HHS Public Access

Author manuscript
Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.
Author Manuscript

Published in final edited form as:

Biomed Microdevices. ; 24(4): 36. doi:10.1007/s10544-022-00636-w.

Digital filtering dissemination for optimizing impedance

cytometry signal quality and counting accuracy
Brandon K Ashley,
Department of Biomedical Engineering, Rutgers, the State University of New Jersey, Piscataway,
NJ, 08854

Umer Hassan*
Author Manuscript

Department of Electrical Engineering, Department of Biomedical Engineering, and Global Health

Institute Rutgers, the State University of New Jersey, Piscataway, NJ, 08854

Abstract
Improving biosensor performances which utilize impedance cytometry is a highly interested
research topic for many clinical and diagnostic settings. During its development, a sensor’s
design and external factors are rigorously optimized, but improvements in signal quality and
interpretation are usually still necessary to produce a sensitive and accurate product. A common
solution involves digital signal processing after sample analysis, but these methods frequently
fall short in providing meaningful signal outcome changes. This shortcoming may arise from
a lack of investigative research into selecting and using signal processing functions, as many
choices in current sensors are based on either theoretical results or estimated hypotheses. While
Author Manuscript

a ubiquitous condition set is improbable across diverse impedance cytometry designs, there lies
a need for a streamlined and rapid analytical method for discovering those conditions for unique
sensors. Herein, we present a comprehensive dissemination of digital filtering parameters applied
on experimental impedance cytometry data for determining the limits of signal processing on
signal quality improvements. Various filter orders, cutoff frequencies, and filter types are applied
after data collection for highest achievable noise reduction. After designing and fabricating a
microfluidic impedance cytometer, 9 μm polystyrene particles were measured under flow and
signal quality improved 6.09 dB when implementing digital filtering. This approached was
then translated to isolated human neutrophils, where similarly, signal quality improved 7.50 dB
compared to is unfiltered original data. By sweeping all filtering conditions and devising a system
to evaluate filtering performance both by signal quality and object counting accuracy, this may
serve as a framework for future systems to determine their appropriately optimized filtering
Author Manuscript

configuration.

Introduction
Highly sensitive and accurate biosensors are continuously researched and sought out to
measure critical and often microscopic analytes, providing decisive and time-dependent
information in biological systems. This imperative form is used in diverse settings, from

*
Corresponding Author: umer.hassan@rutgers.edu.
 Ashley and Hassan Page 2

critical care disease diagnostics (Ashley and Hassan 2021a; C. Murdock et al. 2017;
Author Manuscript

Lu et al. 2015b; Zhang et al. 2020), continuous and discrete physiological monitoring
(Ashley et al. 2019; Brown et al. 2018; Gao et al. 2016; Koh et al. 2016), environmental
monitoring (Alam et al. 2020; Lu et al. 2015a; Marinov et al. 2018), and manufacturing/
product development quality control (Cinti et al. 2017; Izadi et al. 2016; Verma and Singh
2003). The assorted materials and compounds to measure additionally prompts a variety
of detection modalities in which to measure them, each with their advantages in specific
conditions. This includes modalities such as fluorescence microscopy (Hu et al. 2014;
Maetzig et al. 2017; Volpetti et al. 2015), acoustic imaging (Gnyawali et al. 2019; Khateib
et al. 2020; Sarimollaoglu et al. 2014), surface plasmon resonance (Sun et al. 2020; Yoo et
al. 2020), and biochemical assays (Cho and Irudayaraj 2013; Han et al. 2007). Among the
vast technologies, electrochemical impedance spectroscopy (EIS) stands out as a versatile,
accurate, and rapid technique and has demonstrated minimal reagent preparation, miniscule
Author Manuscript

sample volumes, and non-destructive sample measurement for a variety of materials (Bredar
et al. 2020; Hassan et al. 2017; Lu et al. 2015b; Stupin et al. 2017). In a subsection of
EIS, microfluidic impedance cytometry is the recordings of microscopic materials flowing
through a microfluidic channel and disrupting a defined electric field, and has been used
extensively for its manufacturing ease and simple data processing in point-of-care settings
(Ashley and Hassan 2021b; Clausen et al. 2018; Colson and Michel 2021; Zhong et al.
2021).

Unfortunately, each of these choices comes with their own competitive unvalued noise
sources which impedes the intended analyte’s detection. For EIS, this originates from
interferences such as surrounding electrical generation, sample volume electrical properties,
external ionization, and sample collection quantization (Antal et al. 2001; Pierce et al.
2015; Ram et al. 2012). Perfecting experimental conditions may reduce noise from these
Author Manuscript

sources, but in many fields these conditions are uncontrollable and may not reduce noise to a
sufficient degree (Ridgway et al. 2007).

