Comparative Study of Data Processing Techniques for Pancreatic Islets

in Organ-on-Chip Applications

Roland Giraud

1 a

, Dorian Chapeau

1 b

, Jochen Lang

2 c

, Matthieu Raoux

2 d

Sylvie Renaud

2 e

and Antoine Pirog

3 f

Univ. Bordeaux, CNRS, Bordeaux INP, IMS, UMR 5218, F-33400 Talence, France

Univ. Bordeaux, CNRS, Bordeaux INP, CBMN, UMR 5248, F-33600 Pessac, France

Junia, Electronics-Physics-Acoustics Department, F-59000 Lille, France

Keywords:

Organ-on-Chips, MEA, Beta Cells, GLR, DWT, FFT.

Abstract:

Organ-on-chip technology presents a promising platform to study complex physiological processes in a con-

trolled environment. However organ-on-chip devices bring considerable constraints to online monitoring in-

strumentation. This study investigates methods for leveraging data from organ-on-chip systems designed for

diabetic studies by processing recorded extracellular signals from pancreatic islets. The signal processing

techniques used are designed to address the inherent constraints of microﬂuidics, particularly to provide on-

line (real-time) readings and operate effectively in low Signal-to-Noise Ratio (SNR) conditions. This study

assesses the performance of different algorithms using several detection approaches. Synthetic and experi-

mental data were utilized to evaluate algorithm robustness to best account for biological variability. Among

the algorithms tested, those based on frequency and time-frequency methods performed best when compared

to conventional ﬁltering and thresholding approaches, especially regarding robustness to noise and biological

variability.

1 INTRODUCTION

Organ-on-Chips (OoC) are miniature systems repli-

cating human organ structure and function on a mi-

croscale. Constructed using microfabrication tech-

niques, these chips house animal or human cells that

mimic organ function in micro- to milli-meter scale

culture chambers. Microﬂuidic channels control the

ﬂow of nutrients, oxygen, and other substances, cre-

ating a dynamic microenvironment. Equipped with

sensors for real-time monitoring, OoC technology is

invaluable for studying physiological processes, mod-

eling diseases, and testing drug responses in a more

physiologically accurate in vitro setting (Clapp et al.,

2021). This innovation holds the potential to advance

drug development and personalized medicine by of-

fering relevant and reliable experimentation models in

https://orcid.org/0009-0009-5639-0697

https://orcid.org/0000-0002-5998-2905

https://orcid.org/0000-0001-6847-5641

https://orcid.org/0000-0001-8791-2034

https://orcid.org/0000-0001-5632-7991

https://orcid.org/0000-0002-7913-003X

a wide range of application cases (Mastrangeli et al.,

2019). However, Organ-on-Chip technology faces

notable challenges (Wikswo et al., 2013). First, in-

tegrating sensors onto the chip poses signiﬁcant chal-

lenges due to its small size and the biological nature

of the cultured material (Fuchs et al., 2021). Second,

microﬂuidic equipment and ﬂow introduce substan-

tial noise during data acquisition. Additionally, the

system requires fast data processing algorithms to ad-

dress the demand for online monitoring (Moya et al.,

2018) and feedback (either automated or performed

by the experimenter).

This study was conducted in the context of OoC

for diabetes and establishes a comparison of algo-

rithmic solutions to process electrophysiological data

from pancreatic islets recorded on MicroElectrode

Arrays (MEA). In response to glucose, pancreatic

beta cells activate and produce an oscillatory electri-

cal signal that synchronizes with neighboring cells via

a gap-junction protein (connexin36) (Lebreton et al.,

2015). The resulting electrical signature resembles

Action Potentials in shape, but with a broader and

slower proﬁle known as Slow Potentials (SP). SP os-

cillatory frequency (ranging between 0.2-2 Hz) corre-

Giraud, R., Chapeau, D., Lang, J., Raoux, M., Renaud, S. and Pirog, A.

Comparative Study of Data Processing Techniques for Pancreatic Islets in Organ-on-Chip Applications.

