Received 8 January 2025, accepted 5 March 2025, date of publication 13 March 2025, date of current version 2 April 2025.

Digital Object Identifier 10.1109/ACCESS.2025.3550949

Convolutional Autoencoder With Sequential and

Channel Attention for Robust ECG Signal
Denoising With Edge Device Implementation
ALIF WICAKSANA RAMADHAN 1 , SYIFA KUSHIRAYATI 1 , SALSABILA AURELLIA 1 ,
MGS M. LUTHFI RAMADHAN 1 , MUHAMMAD HANNAN HUNAFA 1 ,
MUHAMMAD FEBRIAN RACHMADI1 , APRINALDI JASA MANTAU1 , (Member, IEEE),
SITI NURMAINI2 , (Member, IEEE), SATRIA MANDALA 3,4 , (Member, IEEE),
AND WISNU JATMIKO 1 , (Senior Member, IEEE)
1 Facultyof Computer Science, University of Indonesia, Depok 16424, Indonesia
2 Intelligent
System Research Group, Sriwijaya University, Palembang 30139, Indonesia
3 Human Centric (HUMIC) Engineering, Telkom University, Bandung 40257, Indonesia
4 School of Computing, Telkom University, Bandung 40257, Indonesia

Corresponding author: Alif Wicaksana Ramadhan ([email protected])

This work was supported by the Katalis Program from the Ministry of Education, Culture, Research, and Technology (Kemendikbud)
under Grant NKB-1010/UN2.RST/HKP.05.00/2024.
This work involved human subjects or animals in its research. The authors confirm that all human/animal subject research procedures and
protocols are exempt from review board approval.

ABSTRACT Electrocardiograms (ECG) are vital for diagnosing various cardiac conditions but are often
corrupted by noise from multiple sources, which can hinder accurate interpretation. Denoising ECG signals
is particularly challenging because noise usually overlaps with the frequency range of the signal of interest.
We proposed a convolutional autoencoder with sequential and channel attention (CAE-SCA) to address
this issue. Sequential attention (SA) is based on long short-term memory (LSTM), which captures causal-
temporal relationships. Meanwhile, channel attention (CA) is used to emphasize important features within
channels. SA is applied to the skip connection of each encoder block, and CA is applied after each decoder
block. We validated the CAE-SCA using the MIT-BIH and SHDB-AF databases as clean ECG signals,
with the MIT-BIH Noise Stress Test Database as the noise source. Experimental results give an average
SNR value of 16.187 dB, RMSE of 0.059, and PRD value of 18.529 in the MIT-BIH database. While in
the SHDB-AF dataset, the model obtained 15.308 dB of SNR, 0.049 of RMSE, and 19.220 of PRD. These
results demonstrate our CAE-SCA outperforms all the state-of-the-art methods across all tested metrics.
For efficiency, CAE-SCA achieved competitive results in the metrics of floating-point operations (FLOPs),
inference time, and total parameters. This allowed CAE-SCA to be implemented in edge devices as we tested
using our custom ECG acquisition circuit. A significance test further confirms a statistically significant
improvement in SNR values achieved by the CAE-SCA compared to baseline models, suggesting the
CAE-SCA’s potential for advancing ECG processing in healthcare applications.

INDEX TERMS Autoencoder, convolutional neural networks (CNN), ECG, signal denoising, LSTM.

I. INTRODUCTION
The rising incidence of heart disease in Indonesia has led
to substantial advances in health screening equipment. The
The associate editor coordinating the review of this manuscript and Indonesian Ministry of Health reported that the number
approving it for publication was R. K. Tripathy . of heart disease cases is increasing every year [1]. As a
2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License.
VOLUME 13, 2025 For more information, see https://creativecommons.org/licenses/by/4.0/ 54407
 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

result, there is still a great demand for heart diagnostic tools, commonly known that the RNN-based model is slower
which motivates vigorous research [2], [3], [4]. According due to its way of sequentially processing the input [22].
to the World Health Organization (WHO), cardiovascular The CNN-based model is designed to process data in a
illnesses were responsible for over 38% of Indonesia’s total parallel way. As stated previously, we created a lightweight
fatalities in 2019 [5], [6]. This highlights the urgent need deep-learning model while maintaining accuracy; thus,
for precise, effective, and affordable diagnostic methods we focused only on CNN-based architecture.
are needed. For people at high cardiovascular risk due to Chiang et al. proposed a fully convolutional denoising
conditions such as hypertension, diabetes, and hyperlipi- autoencoder (FCNDAE) for a simple CNN-based model for
demia, early identification is essential to enable prompt the signal denoising task [16]. Their proposed autoencoder
intervention [7]. was constructed using convolutional blocks as the encoder
By capturing the electrical activity of the heart, elec- and convolutional blocks as the decoder. An improvement
trocardiograms (ECGs) can be used for the diagnosis to this was proposed by Chen et al. by adding skip
of cardiac conditions. Important clinical information for connections between each encoder and decoder block [23].
identifying arrhythmias is provided by key ECG components, Later improvement has also been shown by [24] in their
including the P wave, QRS complex, T wave, and U work [24]. They added a convolutional block attention
wave [8], [9]. To improve arrhythmia detection, the Holter module (CBAM) to the autoencoder network. The results
monitor, which is the gold standard ambulatory electrocar- demonstrate improvement compared to the vanilla UNet-
diogram equipment, provides extended monitoring periods based autoencoder.
(24-72 hours) [7]. Nevertheless, recording outside of a hos- Furthermore, [19] proposed another approach by adding
pital increases the risk of noise from electrical interference, the transformer encoder as the bottleneck of the autoencoder,
muscle activity, or electrode movement, all of which can and custom other modules named adaptive parametric ReLU
lower the quality of the diagnosis [9]. Accurate recordings module (AP-ReLU) and dynamic feature aggregation module
depend on the use of efficient noise reduction techniques, (DAM) [19]. The proposed AP-ReLU was inspired by the
such as ECG denoising and acceptability evaluation [10]. ReLU activation function. The DAM network was also
Many deep learning-based denoising approaches have inspired by the CBAM proposed in several modules in the
been proposed using various models. A popular model encoder and decoder blocks.
for this task is the denoising autoencoder (DAE) [11]. To obtain insight into the previously available denoising
An autoencoder comprises two parts: an encoder and a architecture model, we conducted preliminary experiments
decoder. The encoder reduces the complexity of the data on it. The environments are different from each other;
while retaining important features. The decoder works thus, we rebuilt them in our environment configurations.
oppositely to the encoder. The compressed data between In the RNN-based model, the LSTM model achieved the
the encoder and decoder is referred to as the latent space. best performance values. In addition, in the CNN-based
By compressing and decompressing the input into a new models, we found that by adding a skip connection to the
representation, the model retains the important features in the CNN-based autoencoder, there was a significant performance
latent space required to reconstruct the new representation. improvement. Of course, it was not free. The number of
Many researchers have proposed denoising autoencoders, computations required for the model also increased, thereby
with various improvements, from simple, fully connected increasing the inference’s latency higher. In addition, the
neural networks to deep neural networks [12], [13] and not addition of attention blocks improves the performance of the
limited to one-dimensional signal data but also for image UNet-based model.
data [14], [15]. We aim to create a new model based on autoencoder
Another challenge in implementing a deep learning-based architectures by combining the advantages of some baseline
approach to Holter monitors is device constraints. The methods. Specifically, we would like to combine the advan-
analysis can be performed from a recorded ECG signal, which tages of the channel attention module and LSTM to capture
can be performed anywhere; there are no limitations in the important features both through channels and sequences.
device specifications. Meanwhile, a Holter monitor is a small The channel attention module demonstrates how humans
device that can accompany the patient under any condition for emphasize useful information while giving less attention
24 to 72 hours. It is impossible to use such a device with super to unimportant details, which improves model efficiency
specifications. To handle this situation, we must develop a and accuracy. In addition, LSTM networks can capture
deep learning model while maintaining a trade-off between long-term causal relationships in sequential data, such as
model inference speed and accuracy. ECG.
To explore various approaches to signal denoising tasks, This study aims to develop a more efficient denoising
we focus on two main model architecture types. These autoencoder model to optimize its functionality in embedded
two groups are architecture based on convolutional neural devices so that it can be implemented directly in Holter moni-
networks (CNN) and architecture based on recurrent neural tor systems used by medical practitioners. This technology is
networks (RNN) [2], [16], [17], [18], [19], [20], [21]. It is expected to enhance early detection and intervention for heart