Without alternatives, a capable approach is processing post-data collection, where property

differences between signal and noise data may be exploited (Hassan et al. 2015).
Transforming the data into frequency space reveals this exploitation, where specific
frequency regimes are notorious for predominate noise. This includes low frequencies
typically below 30 Hz, where high amplitude, low frequency noise manifests into data drift
over time, corrupting the detection range and primarily disrupting accurate desired signal
recognition (Atakan et al. 1980; Chouhan and Mehta 2007; Liu et al. 2014; Sun et al. 2007).
Contrastingly, high frequencies above 100 kHz also constitute majority noise results, as
frequency data from material recordings generally fall within 0.03 to 100 kHz (Ashley and
Author Manuscript

Hassan 2021c; Hassan et al. 2015). This high frequency noise produces a wide noise band
in the data, where signal identities close to a device’ detection limit may fall underneath and
become degraded (Ashley and Hassan 2021b). While these are the general values, in practice
the specific frequencies which represent analyte data has not been fully explored, being only
defined from theoretical hypotheses (Antal et al. 2001; Stupin et al. 2017). There lies a need
for understanding and experimental justification for specific frequency filtering which may
improve sensor data signal quality and accuracy.

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 3

Often, digital filtering is used with many detection modalities, including EIS (Hassan et al.
Author Manuscript

2015; Manikandan and Soman 2012; Ram et al. 2012; Redhyka et al. 2015). However, when
critical materials must be measured near the technology’s detection limit, the smallest of
margins and signal quality optimizations may correspond to significant detection accuracy
improvements. While applied, there is inadequate research on experimental impedance
cytometry data related to which digital filtering conditions have the highest performance.
Therefore, this article aims to execute a library of digital filtering conditions and provide
a structure for defining signal quality in impedance cytometry data. Polystyrene particles 9
μm in diameter are measured through a microfabricated, microfluidic impedance cytometer
(Fig. 1a, 1b). As shown by Figure 1c, when assessing filtering performance, emphasis
will be placed on counting accuracy first, followed by higher signal-to-noise ratios (SNR)
related to higher signal quality. After transimpedance amplification, a litany of filter types,
filter orders, and cutoff frequencies will be assessed in transforming the original unfiltered
Author Manuscript

result (Fig. 1d) to its defined, most optimized form (Fig. 1e). This approach will further
be presented with biologically relevant samples of isolated human neutrophils to support
its formulation. While systematically defined, this article is the first to explore digital
filtering variations in experimental EIS data and provide guidance on future and more
diverse filtering selections.

Materials and methods

Materials
Phosphate buffered saline (PBS, 1X and 10X), Ficoll-Paque density gradient, (3-Amino-
propyl)triethoxysilane (APTES), and Roswell Park Memorial Institute medium 1640
(RPMI) were purchased through Sigma Aldrich (St. Louis, MO, USA). A NE-300 syringe
pump was purchased from Southpoint Surgical Supply (Coral Springs, FL, USA). A HF2LI
Author Manuscript

lock-in amplifier and HF2TA current amplifier was purchased through Zurich Instruments
(Zurich, SUI). Silver conductive epoxy was purchased from Digi-Key (Thief River Falls,
MN, USA). Unidentifiable human blood was obtained from Robert Wood Johnson Medical
Hospital (New Brunswick, NJ, USA) through an institutional review board (IRB) study.
LabView software was purchased and installed through National Instruments (Austin, TX,
USA). MATLAB version 2020B was purchased and installed through Mathworks (Natick,
MA, USA).

Microelectrode and Microchannel Fabrication

Gold-electrodes are microfabricated on borosilicate glass wafers above a deposited
chromium adhesion layer following the manufacturers protocol for s1813 photolithography.
Microfluidic channel molds are also microfabricated on silicon wafers following the
Author Manuscript

manufacturers protocol for SU-8 3025 photolithography. Polydimethylsiloxane (PDMS) is

poured and cured over SU-8 channel molds, and after removal a biopsy punch creates
inlet and outlet holes to facilitate media perfusion from syringe tubing and a syringe
pump. PDMS and gold electrode surfaces are treated with oxygen plasma and bonded,
where focusing regions are aligned across electrodes to increase polystyrene particle
volume fraction in the electric field space. Key channel dimensions include 100 μm in
width, with focusing regions reducing the width to 30 μm, 22 μm in height, and is

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 4

3 cm long. Microelectrodes are 0.5 μm thick, are 100 μm in width, and are 150 μm
Author Manuscript

apart (Fig. 1b) (Holmes et al. 2006). More specific microfabrication procedures with
this configuration have been previously published, and further detail is included in the
Supplemental Information (Ashley et al. 2021; Ashley and Hassan 2021b).

Signal Acquisition, Demodulation, and Data Processing

Silver conductive epoxy adheres gold electrodes to a custom printed circuit board using
electrode contact pads, visualized in Figure 1a. A 5V AC input is applied to the center
electrode using the HF2LI signal amplifier, and the HF2TA current amplifier increase
recorded signal contributions from the device. 9 μm polystyrene microparticles are diluted
in phosphate buffered saline at a 35 particle/μL concentration and undergo flow at a rate
of 15 μL/min. The signal is recorded at a 250 kHz sampling rate using a PCIe-6361
data acquisition card, and dual detecting (grounded) electrode recordings are subtracted to
Author Manuscript

produce a bipolar impedance pulse.