DOI: 10.5220/0013255700003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 957-963

ISBN: 978-989-758-731-3; ISSN: 2184-4305

957

lates with the insulin production of Langherans islets

and therefore gives precious insight of islet physiol-

ogy (Jaffredo et al., 2021). This research aims at

ﬁnding a precise, robust, and real-time method of

measurement for the Slow Potential (SP) frequency,

within the context of monitoring OoC devices for di-

abetes research.

2 MATERIAL AND METHODS

2.1 Electrophysiology Setup and Data

Acquisition

The signal generated by beta cells was recorded us-

ing a Micro Electrode Array bonded to a PDMS

microﬂuidic chip. Electrophysiological data were

recorded using two widely used acquisition frontends

: MCS 1060 preampliﬁer and USB-ME64 acquisition

board as described in (Lebreton et al., 2015), or Intan

RHD2132 pre-ampliﬁers (Abarkan et al., 2022) con-

nected to a FPGA board for recording. All signal ac-

quisition was performed at a sampling rate of 10 kHz.

This study presents results from diverse (static and

microﬂuidic) experiments and protocols.

2.2 Pancreatic Islet Signal Preliminary

Analysis

Preliminary studies were carried out to determine the

relevant parameters for SP identiﬁcation. Fig. 1 illus-

trates key identiﬁcation criteria for SP. In Fig. 1.A, the

islet signal exhibits a pulsatile behavior, alternating

between silent and active phases. SP are character-

ized by a speciﬁc high depolarization phase (Fig. 1.B,

covered electrodes with SP) and occur in a frequency

band of 0.2 Hz to 2.0 Hz (Fig. 1.C). With islet

activity being modulated by glucose concentration,

these measurements were conducted using a culture

medium containing 11 mMol/l of glucose in which

islets exhibit sustained activity. A Principal Com-

ponent Analysis (PCA) and dbscan clustering was

performed on metrics extracted from detected events

: amplitude, frequency, amplitude of neighbouring

events, frequency of neighbouring events, and wave-

form (waveform data points resampled to the dimen-

sions of the shortest waveform detected) (Fig. 1.D).

This is intended to highlight, based on event prop-

erties, clusters separating false detections of noisy

events from physiological SP events. This cluster-

ing proved highly variable and did not - by itself -

provide a robust method for separating biological sig-

natures (SPs) from noise, which indicated that noise

and SP share similar characteristics ; therefore, more

sophisticated detection algorithms are needed to dif-

ferentiate them.

2.3 Algorithmic Solutions

Taking into account signal characteristics, several al-

gorithmic solutions were tested to evaluate the fre-

quency of SPs. All signal processing and algorithmic

tests were performed in Python 3.11.1 and libraries

Scipy 1.10.0, Numpy 1.24.1, and PyWavelets 1.4.1.

The algorithmic solutions (described individually

below) were tested and compared via parametric anal-

yses mimicking a range of acquisition parameters

(e.g. signal-to-noise ratio) and detection parameters

(e.g. detection threshold). For a fair comparison of

performance between algorithmic solutions, their de-

tection parameters were normalized relative to their

value of best performance (normalized parameter ex-

pressed as ∆/∆

where ∆ is the absolute value of

the parameter and ∆

its value at best performance).

Recorded signals were downsampled to 100 Hz to en-

hance computational efﬁciency without loss of rele-

vant biological information considering the frequency

range of interest for SP (0.2 Hz to 2.0Hz).

2.3.1 Ofﬂine Peak Detection

Slow Potentials were detected as local minima in the

signal, found using Scipy’s find peaks (further re-

ferred to as FP) algorithm from Scipy 1.10.0 This

function ﬁnds all local maxima by comparison of

Figure 1: Characteristics of recorded islets electrophysio-

logical data: (A) Filtered data from one MEA channel, (B)

Amplitude histogram of three different channels (C) Spec-

trogram of A (D) Principal Component Analysis and DB-

SCAN clustering. True detections of SPs are shown in dark

red, false detections caused by noise are shown in black and

light red shows detections that could not be identiﬁed as ei-

their noise or true detection.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

958

neighboring values. As a baseline, prominence pa-

rameters were manually tuned to best match signal

characteristics and maximize detection rate. Before

peak detection, signals were ﬁltered using Butter-

worth bandpass ﬁlter (0.2 Hz, 1

order to 2 Hz, 2

order).