54408 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

disease, especially in areas with limited access to medical of the encoder and decoder block. The final result shows
facilities. The main contributions of this study are as follows: that this model outperforms the other models, FCNDAE [16]
• We introduced a convolutional autoencoder architecture and ACDAE [25], on the sum of squared distances (SSD),
with sequential and channel attention (CAE-SCA) maximum absolute distance (MAD), and cosine similarity
for denoising ECG signals by combining the metrics. The models were tested on the denoising ECG
1D-UNet-based network, LSTM-based RNN, and task using the Medical Information Mart for Intensive Care
channel attention module. (MIMIC-III) Database [29] with noise added from the
• Our CAE-SCA outperformed all accuracy metrics MIT-BIH noise stress test database [27].
while maintaining competitive performance in terms of While ACDAE [25] and CDAE-BAM [24] tried to
computational efficiency metrics. give global attention using channel and spatial attention
mechanisms. On the other hand, study [19] used a transformer
The remainder of this paper is organized as follows. layer in the bridge of the encoder-decoder block. They
Section II provides related work. Section III explains the also used a custom activation network in place of a
details of the proposed CAE-SCA. Section IV gives a brief classical activation function called adaptive parametric ReLU
overview of the dataset used in this study. Section V explains (AP-ReLU) [30] for better involvement of negative value
the experimental design and experimental results. Section VI results for each output of the network layer. They also used
concludes this paper. the Dynamic Aggregation Module (DAM) on the last layer
of the decoder block, whose mechanism is similar to the
II. RELATED WORKS channel and spatial attention module used in [24]. research.
Several researchers have addressed the task of denoising ECG The proposed method is APtrans-CNN. This research gives
signals using an autoencoder (AE) [16], [19], [24], [25]. a result by combining an AE-based CNN with transformers
Research by Chiang et al. proposed a fully convolutional and DAM and using AP-ReLU as an activation network to
network autoencoder (FCNDAE) [16]. This model consists perform better denoising tasks, instead of only one or two
of 6 convolution layers on the encoder side; each encoder combinations of them, tested on the MIT-BIH arrhythmia [26]
block consists of a convolutional, batch normalization, and and atrial fibrillation database [31] with noise added from the
activation Exponential Linear Units (ELU) layer. On the MIT-BIH noise stress test database [27].
decoder side, 7 decoder blocks built with one deconvolution From the related works on denoising tasks using autoen-
layer were used. The results demonstrate that FCNDAE coders, we hypothesize that autoencoder performance can
outperforms traditional DNN and CNN approaches. This is be significantly improved by incorporating skip connections
demonstrated by a higher SNR value, a lower RMSE value, between encoder and decoder blocks and adding local or
and a PRD value on testing using the MIT-BIH arrhythmia global attention mechanisms. This research aims to combine
database [26] with noise added from the MIT-BIH noise stress these strategies while optimizing for efficient resource
test database [27]. utilization.
Several studies have added an attention mechanism
on the decoder side and a skip connection between the III. PROPOSED METHOD
output for each encoder-decoder block [19], [24], [25] like Fig. 1 explains the overall system design of our ECG device
U-Net Architecture [28]. The attention block demonstrates and the proposed deep learning-based ECG signal denoising
how humans emphasize useful information while giving filter. In addition to the deep learning-based filter, we also
less attention to unimportant details. The attention-based used an analog filter for the ECG signal acquisition circuit.
Convolutional Denoising Autoencoder (ACDAE) used the This is a common implementation in any signal-acquisition
CNN layer followed by the leaky rectified linear unit (ReLU) device, including ECG.
activation function and the maximum pooling layer for each The recorded raw signal, x(n) has two main com-
of the 4 encoders and the transposed CNN layer followed by ponents: the signal of interest or clean signal, s(n),
the Leaky ReLU activation function, the maximum pooling and the noise or unwanted signal η(n), represented
layer, and the channel attention module for each of the in (1). The signal-denoising task can be considered a system
4 decoder blocks. The results show that the skip connection with the same input and output domains. The input is a
mechanism can help recover the detailed features of the noised ECG signal x(n), and the output is a predicted clean
ECG signal that might be lost when compressed with an ECG signal ŝ(n). Both signals are ECG signals in the time
encoder block. In addition, the channel attention module domain.
hardly improved the model’s performance if placed on the
last layer of the decoder [25]. x(n) = s(n) + η(n), n ∈ Z≥0 (1)
Another model with a similar architecture as ACDAE [25]
is Convolutional Denoising Autoencoder with Block Atten- The output of the denoised signal is visualized in end-user
tion Module (CDAE-BAM) [24]. The significant difference devices, such as smartphones and desktops. The results are
is the additional spatial attention module in the attention sent to the Raspberry Pi 5 as the core processor device through
block, where the attention block is placed on every last layer a Wi-Fi connection.