Impedance cytometry recordings are saved through a LabView control program and
interpreted in a custom MATLAB script. A region manually selected which is absent of
particles is used as the noise reference values, which is the standard deviation of the selected
region. A threshold is then applied for the absolute value of data 4 times this noise standard
deviation to define the particle pulse, construing the data as 9 μm polystyrene microparticles.
If the average of the threshold triggering data point and the next 10 data points is less than
the threshold, the pulse is neglected, which removes high frequency noise data that may have
triggered a false positive hit above the threshold. Additionally, after collecting all the particle
pulses and obtaining their average peak-to-peak amplitudes, pulses which are 2 times greater
than average are neglected, as these represent two or more particles traveling across the
electric field regime simultaneously. For this study, we are only interested in evaluating
Author Manuscript

the signal quality for individual 9 μm polystyrene microparticles, as more than one particle
disrupting the electric field at once may artificially inflate the average signal quality of the
sample. More detail on instrumentation and microfluidic channel framework may be found
in the Supplemental Information (SI Fig. 1).

Applying and evaluating digital filtering

Using MATLAB, when a data value goes beyond the defined threshold, the magnitudes
of the highest and lowest values of the next 500 data points are combined to produce the
peak-to-peak bipolar amplitude (ΔVT) for an individual particle:

ΔV T = ΔVmax − ΔVmin (Equation 1)

Author Manuscript

Following this, the standard deviation of the previous 200 data points is measured which
defines the local noise prior to particle pulse detection:

1
n∑
σBG = xi2 (Equation 2)
i

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 5

The signal-to-noise ratio for each particle is then quantified as the logarithm ratio of bipolar
Author Manuscript

amplitude with data standard deviation prior to pulse detection, standardized to a decibel
scale:

ΔV T
SNR = 20 log (Equation 3)
σBG

For a 60 second recording segment and from the expected diluted concentration of 35
particles/μL which flows through the channel at 15 μL/min, the expected particle count
is approximately 525 particles. Accounting for a 10% error given micro-changes in
concentration under flow, different filtering conditions described next are rated by increasing
SNR if the counted number of particles is within this 10% error range (Fig. 1c). If the
recorded particle counts deviate 10% away from the 525-particle count average, filter
Author Manuscript

conditions are rated by the closest counts to expected, as these conditions are rated less
favorably since many false positive or false negative measurements occur, indicating an
unoptimized system. Future experiments are underway with simultaneous high speed video
microscopy to capture true object incidence, which can reduce the % error range for defining
counting accuracy.

The original unfiltered data is shown in Fig. 1d, with examples of baseline drift represented
along with an apparent noise band. Dual detecting electrode recordings are subtracted to
produce a bipolar impedance pulse shown by the zoomed-in insert. Butterworth, Chebyshev
Type I (Cheby1), and Chebyshev Type II (Cheby2) digital filters are applied in MATLAB
for both high pass and low pass filtering conditions. Both Cheby1 and Cheby2 filters are
operating with a 5 dB ripple for the passband (Cheby1) or stopband (Cheby2). For each
experiment, powerline interference filtering is also applied with 4th order band-stop filters
Author Manuscript

at 60 Hz and its 120 Hz second harmonic. Figure 1e represents the most optimized filtering
conditions achieved with this approach of the 9 μm polystyrene particle samples, revealing a
smaller noise-band and removed baseline drift, and will be determined and discussed in the
following sections. Data Fourier transformations is performed in MATLAB using the Fast
Fourier Transform or fft command.

From previous reports, particle pulse data collected from micro flowing impedance
cytometry lies between 50 Hz and 90 kHz when flowing at an approximate 15 μL/min
rate (Hassan et al. 2015). Therefore, a significant majority of amplitudes above 90 kHz is
contributed from high frequency noise and produces the large time-domain noise band. It
is expected that eliminating frequencies near or above this 90 kHz value may reduce this
band and improve signal quality without significantly reducing desired object pulse data.
Author Manuscript

Thus, for cutoff frequency selection for digital filtering in the subsequent sections, cutoff
frequencies are varied from 5 Hz to 50 Hz at different filter orders with a 5 Hz step for high
pass filtering. Cutoff frequencies are varied from 60 kHz to 125 kHz with a 15 kHz with low
pass filtering. Along with filter types, filter orders for each type with each filtering pass are
modulated from 1st to 4th order and disseminated to determine the highest achievable SNR
with measured particle counts close to expected particle counts. Isolating neutrophils from
whole blood

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 6

Deidentified whole blood was obtained from patient samples and Robert Wood Johnson
Author Manuscript

University Hospital through an IRB study (Wagner et al. 2021). After collection, blood
was combined with 1X PBS at equal volumes, and layered above Ficoll-Paque density
gradient at a 3 to 4 ratio. This amalgam is centrifuged for 30 mins at 400 g which exploits
density differences to separate platelets, blood cells, and plasma. Platelets and plasma are
aspirated in the supernatant, with the blood cell pellet exposed to deionized water for 15
seconds to dissolve non-neutrophil mononuclear cells. Tonicity was rebalanced with 10X
PBS, and the solution was again centrifuged for 5 mins at 300g to separate red blood
cells from neutrophils. This process is repeated until a gray pellet appears, representing
isolated neutrophils, and this pellet was resuspended in RPMI 1640 media with 50 μL of
stock neutrophils to 5 mL of RPMI 1640 media. Immediately prior to impedance cytometry
experiments, neutrophils are diluted in 1X PBS.
Author Manuscript

