

8
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

Fig. 15. A typical measurement set-up of OxiSense in hospital environment.

Fig. 17. Absolute deviation of readings with respect to CPO over 20 subjects (Each blue
box corresponds to 50% of total readings and the horizontal red line corresponds to
median value of the readings.). (For interpretation of the references to colour in this
• Retain minimum of notch minima points, that is figure legend, the reader is referred to the Web version of this article.)
min{S p [kli1 ], S p [kli2 ], … , S p [kliq ]} as notch minima.
We have identified the key features, beat and notch maximas and
minimas, in the post digital filtered signal. Using these features, (13 subjects). Fig. 15 shows a typical measurement set-up in hospital
S pO2 and heart-rate will be computed. If signal enhancement for environment. For reference, we used the CPO PX701 manufactured by
display of PPG is required, these feature points are connected Newnik Lifecare Pvt. Ltd., Bengaluru, India [23] which is CE certified
through a degree-3 spline-polynomial interpolation. A typical exam- and FDA approved. Comparison of the readings from OxiSense and the
ple is shown in Fig. 14b, where the red trace shows the enhanced reference CPO is shown in Fig. 16. It is clear that the output of OxiSense
PPG. closely follows the reference CPO with deviation of <2%.
3. Metric computation – S pO2 and heart-rate (HR) metrics are com- Fig. 17 shows the absolute deviation of the proposed PO with respect
puted as follows. to the CPO. The blue box limits are 25th and 75th percentile of the read-
|S p [k𝛼 ] − S p [l𝛼 ]|(R)
S pO2 = 108 − 20R − 0.375R2 , where R = |S p [k𝛼 ] − S p [l𝛼 ]|(IR)
, averaged ings. The red line inside the blue box is the median value. The readings
over 10 beats [4]. from the proposed PO have a maximum deviation of <2% of S pO2 .
HR = 60 × Sampling
k −k
frequency
, averaged over 10 beats. This is acceptable limit for medical applications [4,24]. The readings
i i−1
from embedded and remote processing modes of operation are found to
be same. To test the robustness of the proposed signal processing algo-
6. Measurement results
rithm, intentional motion artifacts were introduced during the measure-
ments. Motion artifact in the form of finger and hand movement at low
In order to determine an accurate relationship between S pO2 and R,
(∼1 Hz) and high (∼2–3 Hz) frequency were intentionally introduced.
calibration of the entire system is necessary. This requires artificial oxy-
Fig. 14 (b) shows one such PPG measurement in presence of motion
gen desaturation of humans which can be quite stressful to the subjects.
artifact. The proposed algorithm is able to accurately determine S pO2
As the objective of this work is not to produce a ready-to-sell commer-
and HR for non-severe artifacts. It was also observed that moving the
cial pulse oximeter, so instead we have used the S pO2 -R relationship
hand as a whole introduces much less artifacts since there is less mod-
as presented in Ref. [4] i.e. S pO2 = 108 − 20R − 0.375R2 . To bench-
ulation of blood flow in arteries due to the motion when compared to
mark the performance of the presented PO, OxiSense, measurements of
finger movement. Sever motion artifact effects PPGs in two different
S pO2 and HR was done on human subjects in the age group of 21–87
ways: 1) due to the relative motion between PO probe and finger and
years having S pO2 in the range of 84%–100% using a commercial pulse
2) due to modulation of arterial blood flow caused by the motion.
oximeter (CPO) and OxiSense in hospital (7 subjects) as well as in lab
The total power consumed by OxiSense-1 with the bluetooth mod-
ule being disabled is about 8 mW. The bluetooth module HC-06 con-
sumes about 160 mW of average power when searching for pairing. The
average power consumption is measured by powering OxiSense-1 from
a Keithley Source-meter 2400. The average power is then calculated
from the average current sunk. Table 2 shows the power consumption

Table 2
Power consumption by individual modules of OxiSense-1.