2.3.2 Online Peak Detection

To account for the real-time constraint of our study

and provide a comparison reference with previous

works, the next algorithms are either real-time or real-

time-compatible. This algorithm (further referred to

as RT), previously described in (Pirog et al., 2018), is

also based on extremum detection with a prominence

parameter. To include a validation period around

every local extremum detected and avoid false de-

tections caused by signal jittering, SP detected with

this algorithms are processed some milliseconds after

their actual appearance. Butterworth bandpass ﬁlter-

ing was applied before detection (0.2 Hz, 1

order to

2 Hz, 2

order).

2.3.3 Discrete Wavelet Transform Filtering

In this detection algorithm, signals were ﬁltered using

wavelet transformation before performing SP detec-

tion. The discrete wavelet transform (further referred

to as DWT) (Unser and Aldroubi, 1996) is a math-

ematical transformation used to break down a signal

into multiple sets of discretely sampled wavelets, that

describe how the signal evolves over time within a

speciﬁc frequency range. Therefore SP can be de-

tected by selecting the coefﬁcient in the frequency

band of SP and applying prominence-based event de-

tection. The DWT was applied to the input signal us-

ing Daubechies wavelets (db4) and 10 transform lev-

els. Then for each set within the frequency band of

interest (levels 0 to 6), a threshold was applied and

only the coefﬁcients above the threshold were kept.

Finally, the ﬁltered signal was constructed by apply-

ing an inverse wavelet transform, and peak detection

was performed. The inputs parameters were the num-

ber of sets and the selected ones for ﬁltering. Equally,

the prominence parameter and a threshold value were

required.

2.3.4 Generalized Likelihood Ratio Statistical

Test

This algorithm (further referred to as GLR) aims

at discerning between baseline noise and electro-

physiological activity using a model-based approach.

The Generalized Likelihood Ratio Test (GLR) as de-

scribed in (Mansouri et al., 2018) was used to com-

pare the goodness of ﬁt of observed data against a sta-

tistical model of noise, thus helping detect changes in

a signal and make decisions. The GLR method calcu-

lated a ratio of likelihoods between a target hypothe-

sis (e.g., presence of a SP) and a reference hypothesis

(e.g., baseline condition). This ratio served as a sta-

tistical test, indicating how well the data aligned with

one hypothesis over another. Applying a threshold on

this ratio signal yielded a SP detection index. This

technique required a calibration phase (performed in

the absence of electrophysiological signals, ie. in-

hibitory conditions for the islets), a detection thresh-

old parameter, and an observation window to perform

the GLR. Before processing, signals were ﬁltered us-

ing a Butterworth bandpass ﬁlter (0.2 Hz, 1

order to

2 Hz, 2

order).

2.3.5 Online Frequency Analysis

This last approach (further referred to as FFT) does

not focus on event detection but rather aims at directly

assessing signal frequency. To that end, a Fast Fourier

Transform (FFT) algorithm was used with a 30 s slid-

ing window and a 1 s step. An adaptive threshold

was set according to a fraction of signal energy over

the observation window, and peaks in FFT coefﬁ-

cients were detected when they exceeded that thresh-

old. The frequency returned by the algorithm was the

weighted average of the detected peaks in each obser-

vation window. Before computing FFT signals were

ﬁltered using a Butterworth bandpass ﬁlter (0.2 Hz,

order to 2 Hz, 2

order).

2.4 Benchmarking

2.4.1 Performance Evaluation Metrics

Two distinct metrics were employed to evaluate the

algorithms performances.

• Root Mean Square error (RMS): RMS error testi-

ﬁes for the difference between a model and an ob-

servation. RMS error was calculated as the square

root of the mean of the squares of the differences

between predicted values and observed values.

• Maximum of cross correlation: Cross correla-

tion is a measure of similarity of two series, as

a function of the displacement of one relative to

the other. Cross correlation was computed us-

ing Numpy 1.24.1 then the maximum value was

taken.