Results and discussion

Data Fourier transformation to determine filtering ranges
Figure 2 outlines the Fourier transform (Fig. 2b, 2c) of the entire unfiltered data (Fig.
2a), highlighting specific regimes which contain significant noise contributions which are
evaluated in this report, including low frequency noise (Fig. 2d) and high frequency
noise (Fig. 2e). The goal with cutoff frequency selections is to produce the highest noise
elimination without additionally eliminated frequency data from particle pulses. Visualized
by Figure 2d, baseline drift is a large frequency amplitude contribution below 5 Hz which
arises from external ionization of surrounding electrical devices and micro-variances in
microfluidic flow (Pierce et al. 2015). These observations affirm the cutoff frequencies used
for both high and low frequency noise filtering for the following sections to improve signal
quality and accuracy.
Author Manuscript

Impact of high pass filtering on signal quality

As mentioned previously, eliminating low frequency noise reduces baseline drift effects, and
predicted to increase object counting accuracy and modestly decrease the noise amplitude
(Atakan et al. 1980; Liu et al. 2014). When assessing measurement performance for
biosensors including this impedance cytometry device, the most important metrics include
high SNRs and accurate object determination. Regarding these experiments, SNR is closely
tied to the bulk noise amplitude, while accurate object counting is related to the expected
object concentration in the sample. Therefore, the number of particles measured in a
constant time measurement, the samples noise amplitude, and the composite SNR will
define the improvements in signal quality brought about by the digital filtering conditions.
When no digital filtering is applied, the referenced sample metrics includes a 0.2468 V noise
Author Manuscript

amplitude, a 20.54 dB SNR, and counted 971 particles which is significantly inaccurate to
the expected 525 particle counting concentration.

Figure 3 details the noise, SNR, and number of 9 μm polystyrene particles counted in
the same sample data with a myriad of high pass digital filtering conditions used without
other simultaneous digital filtering. What remains consistent across filter types is a higher
likelihood of reduced counting accuracy at higher cutoff frequencies (see filter orders 1

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 7

and 2 in Fig. 3a, filter order 2 in Fig. 3d, and filter orders 1 and 3 in Fig. 3g). This
Author Manuscript

sharp rise in false positive counts results from removing particle pulse frequency data in
this low frequency regime and producing a thresholding value less likely to discern pulse
data from high amplitude noise data. As a result, inaccurate and over-counting trickles
down to impact average samples noise amplitude, and with a reduced average bipolar pulse
amplitude measured that includes lower false positive noise amplitude, the SNR is lower
relative to other conditions (see filter orders 1 and 2 in Fig. 3c, filter order 2 in Fig. 3f,
and filter orders 1 and 3 in Fig. 3i). It would be expected then that the most optimal cutoff
frequency across the filters would be the highest value that does not trigger an inaccurate
count and thereby eliminating baseline drift noise to its fullest degree. Otherwise, there are
no significant trends between filter orders which dictate filtering roll-off steepness, while
Cheby1 filters have a marginally higher average SNR than Butterworth or Cheby2 filters at
the same cutoff frequency and filter order.
Author Manuscript

Using the ranking system outlined in Figure 1c, each filter order, cutoff frequency, and
filter type iteration was arranged in a hierarchy from most to least optimized to improve
signal quality (Table 1 and SI Table 1). Based on noise reduction, relative SNR increases,
and counting particles within a 10% margin of error to the expected counts, the high pass
filtering alone which improved signal quality the most was the 3rd order Cheby1 filter with
a 30 Hz cutoff frequency. This aligns with our expected midpoint cutoff frequency value
between 5 and 50 Hz, delivering the highest SNR of the over 100 different filter parameter
combinations (SI Table 1). The highest performing Butterworth filter was ranked 11th,
which was a 3rd order filter with a 15 Hz cutoff frequency, while the highest performing
Cheby2 filter was ranked 39th with a 4th order filter at 10 Hz. Of note, Cheby1 filters
typically outperformed Butterworth and Cheby2 filters with similar conditions, and were
less likely to experience the inaccurate counting issue as Cheby2 filters made up 10 of the
Author Manuscript

15 configurations which counted particles outside of the 10% margin of error (SI Table 1).
Otherwise, there were no apparent trends related to filter order throughout the ranks.

Impact of low pass filtering and the most optimized high pass filtering on signal quality
While high-pass filtering primarily eliminates baseline drift, low-pass filtering is expected
to greater reduce the noise band and increase SNR. When evaluating signal quality alone,
however, the lack of baseline drift negation heavily impairs the counting accuracy and signal
quality achieved from even the most optimized low pass filtering conditions (SI Fig. 2). As
such, the next results will focus primarily on modulating low-pass filtering conditions in
conjunction with the most optimized high-pass filtering parameters achieved in the previous
section (3rd order Cheby1 filter with 30 Hz cutoff frequency).
Author Manuscript

In evaluating low-pass filtering effects, Figure 4 presents trends with cutoff frequency across
different filter orders and filter types. Similar to Figure 3, the graphs are presented as the key
metrics in determining signal quality; the number of counted particles, average background
noise amplitude, and SNR. Other than 2 iterations with the 4th order Cheby1 filter below 75
kHz, all configurations were within 10% of the counting error and had adequate counting
accuracy. Unlike the high-pass alone conditions, there is a trend with cutoff frequency and
SNR/noise amplitude, as across the filters and orders a lower cutoff frequency corresponded