VOLUME 13, 2025 54409

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

may experience aliasing. In addition, adding an analog filter

also helps optimize the digital filter’s performance.
An instrumentation amplifier is commonly used for
ECG signal conditioning [32], [33], [34]. ECG signals are
susceptible to interference from common-node noise sources,
such as power-line noise, motion artifacts, and other electrical
disturbances. The instrumentation amplifier was used for
its high gain with low noise, high input impedance, and
excellent common-mode rejection ratio (CMMR). The driven
right leg (DRL) was also used to optimize the use of the
instrumentation amplifier [35]. The DRL circuit actively
reduces common-mode noise at the source by driving the
common-mode voltage (measured across the body) to near
zero. This is achieved by injecting an inverted version of the
common-mode signal back into the patient through the right
leg electrode, creating a feedback loop that cancels the noise.
We also added a passive high-pass filter (HPF) to reduce
noise produced by the patient’s body which has a low
frequency. Other noise sources also come at high frequencies.
Aliasing can cause a high-frequency fold in the interest
signal and distort the digitized signal [36]. To avoid this, the
captured signal should have a range of frequencies below
half the sampling rate used by the ADC, according to the
Nyquist theorem [37]. This is achieved by adding a low-pass
filter (LPF) to the ECG acquisition circuit. In our circuit,
we use 3rd order Butterworth filter rather than a common LPF.
The Butterworth filter is designed to have a maximally flat
response in the pass band, meaning that there are no ripples
or variations in the amplitude of the signal within the pass
band [38]. This ensures that the desired ECG signal is not
distorted or attenuated.

B. DIGITAL FILTER DESIGN

A digital system can only process distinct signal data. For
a continuous signal, such as an ECG signal in our case,
it is common to divide the signal into a short time frame
or window to make it distinct data. Each resulting frame
reflects a segment of the signal sequence in a short period.
The signal is then processed frame-by-frame and treated as a
sequence. This approach is called windowing, which allows
us to analyze the local characteristics of the signal over time
and has also been implemented in many diverse research
studies [3], [39], [40].
In this paper, we have proposed a deep learning-based
FIGURE 1. System design.
filter due to its success in various problems. This deep
learning-based filter can process each segment of the ECG
signal. The output of each filter operation is then arranged
A. ELECTRONIC CIRCUIT DESIGN to form a new continuous sequence of filtered ECG signals.
Fig. 2 shows the visualization of these processes.
The upper side of Fig. 1 shows the ECG signal-acquisition
circuit design. This circuit is used to amplify a very low ECG
signal to the desired range so that it can be measured by an C. MODEL ARCHITECTURE
analog-to-digital converter (ADC) accurately. Even though This study proposes a combination of a CNN-based autoen-
we have already implemented a deep learning-based filter, coder with LSTM (Long Short-Term Memory) on the encoder
we still need an analog filter in this circuit. First, a deep gate at the skip connection and a channel attention block on
learning-based filter is a digital filter with limitations such as, the decoder gate at the skip connection. The basic mechanism

54410 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 2. Windowing filter process.

of the autoencoder is mapping the input x̂ into several latent

variables or hidden representations z through deterministic
function f (x̂), and mapping back to the reconstructed vector
ŝ through deterministic function g(z) [11]. In denoising,
the vector x̂ is a segmented, noised ECG signal, and the FIGURE 3. Activation block architecture.
reconstructed vector ŝ is a predicted, segmented, clean ECG
signal. The autoencoder of the proposed model can be seen
in Fig. 4. The function f (x̂) is equal to a sequence of several such as ECG signals. The LSTM architecture used in this
encoder blocks, and function g(z) is equal to a sequence of study employs a unidirectional LSTM-based RNN (Recurrent
several decoder blocks. The encoder block comprises a CNN Neural Network) by [43]. Third, we employ a channel
followed by a batch normalization, activation block, dropout, attention block, illustrated in Fig. 5. This module is applied
and maximum pooling layer. The decoder block comprises at the end of each decoder block layer to emphasize essential
transposed 1D CNN followed by a batch normalization block, information across the channel dimension. LSTM focuses
activation block, and dropout layer. The encoder and decoder on sequential causal-time information, the channel attention
blocks contain five blocks. The architecture of the activation block generalizes critical features across channel or feature
block is illustrated in Fig. 3. dimensions. The combined architecture of these strategies,
Bidirectional LSTM inspired the architecture to play a depicted in Fig. 4, effectively harnesses both sequential
spatial or temporal attention role in skip connection for and channel-specific information to optimize ECG signal
denoising PCG (phonocardiogram) signals [41]. The channel denoising.
attention module at the end of each decoder block before
entering the skip connection was adopted from the usage IV. DATASET
on [25]. Then, to better capture the negative value from A. MIT-BIH ARRHYTHMIA DATABASE AND SHDB-AF
each output of the network layer, we use an activation block DATABASE
adopted from [30] instead of a single activation function. The The MIT-BIH Arrhythmia Database is a well-known database
overall architecture can be seen in Fig. 4. for biomedical applications, specifically for the detection
An ECG signal is a quasi-periodic signal characterized and classification of arrhythmias in electrocardiogram (ECG)
by a periodic pattern with slight variations in both time signals [3], [18], [44], [45]. This database was built by
and amplitude [42]. Based on this characteristic we deliver the Massachusetts Institute of Technology (MIT) [26]. The
three strategies. First, we use a CNN-based autoencoder as dataset contains 48 records of two-channel ambulatory ECG
the foundational framework of the proposed model. The signals, each lasting an hour for each. The records were
proposed framework is designed to generalize patterns across obtained from 47 subjects recorded between 1975 and 1979,
segmented ECG sequences [16], [19], [24], [25]. Second, representing a range of cardiac events that are mandatory
we apply attention mechanisms to mitigate the loss of for arrhythmia clinical assessment. Each record in the
detailed information at the autoencoder’s bottleneck. The MIT-BIH Arrhythmia Database was acquired at a sampling
first attention mechanism incorporates LSTM networks to rate of 360 Hz and a resolution of 11 bits.
enhance the skip connection input from the encoder by Meanwhile, another ECG dataset was obtained from
capturing long-term causal relationships in sequential data, Japan. In contrast with the MIT-BIH Arrhythmia database,

VOLUME 13, 2025 54411

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 4. CAE-SCA architecture.

developed to evaluate the robustness of deep learning models

across various demographic characteristics, including sex,
age, and ethnicity. The SHDB-AF fills a significant void in
international ECG research databases because it is one of the
only datasets that includes Japanese patients. In this study,
both the MIT-BIH Arrhythmia and SHDB-AF databases were
used for benchmarking.