2.4.2 Test Scenarios

Detection algorithms were benchmarked across sev-

eral test scenarios, to account for the inherent sig-

Comparative Study of Data Processing Techniques for Pancreatic Islets in Organ-on-Chip Applications

959

nal processing difﬁculties encountered in electropysi-

ological signals recorded in microphysiological sys-

tems. First, synthetic signals mimicking the prop-

erties of electrophysiological data were used as test

inputs : as the frequency of the test signal was well

deﬁned for each instant, it could be compared to mea-

sured outputs. The test signal consisted in a sine mod-

ulated in frequency and amplitude within the physio-

logical range of the signal (ie. modulated frequency

from 0.2 to 1.2 Hz and amplitudes between 5 to

30 µV ) and reproducing the alternating silent/active

phases. Gaussian white noise was added (SNR ∼

5 dB) to mimic measurement noise.

Parametric Study of Detection Parameters. The

sensitivity and variability of measured outputs in re-

sponse to changes in measurement parameters was

evaluated in a parametric study of their detection pa-

rameters. First, a preliminary parameter space study

was conducted to evaluate the settings leading to best

detection performance. Then, the most inﬂuential pa-

rameter (event-based methods FP, RT, DWT, GLR :

the detection threshold ; frequency-base method FFT

: the adaptive threshold) was systematically varied in

a deﬁned range (0.1 to 10 times) around its point of

best performance. This study aimed at highlighting

the immunity, or lack thereof, of the benchmarked al-

gorithms to their detection parameters in an effort to

minimize biases caused by algorithm conﬁguration.

Parametric Analyses on Signal-to-Noise Ratio.

To study the behaviour of the tested algorithms as the

signal-to-noise ratio (SNR) deteriorates, a parametric

study was conducted with added white noise, for a

SNR ranging from 45dB to -10dB. The white noise

applied during this study replaces the Gaussian white

noise originally set in the test scenario.

Validation Study. Finally, all algorithms were

tested on electrophysiological recordings of pancre-

atic islets, to account for experimental and biological

variability. To provide a reference for performance

evaluation, recorded signals were manually annotated

with SP events and converted to a reference signal de-

scribing the expected frequency measurements. Fur-

thermore, to emulate increasingly poor recording con-

ditions and monitor detection performance loss, a

parametric study was again done with added white

noise of increasing intensity (SNR from 40 to 0 dB).

3 RESULTS

3.1 Parametric Study on Synthetic Data

The robustness of each algorithm against noise and

biological variability was evaluated by adjusting their

most sensitive parameter: the detection threshold.

Fig. 2.A highlights the detection performance when

tuning this parameter. The prominence-based algo-

rithms RT and FP yielded low error when the thresh-

old parameter was properly tuned, but were highly

sensitive to its variations. Indeed, false detections

abruptly occured when small variations were applied

to the detection threshold. FFT exhibited lower per-

formance but demonstrated very strong resilience to

variations in the threshold parameter, as performance

indicators remained almost constant regardless of the

threshold value. On the other hand, DWT and GLR

consistently delivered low error rates when compared

to the other algorithms, and maintained excellent per-

formance across a broad range of threshold values.

Fig 2.B reports the inﬂuence of noise on detection

performance. Results are consistant with Fig 2.A as

the FP and RT algorithms showed high sensitivity to

noise contrarily to FFT that showed very little pertur-

bation. GLR and DWT still performed better than FP

and RT at low SNR but were less robust than FFT.

Figure 2: Evaluation of the algorithms (FO, RT, DWT,

GLR, FFT) on synthetic signals. (A) Parametric analysis on

the detection threshold (detection threshold was normalized

according to ∆/∆

where ∆

is the threshold at best perfor-

mance and ∆ the threshold). (B) Parametric analysis on the

test signal SNR.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

960

3.2 Parametric Study on Recorded Data

Next, the same technique was applied to recorded

electrophysiological data. Recorded signals were

manually annotated with SP events to serve as a ref-

erence for performance evaluation (Fig. 3).

As depicted in Fig. 3, the Fourier transform algo-

rithm (FFT) appeared to be the most robust to noise.