Power consumption Value (mW)

Custom PO ASIC 0.2

Analog buffer in ASIC 0.64
Op-amps 0.25
Level Translator ICs 0.9
Probe - LEDs 1.39
Total 3.38
MSP430 - Launch Pad (MSP - 4.51 mW (in Embedded processing mode
EXP430F5529L) for processing)
3.94 mW (in remote processing mode -
Fig. 16. Comparison of measured S pO2 values using CPO and OxiSense. comm. with PC via USB)

9
D.M. Das et al. Microelectronics Journal 72 (2018) 1–10

Table 3 Appendix A. Supplementary data

Comparison of system level power consumption.

Pulse oximeter Power consumption (mW) Supplementary data related to this article can be found at https://
doi.org/10.1016/j.mejo.2017.12.001.
This work 8a
[4] 4.8b
[25] 10 References
[26] 13.2
[27] 20.1 [1] https://www2.deloitte.com/content/dam/Deloitte/global/Documents/Life-
[10] 45.6 Sciences-Health-Care/gx-lshc-2015-health-care-outlook-india.pdf, 2016. (Online;
WristOx® 3100 ∼55 Accessed 26 April 2017).
Avant® 4000 71 [2] S. Salvi, K. Apte, S. Madas, M. Barne, S. Chhowala, T. Sethi, K. Aggarwal, A.
PalmSAT® 2500 ∼80 Agarwal, J. Gogtay, Symptoms and medical conditions in 204912 patients visiting
primary health-care practitioners in India: a 1-day point prevalence study (the
Onyx® 9500 ∼120
POSEIDON study), Lancet Glob. Health 3 (12) (2015) 776–784.
8500 Series ∼130
[3] J.G. Webster, Medical Instrumentation - Application and Design, third ed., Wiley,
a Including power for digital processing. 2009.
b
Excluding power for digital processing. [4] M. Tavakoli, L. Turicchia, R. Sarpeshkar, An ultra-low-power pulse oximeter
implemented with an energy-efficient transimpedance amplifier, IEEE Trans.
Biomed. Circuits Syst. 4 (2010) 27–38.
[5] K. Glaros, E. Drakakis, A sub-mW fully-integrated pulse oximeter front-end, IEEE
break-up of the individual modules. Table 3 shows that the presented Trans. Biomed. Circuits Syst. 7 (2013) 363–375.
prototype OxiSense is comparable with some of the recent works from [6] S. Gubbi, B. Amrutur, Adaptive pulse width control and sampling for low power
pulse oximetry, IEEE Trans. Biomed. Circuits Syst. 9 (2015) 272–283.
power consumption standpoint. It should be noted that this data also [7] Z. Lu, X. Chen, Z. Dong, Z. Zhao, X. Zhang, A prototype of reflection pulse
depends on the probe characteristics. Table 3, compares OxiSense with oximeter designed for mobile healthcare, IEEE J. Biomed. Health Inf. (2015) 1–1.
different commercially available low-power POs from only power dissi- [8] C. Scully, J. Lee, J. Meyer, A. Gorbach, D. Granquist-Fraser, Y. Mendelson, K.
Chon, Physiological parameter monitoring from optical recordings with a mobile
pation standpoint. These numbers exclude display power and includes
phone, IEEE Trans. Biomed. Eng. 59 (2012) 303–306.
only LED and processing power [4]. In Ref. [6], the power consumption [9] P. Reddy, A. Kumar, S. Rahman, T. Mundra, A new antispoofing approach for
claimed is only 0.8 mW. However, it depends heavily on the PLL track- biometric devices, IEEE Trans. Biomed. Circuits Syst. 2 (2008) 328–337.
[10] R. Haahr, S. Duun, M. Toft, B. Belhage, J. Larsen, K. Birkelund, E. Thomsen, An
ing algorithm which is only tested on 3 subjects in controlled laboratory
electronic patch for wearable health monitoring by reflectance pulse oximetry,
environment. Worst case power consumption in that case is 16 mW. IEEE Trans. Biomed. Circuits Syst. 6 (2012) 45–53.
Hence, it not included in the comparison table. Except this work, none [11] K. Li, S. Warren, A wireless reflectance pulse oximeter with digital baseline control
of the other POs in Table 3 are IoT enabled. for unfiltered photoplethysmograms, IEEE Trans. Biomed. Circuits Syst. 6 (2012)
269–278.