B. MIT-BIH NOISE STRESS TEST DATABASE

The MIT-BIH Arrhythmia and SHDB-AF databases are well-
known arrhythmia detection and classification databases.
Meanwhile, this research aims to develop a novel ECG signal-
denoising method. Hence, pairs of clean and noise signals
are required. In this research, instead of generating noise
signals synthetically, we used the MIT-BIH Noise Stress
Test Database [27] since it was recorded under real-noise
conditions using ambulatory electrocardiogram devices. This
noise-signal generation approach has also been proposed in
several other studies [3], [47], [48].
This database incorporates 12 records of ECG signals with
a duration of 30 minutes each and three records of typical
FIGURE 5. Channel attention block architecture.
noise in ambulatory ECG recordings with the same duration.
The noise signals were obtained using physically active
participants and standard ECG devices, including recorders,
SHDB-AF provides newer Holter ECG records, which electrodes, and leads. The electrodes were configured on
particularly target atrial fibrillation (AF) detection [46]. The the limbs in positions where the subjects’ ECG readings
data were recorded with a sampling rate of 200 Hz from could not be seen. Each of the three noise signals contains
100 patients with paroxysmal AF. The proposed dataset was different dominant noise components, which are baseline

54412 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

wander (BW), motion artifact (MA), and electrode motion Algorithm 1 Noise Addition to Signal Clips
(EM). These three noises are the most typical in ECG signal Input: clean ECG signal s(n), 0 < n < N
recordings. BW noise is caused by electrode misplacement, Output: noised ECG signal x(n), 0 < n < segment_size
unwell-prepared skin, subject breathing movement, and 1: for i = 0 to (N−segment_size) step stride=512 do
insufficient contact between electrode cables [49]. These 2: signal_clip ← signal[i : i + segment_size, :]
types of noise range from 0.05 to 3 Hz. In addition, MA noise 3: em_clip ← em_noise[i : i + segment_size, :]
is usually generated by the patient’s body movement; thus, 4: ma_clip ← ma_noise[i : i + segment_size, :]
it interferes with the ECG signal features [50]. Finally, 5: bw_clip ← bw_noise[i : i + segment_size, :]
EM noise is commonly caused by the movement of the 6: if clip_size < segment_size then
electrodes. This is often considered the most problematic 7: break
noise because it appears as an ectopic beat. This makes EM 8: end if
noise more difficult to remove by simple filters than other 9: r1, r2 ← two random numbers with sum equal to 1
types of noise. 10: a ← r1
11: b ← r2 − r1
C. NOISED ECG DATASET GENERATION 12: c ← 1 − r2
The MIT-BIH Arrhythmia dataset and the MIT-BIH Noise 13: snr ← random.randint(−5, 10)
Test Stress dataset have a sampling rate of 360 Hz. The 14: noise_clip ← (a × em_clip) + (b × ma_clip) + (c ×
SHDB-AF dataset was sampled at 200 Hz; thus, the signal bw_clip)
in that dataset was resampled to the MIT dataset sampling 15: power_sclean ← equation (2)
rate using the Fourier method [51]. This synchronized step 16: power_snoise ← equation (3)
ensures the temporal consistency of the input to the noise 17: noise_coef ← equation (4)
reduction model. After the sampling rates were synchronized, 18: noised_signal_clip ← equation (5)
we generated a noised ECG signal by the Algorithm 1. This 19: Save noised_signal_clip and signal_clip
algorithm was inspired by the use of the MIT-BIH Noise Test 20: end for
Stress dataset in their original paper [27]. While the original
approach to generate a noise signal was by defining a certain
SNR value, we randomized the SNR between a certain range.
This was done to obtain more diverse and various types of Arrhythmia [26] dataset and the SHDB-AF dataset [46].
noise signals. After all the samples from the mentioned dataset have been
From the dataset, we obtained BW (ηBW ), MA (ηMA ), and processed using the process in subsection IV-C, we choose all
EM (ηEM ) noise. Inspired by this research [24], [49], three the records with channel MLII from the MIT-BIH Arrhythmia
types of noises with specific weight a, b, c into a single noise dataset and 50 records (001-050 record) with channel ECG1
signal. The sum of a, b, c is equal to 1. Then, the combined from the SHDB-AF dataset to be segmented and added
noised signal was merged into the clean ECG signal from the with some noise using Algorithm 1. The length of each
MIT-BIH Arrhythmia or SHDB-AF dataset, so the combined noised-clean ECG segmented signal was 1024 data points,
signal had a signal-to-noise ratio (SNR) value between −5 dB or equal to 2.84 seconds in duration. We selected this number
and 10 dB as in [9]. This was achieved using the following based on the following considerations:
equation: • The lower limit of frequency noise we want to reduce is
N 0.05 Hz.
1 X
The lowest heart rate is around 0.67 Hz, but in practice,
Pclean = s2i (2) •
N if we use frequency domain analysis, 0.05 Hz is
i=1
N recommended as a cut-off frequency of a high-pass
1 X 2
Pnoise = ηi (3) filter [52].
N • The larger the last segment, the higher the time and
i=1
s computational complexity.
Pclean
K= (4)
SNR
10 10 · Pnoise The total number of segments generated from the MIT-BIH
Arrhythmia dataset was 119,192, and from the SHDB-AF
x(n) = s(n) + K [a · ηBW (n) + b · ηMA (n) + c · ηEM (n)]
dataset was 63,400. The generated noise-free ECG signals
(5) from the MIT-BIH Arrhythmia dataset are used as data for
training, validation, and testing at a ratio of 8:1:1. In contrast,
V. EXPERIMENT RESULTS AND DISCUSSION the generated signal from the SHDB-AF dataset was used for
A. EXPERIMENTAL SETUP AND DESIGN testing only. The proposed approach was designed to evaluate
1) EXPERIMENTAL SAMPLES the generalization capabilities of CAE-SCA, specifically, its
For the experiment in this research, we use 2 datasets as ability to perform effectively on unseen data that were not
resources of clean ECG signals: the MIT-BIH included during the training phase.