Both peak-ﬁnding algorithms (FP and RT) perfor-

mances degraded quickly under higher noise condi-

tions, despite maintaining a good overall correlation

with the reference signal. The GLR algorithm per-

formed the least favorably when compared to FFT and

DWT. DWT showed good error results but exhibited

a slightly lesser correlation ﬁt compared to FFT.

Figure 3: Evaluation of the algorithms (FO, RT, DWT,

GLR, FFT) on recorded data with added white noise to

change SNR. (A) RMS error. (B) Maximum of cross-

correlation.

3.3 Validation Study

Finally the algorithms were tested on three recordings

from different experiments and electrodes to account

for the experimental and biological variability.

The overall behaviour of each algorithm matches

the frequency variation of the signal but their perfor-

mance level varies (Fig. 4). The detection follows

the variations in glucose concentration ie the SP fre-

quency increases from 4 to 12 mM and decreases

from 12 to 4 mM. The lesser performance of GLR

can be explained by a poor match between the noise

Figure 4: Comparison of the different algorithmic outputs

to a manually constructed reference signal (both averaged

over a 30 s window). (A) Single channel at constant glu-

cose concentration (11 mM). (B) Average from two chan-

nels with variable glucose level (4 to 12 mM).

hypotheses made and the actual distribution recorded

during the silent phases used for calibration.

4 DISCUSSION

This study offers insights into signal processing ap-

proaches for pancreatic islets in organs-on-chips. Re-

sults in Fig.5 show clear superiority of the FFT algo-

rithm exhibiting lower RMS error and higher correla-

tion with the reference. Performance with GLR was

comparable with other algorithms, provided the noise

model adequately matched the noise distribution of

the experiments, and otherwise declining sharply. FP

and RT showed good performances if correctly tuned

but were highly dependent of SNR shown in Fig. 2

and 3. In comparison DWT ﬁltering showed simi-

lar performances but higher immunity to noise. Nev-

ertheless, the evaluation process on recorded data is

limited by the quality of reference signals. Consid-

ering the complexity and variability of organ-on-chip

devices, generating a reliable gold reference is a sig-

niﬁcant challenge. Fig.5 was obtained using three ref-

erence signals, where SP were hand marked, a com-

mon procedure in electrophysiology. Therefore the

Comparative Study of Data Processing Techniques for Pancreatic Islets in Organ-on-Chip Applications

961

so-called reference can only be used as an indicator

subject to human error and not as an absolute refer-

ence.

Figure 5: Performance evaluation on experimental data

(mean of 3 recordings).

Through this study, we also compared two mea-

surement paradigms, based either on event detec-

tion (FP, RT, DWT and GLR) or frequency detec-

tion (FFT). Each of the tested algorithms employs

a distinct measurement approach : RT and FP are

simple prominence-based detection methods imple-

mented online and ofﬂine, respectively; GLRT uti-

lizes a statistical test based on noise characteristics;

DWT is a time-frequency method used as a ﬁlter-

ing tool; FFT is a purely frequency-based approach.

These approaches have speciﬁc advantages and draw-

backs. RT and FP offer excellent temporal resolu-

tion, dynamics, and computational efﬁciency but ex-

hibit poor robustness to noise and parameter changes.

GLRT is more reliable than FP and RT provided

that the excluded noise has Gaussian characteristics.

DWT ﬁltering presents a good trade-off between tem-

poral resolution and performance; however, it still re-

quires ﬁne tuning relative to the signal. Finally, FFT

consistently shows good performance : the sliding

window implemented within the FFT approach re-

duces sensitivity to isolated noise events, as the anal-

ysis is performed over a window of samples rather

than a single point. However this robustness comes

at the expense of poor dynamics and temporal reso-

lution, and therefore does not perform optimally dur-

ing transient states. Another beneﬁt of this algorithm

is the weighted average of frequencies performed af-

ter the fft algorithm, which may be more relevant

than threshold-based approaches. Indeed, SPs likely

comprise multiple electrophysiological couplings at

slightly different frequencies. To that extent, while

event-based methods showed excellent performance,

the frequency-based method may reveal supplemen-

tary information; indeed, islet behaviour result from

multiple beta cell signals forming clusters of activa-

tion (Luchetti et al., 2023)(Jaffredo et al., 2018), with

periodic behaviours well suited to frequency analy-

sis. A frequency-based approach giving insight on the

frequency spread of a signal rather than a single fre-

quency measurement may thus be especially relevant

to fully characterize islet behavior.