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 8

with a lower noise amplitude and higher SNR. This is consistent with the ranked filtering
Author Manuscript

conditions (Table 2), as each filter type’s highest rank was with a 60 kHz cutoff frequency.
For these parameters, the highest rank overall used a 1st order Butterworth filter, while
the highest Cheby1 filter was ranked 4th and Cheby2 filter ranked 7th. Contrasting to the
high-pass filtering results alone, there is an even distribution of filter types in the top ranks
of low-pass filtering signal quality (SI Table 2). However, filter order results did not have
consistent relationships to filter type, as cutoff frequency dominated the correlation with
signal quality ranks. With this combinatorial filtering, SNR increased 6.09 dB and placed the
device within an acceptable counting range compared to the unfiltered data.

It was previously believed that frequencies below 90 kHz for this device’s configuration
also included considerable desired signal components, and therefore cutoff frequencies for
a low pass filter below 90 kHz would decrease signal quality relative to a higher cutoff
frequency (Hassan et al. 2015). However, the results presented here reflects a greater noise
Author Manuscript

reduction offset down to 60 kHz, even as particle bipolar amplitude slightly declines as well
(Fig. 1d). This proves the power noise reduction has on SNR relative to the measured signal
amplitude, and future directions may research filtering conditions which go below this 60
kHz low pass cutoff frequency as well to determine optimal signal quality.

In both high-pass and low-pass filtering analysis, the Cheby1 filter generally outperformed
the Butterworth and Cheby2 filters in noise reduction at identical filter order and cutoff
frequency conditions. This is expected, as ideal Cheby1 filters have a larger stopband
attenuation magnitude farther halfway through the stopband frequency range compared to
a Butterworth filter with the same cutoff frequency and filter order (Taylor and Williams
2006; Weinberg and Slepian 1960). Additionally, the Cheby2 filter performed the worst in
attenuating baseline drift, most likely due to the ripple in the stopband inhibiting greater
Author Manuscript

noise reduction as opposed to the ripple being in the passband for Cheby1 filters (Bianchi
and Sorrentino 2007; Singh et al. 2010). The severity of this effect in the low-pass results
was not as apparent, as the stopband at lower cutoff frequencies also included polystyrene
particle pulse data. One consideration may be computational time for applying digital
filters, as for identical conditions Butterworth filters have a faster step response (Singh
et al. 2010). However, for the reported analysis, additional time for digital filtering did
not exceed 3 seconds for any experimental condition, including combinatorial high-pass
and low-pass experiments. When optimal filtering conditions are determined and real-time
filtering is applied to future experiments, computer memory buffers settings may be applied
to discretize smaller data sections and accommodate for the required filter computing time.

Applying filtering conditions to isolated neutrophil data

Author Manuscript

To evaluate the procedural robustness of optimizing digital filtering conditions for time-
domain impedance cytometry data, this approach was translated to more heterogenous
samples with greater biomedical applications. Specifically, isolated neutrophils from human
whole blood were measured in the designed microfluidic impedance cytometer, and the
same iterative process was used to determine the low and high pass digital filtering
conditions that generated the highest SNR while also remaining within ± %10 of the
expected number of neutrophils counted across the sample recording. Here, neutrophils were

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 9

diluted to 5×104 cells/mL, and over a 60 second recording with the same 15 μL/min the
Author Manuscript

expected number of cells to count was 750.

Figure 5 represents the time-domain impedance data of counted neutrophils measured with
the microfluidic impedance cytometer before (Fig. 5a) and after (Fig. 5b) digital filtering
was applied. Here, there is a more apparent baseline drift consideration, although this did
not greatly impact the neutrophil counting results the number of counted cells at 769 was
still within ± %10 of the expected 750 count. Additionally, only a select number of filters
were evaluated based on filter conditions from this group’s previous publications as well as
the highest-performing filter conditions from the previous sections in this article (Ashley and
Hassan 2021b; Hassan et al. 2015). It is notable that our previous digital filtering conditions
delivered the lowest increase in SNR (4th order Butterworth high pass filter with a 20 Hz
cutoff, 4th order Butterworth low pass filter with a 120 kHz cutoff), and changes in filter
cutoff frequencies appeared to have most significant SNR impacts.
Author Manuscript

Many—but not all—trends remain consistent to the 9 μm polystyrene test particle results.
This includes lower low-pass filtering cutoff frequencies between 60 to 75 kHz delivered
higher SNRs, the 25–30 Hz cutoff frequency range for high-pass filtering was optimal to
remove baseline drift without greatly disrupting neutrophil signal quality, and the Cheby1
filters were found higher ranking than Butterworth or Cheby2 filters (Table 3). However, the
most optimal filtering conditions for the neutrophil data was not identical to the previous
section results: here, the most optimal filtering conditions were a 3rd order Cheby1 high-pass
filter with a 30 Hz cutoff and a 3rd order Cheby1 low-pass filter with a 60 kHz cutoff, while
the filtering conditions most optimal for the 9 μm polystyrene test particle results ranked
3rd in signal quality for the neutrophil data. This most likely arises from the small margins
which separate all these conditions, as only a 0.11 dB change in SNR exists between
Author Manuscript

these two conditions, so the smallest impacts in bipolar amplitude frequency data between
neutrophils and polystyrene particles, flow perturbations, or local media conductivity may
change the final numerical result. Nonetheless, the aim of these results is to highlight the
methodology for determining the most optimal filtering conditions and which signal quality
metrics contribute to those decisions.