VOLUME 13, 2025 54413

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

TABLE 1. Experimental result model training (MIT-BIH Dataset). TABLE 3. Testing result on SHDB-AF dataset.

were configured to be 0.7 and 0.3, respectively, according

to the proposed configuration used in original research [53].
We should select a value β that is not too high because it can
TABLE 2. Testing result on MIT-BIH dataset. cause the denoised signal to become unsmooth. Then, mean
square error (MSE) is applied. The MSE loss is computed as
the square of the RMSE, as shown in (10).
v
u
u1 X N
RMSE = t [ŝi − si ]2 (10)
N
i=1

3) PERFORMANCE METRICS
Before the segmented, noise-free ECG signal enters the Researchers commonly use several performance measure-
model, a minimum-maximum normalization will be applied ment parameters, such as SNR, root mean square error
to the segmented signals with the following equation: (RMSE), and percent root mean square difference (PRD)
[18], [45], [53] to assess the performance of noise reduction
si − smin methods. SNR measures the ratio between the clean signal
s∗i = (6)
smax − smin and the noise in the signal. RMSE measures the variance
where s∗i is the normalized signal of the segmented input between the denoised and noise-free signals, and PRD
signal, smin , and smax are the minimum and maximum measures the quality of the signal recovery. This research
values of the segmented input signal, respectively. This aimed to achieve high SNR, low RMSE, and low PRD value.
normalization improves the convergence speed and stability The equations for SNR, RMSE, and PRD are in (11), (10),
of the model during the learning period [3]. and (12) respectively.
PN 2 !
s
2) LOSS FUNCTION SNRoutput = 10 · log10 PN i=1 i (11)
2
A distance loss function was utilized to evaluate global i=1 (ŝi − si )
v
similarity across the entire signal and the maximum loss u PN
u [ŝi − si ]2
function to evaluate local differences [24], [53]. (7) shows PRD = 100 · t i=1 PN 2 (12)
the distance loss function. i=1 si
v
u N In addition to model performance, computational costs
2
uX
l =t
dist ŝ − si i (7) were used to evaluate models that deliver good performance
i=1 with efficient computation. To assess computational cost,
lmax = max |ŝ1 − s1 |, |ŝ2 − s2 |, . . . , |ŝN − sN | we used FLOPs (floating point operations), memory size,


(8)
and inference time [54], [55]. FLOPs estimate the total
where ldist is the loss value using the maximum loss function; computational load by summing all arithmetic operations,
it ŝi is the ground truth signal’s sample at time/step i, and si such as addition, subtraction, division, and multiplication,
is the predicted clean signal output’s sample at time i. The in the algorithm on the model. In contrast, the inference time
maximum loss value generated in (8) is notated as lmax , which measures the actual time taken for model execution. The
is calculated by selecting the highest ground truth distance memory size is the estimated CPU memory consumption.
and predicted clean signal. The inference time and memory size are evaluated while
running the model on a Raspberry Pi device to simulate edge
Loss = λldist + βlmax (9)
computing deployment scenarios.
These two loss functions are then combined into a single
loss function by weighting them, as shown in (9), with the 4) TEST OF SIGNIFICANCE
first λ as the weight of the distance loss and the second β A test of significance was used to gather evidence from a
as the weight of the maximum loss. The value of λ and β sample of the population to evaluate a hypothesis regarding

54414 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 6. Signal denoising result visualizations.

TABLE 4. Performance comparison.

the population. In this study, the population consisted of normal distribution [56]. To strengthen the assumption of a
ECG signals, and the sample is the test dataset selected normal distribution, the graphical distribution of the SNR
randomly, as described in V-A1. The size of the test dataset values obtained from the CAE-SCA and the comparison
is 11,920 which is sufficiently large to apply the central limit model is examined. Based on these considerations, a right-
theorem. The theorem states that as the sample size increases, sided t-test was applied to determine whether the CAE-SCA
the sampling distribution of the sample mean approaches a leads to significantly higher SNR than another model.

VOLUME 13, 2025 54415

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

5) GENERAL HYPERPARAMETERS AND THE COMPUTING

PLATFORM
Several hyperparameters were used for all deep learning
models in this research. Each model was optimized using
the Adam optimizer with a learning rate of 0.0001 and batch
size of 32. The number of epochs for each training session
depends on the change in validation MSE loss between the
current and previous epochs. If the change is less than 0.0001
10 times in a row, the training is halted. All experiments,
including the training, validation, and testing of the models,
were conducted on an NVIDIA Tesla V100. We used Python
3.10 and PyTorch 2.5 as the machine learning framework.

B. MODEL EVALUATION
1) TRAINING RESULT
Table 1 is the overview of the training results for each tested
model. The results indicate varying levels of effectiveness
among different model architectures for ECG denoising on
the MIT-BIH dataset. The UNet model, which incorpo-
rates residual connections into its autoencoder architecture,
demonstrated lower training and validation losses than
the FCNDAE model, which lacks residual connections.
Among the autoencoder-based models with residual con-
nections, excluding CAE-SCA, the CNN-Attention Block
model achieved the lowest validation loss, outperforming
FIGURE 7. Contribution’s position visualization.
UNet, APTrans, and DRNN. These results indicate that the
integration of residual connections and attention mechanisms
in both the channel and spatial dimensions significantly
contributes to improving the performance of denoising which indicate an SNR of 16.187 ± 5.128 dB, RMSE of
models. The recurrent models, including GRU and LSTM, 0.059 ± 0.041, and PRD of 18.529 ± 12.306% on MIT-
demonstrate reasonably balanced performance, with LSTM BIH. The performance of CNN-Attention Block and LSTM
achieving the lowest validation loss within this group. This
was marginally worse. The performance of APTrans, DRNN,
suggests that recurrent models can capture temporal or spatial FCNDAE, GRU, and UNet was below average. The best
dependencies in ECG signals, which is crucial for effective SNR, RMSE, and PRD values were obtained by CAE-SCA
denoising. on SHDB-AF, demonstrating its resilience and versatility in
The proposed CAE-SCA, which combines CNN, LSTM, unknown datasets. Fig. 6 displays the example segmented
and channel attention block architectures, achieved the lowest
signal following denoising using the tested models.
loss values overall, which indicates that it is the most effective Each model’s predicted signal was compared against
method for ECG denoizing in this study. Combining the the input and ground truth signals to assess the accuracy
spatial feature extraction capabilities of CNNs with the of peak detection, noise suppression, and overall fidelity.
LSTM’s ability to capture long-term dependencies enables The CAE-SCA demonstrated superior performance, closely
the CAE-SCA to effectively address the complexities of ECG aligning with the ground truth signal while effectively
signals, leading to superior denoising performance. In addi-
capturing peaks and subtle variations. This model excelled at
tion, incorporating a channel attention block alongside LSTM reducing noise and maintained consistency even in regions
within skip or residual connections allows the model to of sudden signal spikes, highlighting its robustness and
effectively handle both temporal and channel dependencies. adaptability to dynamic signal changes.
The CAE-SCA model is the best-performing model, followed Despite experiencing slight lags in high-noise regions,
by CNN-Attention Block and LSTM, which also demonstrate particularly around sample number 800, the U-Net model
strong generalization.
effectively detected peaks and preserved the overall signal
structure, with notable accuracy observed near sample
2) TESTING RESULT numbers 200 and 600, as illustrated in Fig. 6h. Although
To investigate generalization capabilities, we used a reserved it had trouble with abrupt changes, APTrans produced
10% of the MIT-BIH dataset and the unseen SHDB-AF smooth forecasts with good noise reduction, as shown in
dataset. CAE-SCA performed better than other models, Fig. 6b. Accurate peak detection and balanced performance
as evidenced by the results in Table 2 and Table 3, were attained by the CNN attention block, whereas minor