[12] M. Zare, M. Maymandi-Nejad, A fully digital front-end architecture for ECG
7. Conclusion acquisition system with 0.5 V supply, IEEE Trans. VLSI Syst. 24 (2016) 256–265.
[13] A. Srivastava, N. Sankar, B. Chatterjee, D. Das, M. Ahmad, R. Kukkundoor, V.
Saraf, A. Jayachandran, D. Sharma, M. Baghini, Bio-Witel: a low power integrated
In this work, we have designed and developed a fully integrated PO wireless telemetry system for healthcare applications in 401–406 MHz band of
ASIC in 180 nm CMOS technology. Considering design for testability, MedRadio spectrum, IEEE J. Biomed. Health Inf. (2016) 1–1.
[14] www.fcc.gov/general/medical-device-radiocommunications/-service-medradio,
we are able to test the functionality of individual modules of the PO
2016. (Online; Accessed 26 April 2017).
ASIC. We envisage that this work will serve as a guiding example for [15] M. Shokouhian, R. Morling, I. Kale, Interference resilient sigma delta-based pulse
developing complex and fully integrated biosignal measurement ASICs. oximeter, IEEE Trans. Biomed. Circuits Syst. 10 (2016) 623–631.
[16] N.S. Kim, T. Austin, D. Baauw, T. Mudge, K. Flautner, J.S. Hu, M.J. Irwin, M.
By considering testability during the initial design phase itself, it can
Kandemir, V. Narayanan, Leakage current: Moore’s law meets static power,
potentially reduce ASIC development time, since from the first tape- Computer 36 (2003) 68–75.
out, post-silicon behavior of individual blocks can be determined. Our [17] Spartan-3E FPGA Starter Kit Board User Guide, Xilinx, 2011. Ver. 1.2.
PO ASIC is comparable with state-of-the-art designs and consumes only [18] MSP430F5529 LaunchPad Development Kit, User Guide, Texas Instruments, 2015.
[19] Buildozer Documentation, 2016, http://buildozer.readthedocs.io/en/latest/
about 200 μW. Using the PO ASIC, we further designed a PO system, quickstart.html[Accessed 26 April 2017].
OxiSense. By employing a novel signal processing on the raw measure- [20] Python-for-Android Project, 2016. https://github.com/kivy/buildozer[Accessed 26
ment signals, SpO2 and heart rate information are extracted. The signal April 2017].
[21] Pyjnius Documentation, 2016, http://pyjnius.readthedocs.io/en/latest/android.
processing algorithm is simple and it can be implemented on a low- html[Accessed 26 April 2017].
power micro-controller itself. It is also immune to small motion arti- [22] W. Karlen, J.M. Ansermino, G. Dumont, Adaptive pulse segmentation and artifact
facts. detection in photoplethysmography for mobile applications, in: Engineering in
Medicine and Biology Society (EMBC), Annual International Conference of the
IEEE, 2012.
Acknowledgment [23] Newnik PX701 Fingertip Pulse Oximeter, 2016, http://www.newnik.com/
products/pulse-oximeter/[Accessed 26 April 2017].
[24] J.G. Webster, Design of Pulse Oximeters, first ed., IOP Publishing Limited, Bristol,
We acknowledge MeitY (Ministry of Electronics and Information
U.K., 1997.
Technology), Government of India for sponsoring the project and Texas [25] B. Sueppel, D. Mortara, Low Power Pulse Oximeter, 2004.
Instruments for providing developmental tools. We also acknowledge [26] J. Patterson, G.-Z. Yang, Ratiometric artifact reduction in low power reflective
Europractice for providing chip fabrication service and Dr. Nambiraj photoplethysmography, IEEE Trans. Biomed. Circuits Syst. 5 (2011) 330–338.
[27] F. Adochiei, C. Rotariu, R. Ciobotariu, H. Costin, A wireless low-power pulse
Konar, Dept. of Anaesthesia, B.K.L. Walawalkar Hospital, for recording oximetry system for patient telemonitoring, in: 7th International Symposium on
facility. Advanced Topics in Electrical Engineering, 2011, pp. 1–4.

10