It has been demonstrated that a lack of experimental analysis produced sub-ideal signal
processing, and the framework put forth in this article can provide clarity for other systems
to find their own signal processing best fits. This digital filtering approach may be used
with minimal modulation, including for systems of varying experimental or instrumental
forms, such as different channel dimensions, different electric field magnitudes, or using
single-ended measurements. While exhaustive filtering conditions were highlighted, future
Author Manuscript

applications of this method can omit some steps like including Cheby2 filters or extreme
cutoff frequency ranges to evaluate filtering results more efficiently. Collecting the highest
feasible signal quality in a biomedical device is greatly important for measuring minute but
critical analyte changes and lowering a systems detection limit, which in turn can increase
the sensitivity and accuracy of these highly depended-upon machines.

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 10

Supplementary Material
Author Manuscript

Refer to Web version on PubMed Central for supplementary material.

Funding information
Research supported by the National Institute of Health T32 GM135141, the National Science Foundation Grant
Number 2002511, and Rutgers University School of Engineering.

References
Alam AU, Clyne D, Jin H, Hu N-X, and Deen MJ, ACS Sens. 5, 412 (2020). [PubMed: 32028771]
Antal T, Droz M, Györgyi G, and Rácz Z, Phys. Rev. Lett 87, 240601 (2001). [PubMed: 11736489]
Ashley BK, Brown MS, Park Y, Kuan S, and Koh A, Biosensors and Bioelectronics 132, 343 (2019).
[PubMed: 30897541]
Ashley BK and Hassan U, WIREs Nanomedicine and Nanobiotechnology 13, e1701 (2021a).
Author Manuscript

[PubMed: 33650293]
Ashley BK and Hassan U, Biotechnology and Bioengineering 118, 4428 (2021b). [PubMed:
34370302]
Ashley BK and Hassan U, in 2021 43rd Annual International Conference of the IEEE Engineering in
Medicine Biology Society (EMBC) (2021c), pp. 1201–1204.
Ashley BK, Mukerji I, and Hassan U, in 2021 43rd Annual International Conference of the IEEE
Engineering in Medicine Biology Society (EMBC) (2021), pp. 7233–7236.
Atakan AK, Blass WE, and Jennings DE, Appl Spectrosc 34, 369 (1980).
Bianchi G and Sorrentino R, Electronic Filter Simulation & Design (McGraw Hill Professional, 2007).
Bredar ARC, Chown AL, Burton AR, and Farnum BH, ACS Appl. Energy Mater 3, 66 (2020).
Brown MS, Ashley B, and Koh A, Front. Bioeng. Biotechnol 6, (2018).
Cho I-H and Irudayaraj J, International Journal of Food Microbiology 164, 70 (2013). [PubMed:
23603219]
Author Manuscript

Chouhan VS and Mehta SS, in 2007 International Conference on Computing: Theory and Applications
(ICCTA’07) (2007), pp. 512–515.
Cinti S, Basso M, Moscone D, and Arduini F, Analytica Chimica Acta 960, 123 (2017). [PubMed:
28193355]
Clausen CH, Dimaki M, Bertelsen CV, Skands GE, Rodriguez-Trujillo R, Thomsen JD, and Svendsen
WE, Sensors (Basel) 18, (2018).
Murdock RC, Gallegos KM, Hagen JA, Kelley-Loughnane N, Weiss AA, and Papautsky I, Lab on a
Chip 17, 332 (2017). [PubMed: 27966711]
Colson BC and Michel APM, ACS Sens. 6, 238 (2021). [PubMed: 33423457]
Gao W, Emaminejad S, Nyein HYY, Challa S, Chen K, Peck A, Fahad HM, Ota H, Shiraki H, Kiriya
D, Lien D-H, Brooks GA, Davis RW, and Javey A, Nature 529, 509 (2016). [PubMed: 26819044]
Gnyawali V, Strohm EM, Wang J-Z, Tsai SSH, and Kolios MC, Scientific Reports 9, 1585 (2019).
[PubMed: 30733497]
Han S-M, Cho J-H, Cho I-H, Paek E-H, Oh H-B, Kim B-S, Ryu C, Lee K, Kim Y-K, and Paek S-H,
Author Manuscript

Analytica Chimica Acta 587, 1 (2007). [PubMed: 17386746]

Hassan U, Ghonge T, Jr BR, Patel M, Rappleye M, Taneja I, Tanna A, Healey R, Manusry N, Price
Z, Jensen T, Berger J, Hasnain A, Flaugher E, Liu S, Davis B, Kumar J, White K, and Bashir R,
Nature Communications 8, 15949 (2017).
Hassan U, Reddy B, Damhorst G, Sonoiki O, Ghonge T, Yang C, and Bashir R, Technology 03, 201
(2015).
Holmes D, Morgan H, and Green NG, Biosensors and Bioelectronics 21, 1621 (2006). [PubMed:
16332434]