54416 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 8. Histogram of SNR Values for MIT-BIH Test dataset.

variations were observed in low-amplitude areas close to are critical. Future research could focus on hybrid approaches
sample number 700 as visualized in Fig. 6c. Although it that integrate the strengths of these models to further improve
created artifacts in low-amplitude regions, the DRRN model the model performance.
successfully recorded peaks, as shown in Fig. 6d. FCN-
DAE, on the other hand, did not perform well, generating
outputs that were flat and did not include important peaks, 3) COMPUTATIONAL COST EVALUATION
as visualized in Fig. 6e. Peak reconstruction was a strong suit On the MIT-BIH and SHDB-AF datasets, a performance
for both the GRU and LSTM models, with the GRU handling comparison was performed to assess the computational
noise well and the LSTM performing marginally better in efficiency and SNR. With SNR values of 16.187 (MIT-BIH)
dynamic areas, as shown in Fig. 6f and Fig. 6g. and 15.308 (SHDB-AF), CAE-SCA is perfect for real-time
Overall, the CAE-SCA emerged as the most reliable edge applications. The proposed method requires only
choice for signal reconstruction, as shown in Fig. 6a, closely 437.029 MFLOPs, 117.83 K parameters, and 345.125 ms
followed by GRU and LSTM, which also demonstrated inference time. Although APTrans provides competitive SNR
strong capabilities. U-Net and CNN attention blocks demon- values (11.989 and 13.606), its practicality is limited by
strated promise but exhibited minor weaknesses in noise- its requirements of 12.497 G FLOPs, 6.1 M parameters,
sensitive regions, whereas the FCN-DAE model struggled and 963.174 ms inference time. The CNN-Attention Block
significantly with the task. These findings underscore the balances efficiency and performance by achieving SNR
importance of model selection in signal reconstruction values of 14.089 and 14.317 with 166.453 M FLOPs, 76.24 K
applications, particularly when accuracy and noise resilience parameters, and 380.911 ms inference time. Although they

VOLUME 13, 2025 54417

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 9. Histogram of SNR Values for SHDB-AF Test dataset.

struggle with complexity, GRU (6.522 G FLOPs, 199.69 K second (FPS) value based on inference time in Raspberry
parameters, 2248.219 ms) and LSTM (9.065 G FLOPs, Pi 5 in both MITBIH and SHDB-AF datasets, respectively.
277.62 K parameters, 8836.908 ms) provide moderate The word frame refers to the segment for each input fed to
performance. the model in each inference. The FPS value is calculated
The other models with poor performance included DRNN, by dividing one by the inference time, resulting the total
FCNDAE, and UNet. With 828.375 M FLOPs, 25.54 K inference frequency happened in one second. These results
parameters, and a quick inference time of 214.394 ms, shed light on the trade-off between denoising performance
the proposed DRNN obtained SNR values of 9.219 and (SNR) and computational efficiency. On both datasets, the
10.613. With 509.805 M FLOPs, 80.223 K parameters, CAE-SCA achieved a notable balance, exhibiting high SNR
and 289.449 ms inference, FCNDAE exhibits poor SNR while maintaining competitive FPS. Overall, the CAE-SCA
values (3.533 and 3.688). Despite its quick inference obtained the highest SNR value compared to the other
(4382.789 ms) and low parameters (1.84 M), UNet yields methods. Although it is still behind FCNDAE and DRRN
moderate outcomes but requires 13.482 G FLOPs. With all in terms of FPS, it still achieves commendable performance
the factors considered, CAE-SCA exhibits the best balance with approximately 3 FPS, which suggests its suitability for
between accuracy, efficiency, and inference speed, which real-time application in embedded devices.
makes it ideal for use in settings with limited resources.
Fig. 7 represents the position of the CAE-SCA’s contri- 4) TEST OF SIGNIFICANCE EVALUATION
bution compared to other models by visualizing the relation The proposed CAE-SCA achieved the best average value
between the obtained SNR value and calculated frames per across all metrics in the model performance evaluation.

54418 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

FIGURE 10. Hardware implementation test result.

TABLE 5. Significance test result. As shown in Fig. 10, the raw signal (above) contains a very
noisy signal. We cannot find the ECG signal features P, Q, R,
S, and T easily. Meanwhile, the filtered signal (bottom) shows
a clear visualization of a common ECG signal, allowing the
signal to be analyzed more easily.

VI. CONCLUSION
This study aims to develop a novel, strong, and efficient
denoising autoencoder architecture for edge devices. The
We conducted a significance test to provide evidence performance experiments were conducted on both a cloud
supporting the superiority of the CAE-SCA over the com- environment and Raspberry Pi 5, with the latter serving as a
pared models. Fig. 8 and Fig. 9 illustrate the differences representative edge device for testing in the context of Holter
in histograms between the CAE-SCA and other models, monitor implementation. While maintaining competitive
revealing that the distribution of SNR values for each model computational efficiency, the CAE-SCA has outperformed
approximates a normal distribution; however, the CAE-SCA other state-of-the-art models in terms of model performance
exhibits a higher mean. This allowed us to perform a metrics. The incorporation of LSTM in the skip connections
right-sided independent sample t-test. The results, shown as a sequential attention mechanism enables CAE-SCA to
in Table 5, indicate that all p-values for SNR comparisons effectively capture the causal dependencies of ECG signals.
between the CAE-SCA and each baseline model are well The CNN-based channel attention mechanism and the overall
below the chosen significance level of 0.001. This confirms autoencoder architecture address the quasi-periodic nature of
the statistically significant difference in the mean SNR values ECG signals, thereby enhancing the model’s performance in
between the CAE-SCA and the other models. In particular, denoising tasks. In the future, it will be beneficial to test
CAE-SCA leads to significantly higher SNR than the the ECG acquisition circuit under different user activities
alternative models. and disturbances. For instance, testing during walking,
sitting, standing, and other activities would provide a more
C. REAL HARDWARE IMPLEMENTATION TEST comprehensive validation of the CAE-SCA’s robustness and
As stated previously, we would like to implement the reliability.
proposed model in an embedded Holter monitor. We powered
our ECG acquisition circuit with two 9-volt batteries for the ACKNOWLEDGMENT
positive and negative signal ranges. The Raspberry Pi was The authors would like to express their gratitude to ChatGPT
also powered as the main processor with a power bank. In the from OpenAI for providing valuable assistance in refining
test scenario, we used a smartphone as the end-user device their manuscript. The tool’s suggestions for language clarity,
to visualize the signal. Fig. 10 shows the visualization of the structure, and coherence significantly enhanced the quality of
real raw captured signal from our ECG acquisition device and the study. Additionally, they extend their sincere appreciation
the filtered signal using our proposed ECG denoising model. to the Katalis Program from the Ministry of Education,