ACKNOWLEDGEMENTS

We thank David Henry for his assistance with the

GLRT algorithm and Julien Gaitan for his help in bi-

ological experiments. This work was funded by the

ANR FUN-NET (ANR-21-CE14-0078) and the ANR

DIAMOCHIP (ANR-22-CE19-0032-05).

REFERENCES

Abarkan, M., Pirog, A., Maﬁlaza, D., Pathak, G., et al.

(2022). Vertical organic electrochemical transistors

and electronics for low amplitude micro-organ sig-

nals. Advanced Science, 9(8):2105211.

Clapp, N., Amour, A., Rowan, W. C., and Candarlioglu,

P. L. (2021). Organ-on-chip applications in drug dis-

covery: an end user perspective. Biochemical Society

Transactions, 49(4):1881–1890.

Fuchs, S., Johansson, S., Tjell, A.Ø., Werr, G., Mayr, T.,

and Tenje, M. (2021). In-line analysis of organ-on-

chip systems with sensors: Integration, fabrication,

challenges, and potential. ACS Biomaterials Science

& Engineering, 7(7):2926–2948. PMID: 34133114.

Jaffredo, M., Bertin, E., Pirog, A., Puginier, E., Gaitan, J.,

Oucherif, S., Lebreton, F., et al. (2021). Dynamic uni-

and multicellular patterns encode biphasic activity in

pancreatic islets. Diabetes, 70(4):878–888.

Jaffredo, M., Pirog, A., Bertin, E., Catargi, B., Renaud,

S., Lang, J., and Raoux, M. (2018). Differential beta

cell coupling patterns drive biphasic activity. In 54th

EASD Annual Meeting of the European Association

for the Study of Diabetes, volume 61, pages S17–S18.

Lebreton, F., Pirog, A., Belouah, I., Bosco, D., Berney, T.,

Meda, P., Bornat, Y., Catargi, B., Renaud, S., Raoux,

M., et al. (2015). Slow potentials encode intercellular

coupling and insulin demand in pancreatic beta cells.

Diabetologia, 58:1291–1299.

Luchetti, N., Filippi, S., and Loppini, A. (2023). Multilevel

synchronization of human β-cells networks. Frontiers

in Network Physiology, 3.

Mansouri, M., Baklouti, R., Harkat, M. F., et al. (2018).

Kernel generalized likelihood ratio test for fault de-

tection of biological systems. IEEE Transactions on

NanoBioscience, 17(4):498–506.

Mastrangeli, M., Millet, S., the ORCHID partners, and van

den Eijnden-van Raaij, J. (2019). Organ-on-chip in

development:towards a roadmap for organs-on-chip.

Preprints.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

962

Moya, A., Ortega-Ribera, M., Guimer

a, X., Sowade, E.,

Zea, M., Illa, X., et al. (2018). Online oxygen mon-

itoring using integrated inkjet-printed sensors in a

liver-on-a-chip system. Lab on a Chip, 18(14):2023–

2035.

Pirog, A., Bornat, Y., Perrier, R., Raoux, M., Jaffredo, M.,

Quotb, A., Lang, J., Lewis, N., and Renaud, S. (2018).

Multimed: An integrated, multi-application platform

for the real-time recording and sub-millisecond pro-

cessing of biosignals. Sensors, 18(7):2099.

Unser, M. and Aldroubi, A. (1996). A review of wavelets

in biomedical applications. Proceedings of the IEEE,

84(4):626–638.

Wikswo, J. P., Block III, F. E., Cliffel, D. E., Goodwin,

C. R., Marasco, C. C., Markov, D. A., McLean, D. L.,

and others. (2013). Engineering challenges for instru-

menting and controlling integrated organ-on-chip sys-

tems. IEEE Transactions on Biomedical Engineering,

60(3):682–690.

Comparative Study of Data Processing Techniques for Pancreatic Islets in Organ-on-Chip Applications

963