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 11

Hu R, Liu T, Zhang X-B, Huan S-Y, Wu C, Fu T, and Tan W, Anal. Chem 86, 5009 (2014). [PubMed:
24731194]
Author Manuscript

Izadi Z, Sheikh-Zeinoddin M, Ensafi AA, and Soleimanian-Zad S, Biosensors and Bioelectronics 80,
582 (2016). [PubMed: 26896793]
Khateib F, Mehaney A, Amin RM, and Aly AH, Phys. Scr 95, 075704 (2020).
Koh A, Kang D, Xue Y, Lee S, Pielak RM, Kim J, Hwang T, Min S, Banks A, Bastien P, Manco
MC, Wang L, Ammann KR, Jang K-I, Won P, Han S, Ghaffari R, Paik U, Slepian MJ, Balooch G,
Huang Y, and Rogers JA, Science Translational Medicine 8, 366ra165 (2016).
Liu X, Li L, and Mason AJ, Philosophical Transactions of the Royal Society A: Mathematical,
Physical and Engineering Sciences 372, 20130107 (2014).
Lu Y, Macias D, Dean ZS, Kreger NR, and Wong PK*, IEEE Transactions on NanoBioscience 14, 811
(2015a). [PubMed: 26584498]
Lu Z, Wang H, Naqvi SR, Fu H, Zhao Y, Song H, and Christen JB, in 2015 28th IEEE International
System-on-Chip Conference (SOCC) (2015b), pp. 240–244.
Maetzig T, Ruschmann J, Lai CK, Ngom M, Imren S, Rosten P, Norddahl GL, von Krosigk N, Sanchez
Milde L, May C, Selich A, Rothe M, Dhillon I, Schambach A, and Humphries RK, Molecular
Author Manuscript

Therapy 25, 606 (2017). [PubMed: 28253481]

Manikandan MS and Soman KP, Biomedical Signal Processing and Control 7, 118 (2012).
Marinov MB, Nikolov GT, and Ganev B, in 2018 IX National Conference with International
Participation (ELECTRONICA) (2018), pp. 1–4.
Pierce KM, Parsons BA, and Synovec RE, in Data Handling in Science and Technology, edited by de
la Peña AM, Goicoechea HC, Escandar GM, and Olivieri AC (Elsevier, 2015), pp. 427–463.
Ram MR, Madhav KV, Krishna EH, Komalla NR, and Reddy KA, IEEE Transactions on
Instrumentation and Measurement 61, 1445 (2012).
Redhyka GG, Setiawan D, and Soetraprawata D, in 2015 International Conference on Automation,
Cognitive Science, Optics, Micro Electro-Mechanical System, and Information Technology
(ICACOMIT) (2015), pp. 72–77.
Ridgway K, Lalljie SPD, and Smith RM, Journal of Chromatography A 1153, 36 (2007). [PubMed:
17313955]
Sarimollaoglu M, Nedosekin DA, Menyaev YA, Juratli MA, and Zharov VP, Photoacoustics 2, 1
Author Manuscript

(2014). [PubMed: 24921062]

Singh Y, Tripathi S, and Pandey M, International Journal of Computer Applications 10, 23 (2010).
Stupin DD, Koniakhin SV, Verlov NA, and Dubina MV, Phys. Rev. Applied 7, 054024 (2017).
Sun LL, Leo YS, Zhou X, Ng W, Wong TI, and Deng J, Materials Science for Energy Technologies 3,
274 (2020).
Sun T, Holmes D, Gawad S, Green NG, and Morgan H, Lab on a Chip 7, 1034 (2007). [PubMed:
17653346]
Taylor FJ and Williams A, Electronic Filter Design Handbook, Fourth Edition (McGraw-Hill
Education, 2006).
Verma N and Singh M, Biosensors and Bioelectronics 18, 1219 (2003). [PubMed: 12835039]
Volpetti F, Garcia-Cordero J, and Maerkl SJ, PLOS ONE 10, e0117744 (2015). [PubMed: 25680117]
Wagner K, Sami MA, Norton C, McCoy J, and Hassan U, RSC Advances 11, 21315 (2021). [PubMed:
35478803]
Weinberg L and Slepian P, IRE Transactions on Circuit Theory 7, 88 (1960).
Author Manuscript

Yoo H, Shin J, Sim J, Cho H, and Hong S, Biosensors and Bioelectronics 168, 112561 (2020).
[PubMed: 32877782]
Zhang S, Li Z, and Wei Q, Nanotechnology and Precision Engineering 3, 32 (2020).
Zhong J, Liang M, and Ai Y, Lab on a Chip 21, 2869 (2021). [PubMed: 34236057]

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 12
Author Manuscript
Author Manuscript
Author Manuscript

Fig. 1:
(a) Custom printed circuit board (PCB) which onboards microfabricated microfluidic
Author Manuscript

impedance cytometry device. Silver conductive epoxy adheres and facilitates electron
transfer between gold microelectrodes and computer-connector interface. (b) Brightfield
microscope image of microfluidic electric field regime. A polydimethylsiloxane (PDMS)-
based channel has focusing regions between gold electrodes which increases particle pulse
amplitude. The middle electrode is voltage stimulated (Vin) and exterior detecting electrodes
produces an electric field in the region. (c) Flowchart of digital filtering selection criteria
and how filtering conditions were ranked for signal quality optimization. (d) Time-domain