VOLUME 13, 2025 54419

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

Culture, Research, and Technology of Indonesia for their [19] J. Wang, S. Pei, Y. Yang, and H. Wang, ‘‘Convolutional transformer-
generous support. While the final interpretations and analyses driven robust electrocardiogram signal denoising framework with adaptive
parametric ReLU,’’ Math. Biosci. Eng., vol. 21, no. 3, pp. 4286–4308,
remain their own, support from both sources was instrumental 2024.
in the completion of this research. [20] K. Antczak, ‘‘Deep recurrent neural networks for ECG signal denoising,’’
2018, arXiv:1807.11551.
[21] M. Dias, P. Probst, L. Silva, and H. Gamboa, ‘‘Cleaning ECG with deep
REFERENCES learning: A denoiser tested in industrial settings,’’ Social Netw. Comput.
[1] BMC Public Health—Bmcpublichealth.biomedcentral.com. Accessed: Sci., vol. 5, no. 6, p. 699, Jul. 2024.
Dec. 17, 2024. [Online]. Available: https://bmcpublichealth.biomed [22] I. Banerjee, Y. Ling, M. C. Chen, S. A. Hasan, C. P. Langlotz,
central.com/ N. Moradzadeh, B. Chapman, T. Amrhein, D. Mong, D. L. Rubin, O. Farri,
[2] A. Rasti-Meymandi and A. Ghaffari, ‘‘A deep learning-based framework and M. P. Lungren, ‘‘Comparative effectiveness of convolutional neural
for ECG signal denoising based on stacked cardiac cycle tensor,’’ Biomed. network (CNN) and recurrent neural network (RNN) architectures for
Signal Process. Control, vol. 71, Jan. 2022, Art. no. 103275. radiology text report classification,’’ Artif. Intell. Med., vol. 97, pp. 79–88,
[3] X. Wang, B. Chen, M. Zeng, Y. Wang, H. Liu, R. Liu, L. Tian, and Jun. 2019.
X. Lu, ‘‘An ECG signal denoising method using conditional generative [23] Z. Chen, M. Wang, M. Zhang, W. Huang, H. Gu, and J. Xu, ‘‘Post-
adversarial net,’’ IEEE J. Biomed. Health Informat., vol. 26, no. 7, processing refined ECG delineation based on 1D-UNet,’’ Biomed. Signal
pp. 2929–2940, Jul. 2022. Process. Control, vol. 79, Jan. 2023, Art. no. 104106.
[4] A. Azzouz, B. Bengherbia, P. Wira, N. Alaoui, A. Souahlia, M. Maazouz, [24] W. Chorney, H. Wang, L. He, S. Lee, and L.-W. Fan, ‘‘Convolutional block
and H. Hentabeli, ‘‘An efficient ECG signals denoising technique based on attention autoencoder for denoising electrocardiograms,’’ Biomed. Signal
the combination of particle swarm optimisation and wavelet transform,’’ Process. Control, vol. 86, Sep. 2023, Art. no. 105242.
Heliyon, vol. 10, no. 5, Mar. 2024, Art. no. e26171. [25] P. Singh and A. Sharma, ‘‘Attention-based convolutional denoising
[5] Cardiovascular Diseases (CVDs)—Who.int. Accessed: Dec. 17, 2024. autoencoder for two-lead ECG denoising and arrhythmia classification,’’
[Online]. Available: https://www.who.int/news-room/fact- IEEE Trans. Instrum. Meas., vol. 71, pp. 1–10, 2022.
sheets/detail/cardiovascular-diseases-(cvds) [26] G. B. Moody and R. G. Mark, ‘‘MIT-BIH arrhythmia database,’’
[6] W. S. P. Harmadha, F. R. Muharram, R. S. Gaspar, Z. Azimuth, Tech. Rep., 1992.
H. A. Sulistya, F. Firmansyah, C. E. C. Z. Multazam, M. Harits, [27] G. B. Moody, W. E. Muldrow, and R. G. Mark, ‘‘The MIT-BIH noise stress
and R. M. Putra, ‘‘Explaining the increase of incidence and mortality test database,’’ Tech. Rep., 1992.
from cardiovascular disease in Indonesia: A global burden of disease
[28] O. Ronneberger, P. Fischer, and T. Brox, ‘‘U-Net: Convolutional networks
study analysis (2000–2019),’’ PLoS ONE, vol. 18, no. 12, Dec. 2023,
for biomedical image segmentation,’’ 2015, arXiv:1505.04597.
Art. no. e0294128.
[29] A. E. W. Johnson, T. J. Pollard, L. Shen, L.-W.-H. Lehman, M. Feng,
[7] S.-K. Chua, L.-C. Chen, L.-M. Lien, H. Lo, Z.-Y. Liao, S.-P. Chao,
M. Ghassemi, B. Moody, P. Szolovits, L. Anthony Celi, and R. G. Mark,
C.-Y. Chuang, and C.-Z. Chiu, ‘‘Comparison of arrhythmia detection
‘‘MIMIC-III, a freely accessible critical care database,’’ Sci. Data, vol. 3,
by 24-hour Holter and 14-day continuous electrocardiography patch
no. 1, May 2016, Art. no. 160035.
monitoring,’’ Acta Cardiologica Sinica, vol. 36, no. 3, pp. 251–259,
May 2020. [30] M. Zhao, S. Zhong, X. Fu, B. Tang, S. Dong, and M. Pecht, ‘‘Deep
residual networks with adaptively parametric rectifier linear units for fault
[8] P. M. Tripathi, A. Kumar, R. Komaragiri, and M. Kumar, ‘‘A review on
diagnosis,’’ IEEE Trans. Ind. Electron., vol. 68, no. 3, pp. 2587–2597,
computational methods for denoising and detecting ECG signals to detect
Mar. 2021.
cardiovascular diseases,’’ Arch. Comput. Methods Eng., vol. 29, no. 3,
pp. 1875–1914, May 2022. [31] G. B. Moody and R. G. Mark, ‘‘MIT-BIH atrial fibrillation database,’’
[9] S. Chatterjee, R. S. Thakur, R. N. Yadav, L. Gupta, and D. K. Raghuvanshi, Tech. Rep., 1992.
‘‘Review of noise removal techniques in ECG signals,’’ IET Signal [32] L. Xiu and Z. Li, ‘‘Low-power instrumentation amplifier IC design
Process., vol. 14, no. 9, pp. 569–590, Dec. 2020. for ECG system applications,’’ in Proc. Eng., vol. 29, pp. 1533–1538,
[10] P. Bing, W. Liu, Z. Zhai, J. Li, Z. Guo, Y. Xiang, B. He, and L. Zhu, Jan. 2012.
‘‘A novel approach for denoising electrocardiogram signals to detect [33] A. Wong, K.-P. Pun, Y.-T. Zhang, and C.-S. Choy, ‘‘An ECG measurement
cardiovascular diseases using an efficient hybrid scheme,’’ Frontiers IC using driven-right-leg circuit,’’ in Proc. IEEE Int. Symp. Circuits Syst.,
Cardiovascular Med., vol. 11, Apr. 2024, Art. no. 1277123. Apr. 2006, pp. 345–348.
[11] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, ‘‘Extracting and [34] W. Wang, ‘‘Design of low-noise low-power ECG amplifier circuit with
composing robust features with denoising autoencoders,’’ in Proc. 25th Int. high integration level,’’ J. Phys., Conf. Ser., vol. 2649, no. 1, Nov. 2023,
Conf. Mach. Learn. (ICML), New York, NY, USA, 2008, pp. 1096–1103. Art. no. 012062.
[12] D. Chen, S. Shi, X. Gu, and B. Shim, ‘‘Robust DoA estimation using [35] B. B. Winter and J. G. Webster, ‘‘Driven-right-leg circuit design,’’ IEEE
denoising autoencoder and deep neural networks,’’ IEEE Access, vol. 10, Trans. Biomed. Eng., vol. BME-30, no. 1, pp. 62–66, Jan. 1983.
pp. 52551–52564, 2022. [36] V. K. Singh, W.-G. Ho, and R. Gharpurey, ‘‘A frequency-folded ADC
[13] M. Scarpiniti, S. S. Ahrabi, E. Baccarelli, L. Piazzo, and A. Momenzadeh, channelizer with digital equalization and relaxed anti-alias filtering,’’
‘‘A novel unsupervised approach based on the hidden features of deep IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 65, no. 7, pp. 2304–2317,
denoising autoencoders for COVID-19 disease detection,’’ Expert Syst. Jul. 2018.
Appl., vol. 192, Apr. 2022, Art. no. 116366. [37] H. J. Landau, ‘‘Sampling, data transmission, and the Nyquist rate,’’ Proc.
[14] C. Gheller and F. Vazza, ‘‘Convolutional deep denoising autoencoders IEEE, vol. 55, no. 10, pp. 1701–1706, 1967.
for radio astronomical images,’’ Monthly Notices Roy. Astronomical Soc., [38] C. Chang, ‘‘Maximally flat amplitude low-pass filter with an arbitrary
vol. 509, no. 1, pp. 990–1009, Oct. 2021. number of pairs of real frequency transmission zeros,’’ IEEE Trans. Circuit
[15] Z. Shang, L. Sun, Y. Xia, and W. Zhang, ‘‘Vibration-based damage Theory, vol. CT-15, no. 4, pp. 465–467, Apr. 1968.
detection for bridges by deep convolutional denoising autoencoder,’’ [39] A. W. Ramadhan, A. Wijayanto, and H. Oktavianto, ‘‘Implementation of
Struct. Health Monitor., vol. 20, no. 4, pp. 1880–1903, Jul. 2021. audio event recognition for the elderly home support using convolutional
[16] H.-T. Chiang, Y.-Y. Hsieh, S.-W. Fu, K.-H. Hung, Y. Tsao, and S.-Y. Chien, neural networks,’’ in Proc. Int. Electron. Symp. (IES), Sep. 2020,
‘‘Noise reduction in ECG signals using fully convolutional denoising pp. 91–95.
autoencoders,’’ IEEE Access, vol. 7, pp. 60806–60813, 2019. [40] N. P. Martono and H. Ohwada, ‘‘Evaluating the impact of windowing
[17] H. Rahman, T. F. N. Bukht, A. Imran, J. Tariq, S. Tu, and A. Alzahrani, techniques on Fourier transform-preprocessed signals for deep learning-
‘‘A deep learning approach for liver and tumor segmentation in CT images based ECG classification,’’ Hearts, vol. 5, no. 4, pp. 501–515, Oct. 2024.
using ResUNet,’’ Bioengineering, vol. 9, no. 8, p. 368, Aug. 2022. [41] S. N. Ali, S. B. Shuvo, Md. I. S. Al-Manzo, A. Hasan, and T. Hasan,
[18] Y. Hou, R. Liu, M. Shu, and C. Chen, ‘‘An ECG denoising method based ‘‘An end-to-end deep learning framework for real-time denoising of heart
on adversarial denoising convolutional neural network,’’ Biomed. Signal sounds for cardiac disease detection in unseen noise,’’ IEEE Access,
Process. Control, vol. 84, Jul. 2023, Art. no. 104964. vol. 11, pp. 87887–87901, 2023.