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 13

impedance detection data of 9-micron polystyrene particles collected from the microfluidic
Author Manuscript

device prior to applying digital filtering. A differential signal is collected from the two
detecting electrodes, followed by transimpedance current amplification which produces a
bipolar pulse for each particle. (e) The same impedance data after digital filtering with
the most optimized conditions (1st order Butterworth low pass filter at a 60 kHz cutoff
frequency and 3rd order Chebyshev Type I high pass filter at a 30 Hz cutoff frequency)
which reveals lower noise band and baseline drift per polystyrene pulse signal amplitude.
Author Manuscript
Author Manuscript
Author Manuscript

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 14
Author Manuscript
Author Manuscript
Author Manuscript

Fig. 2.
(a) Full time-domain recording results of unfiltered 9 micron polystyrene impedance pulses.
Author Manuscript

(b) Full Fourier transformation of recordings with highlighted high-noise frequency regimes:
low noise amplitude (c), powerline interference (d), and high frequency noise above
approximately 90 kHz (e).

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 15
Author Manuscript
Author Manuscript
Author Manuscript

Fig. 3:
High pass filtering results alone when varying the cutoff frequency (10 Hz to 50 Hz) with
different filter types (Butterworth (a-c), Chebyshev Type I (d-f), Chebyshev Type II (g-i)
Author Manuscript

filter types) and different filter orders: 1st order (gray), 2nd order (red), 3rd order (blue),
4th order (green). Results compared with the number of particles counted through the
impedance detection recording (a, d, g), the average background noise (b, e, h), and the SNR
or signal to noise ratio (c, f, i) across the different filter types, orders, and cutoff frequencies.
Black dotted lines represent unfiltered values for comparison.

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 16
Author Manuscript
Author Manuscript
Author Manuscript

Fig. 4:
Low pass filtering results when varying the cutoff frequency (60 kHz to 120 kHz) with
different filter types (Butterworth (a-c), Chebyshev Type I (d-f), Chebyshev Type II (g-i)
Author Manuscript

filter types) and different filter orders: 1st order (gray), 2nd order (red), 3rd order (blue),
4th order (green). Results compared with the number of particles counted through the
impedance detection recording (a, d, g), the average background noise (b, e, h), and the SNR
or signal to noise ratio (c, f, i) across the different filter types, orders, and cutoff frequencies.
Statistics presented also simultaneously include filtering with the most optimized high pass
filtering results (Cheby1, 3rd order, 30 Hz). Black dotted lines represent unfiltered values for
comparison.

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 17
Author Manuscript
Author Manuscript
Author Manuscript

Fig. 5:
Time domain-data with isolated neutrophil impedance pulses before (a) and after (b)
optimized digital signal filtering (3rd order Chebyshev Type I high pass filter at a 30
Hz cutoff frequency and 3rd order Chebyshev Type I low pass filter at a 60 kHz cutoff
frequency)
Author Manuscript

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 18

Table 1:

Highest ranking high pass alone filtering conditions for each filter type
Author Manuscript

Filter type, order, and cutoff frequency # of particles counted Noise (V) SNR (dB) Overall rank

Cheby1, 3rd order, 30 Hz 487 0.1317 25.1761 1

Butterworth, 3rd order, 15 Hz 520 0.1679 25.0369 11

Cheby2, 4th order, 10 Hz 515 0.1599 24.8857 39

Author Manuscript
Author Manuscript
Author Manuscript

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 19

Table 2:

Highest ranking low pass alone filtering conditions for each filter type including the highest ranking high pass
Author Manuscript

filtering conditions (Cheby1, 3rd order, 30 Hz)

Filter type, order, and cutoff frequency # of particles counted Noise (V) SNR (dB) Overall rank

Butterworth, 1st order, 60 kHz 544 0.1211 26.6293 1

Cheby1, 3rd order, 60 kHz 538 0.1255 26.2322 4

Cheby2, 2nd order, 60 kHz 530 0.1421 25.4292 7

Author Manuscript
Author Manuscript
Author Manuscript

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.

 Ashley and Hassan Page 20

Table 3:

Highest ranking low and high pass filtering conditions for isolated human neutrophil impedance data.
Author Manuscript

Filter type, order, and cutoff frequency # of particles counted Noise (V) SNR (dB) Overall rank

High pass: Cheby1, 3rd

order, 30 Hz 716 0.1425 26.6806 1
Low pass: Cheby1, 3rd order, 60 kHz

High pass: Cheby1, 3rd order, 30 Hz 726 0.1320 26.6108 2

Low pass: Cheby1, 4th order, 75 kHz

High pass: Cheby1, 3rd order, 30 Hz 719 0.1425 26.5687 3

Low pass: Butterworth, 1st order, 60 kHz

High pass: Cheby1, 1st order, 25 Hz 724 0.1455 26.4198 4

Low pass: Butterworth, 1st order, 60 kHz

High pass: Butterworth, 4th order, 20 Hz 726 0.1515 26.4095 5

Low pass: Butterworth, 4th order, 120 kHz

Unfiltered 769 0.1770 19.1833 6

Author Manuscript
Author Manuscript
Author Manuscript

Biomed Microdevices. Author manuscript; available in PMC 2023 October 28.