54420 VOLUME 13, 2025

A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

[42] Alsaade and Bhuiyan, ‘‘On development of fuzzy filter for de-noising ECG SYIFA KUSHIRAYATI received the bachelor’s
signal,’’ Int. J. Comput. Sci. Netw. Secur., vol. 20, no. 12, pp. 252–257, degree in biomedical engineering from Bandung
Dec. 2020. Institute of Technology, Indonesia. She is currently
[43] H. Sak, A. Senior, and F. Beaufays, ‘‘Long short-term memory based a Research Assistant with the Faculty of Computer
recurrent neural network architectures for large vocabulary speech Science, University of Indonesia. Her research
recognition,’’ 2014, arXiv:1402.1128. interests include digital signal processing and
[44] H. Li and P. Boulanger, ‘‘A survey of heart anomaly detection using artificial intelligence.
ambulatory electrocardiogram (ECG),’’ Sensors, vol. 20, no. 5, p. 1461,
Mar. 2020.
[45] I. Hermawan, N. Sevani, A. F. Abka, and W. Jatmiko, ‘‘Denoising ambula-
tory electrocardiogram signal using interval dedependent thresholds based
stationary wavelet transform,’’ JOIV, Int. J. Informat. Visualizat., vol. 8,
no. 2, p. 742, May 2024.
[46] K. Tsutsui, S. B. Brimer, N. Ben-Moshe, J. M. Sellal, J. Oster, H. Mori,
Y. Ikeda, T. Arai, S. Nakano, R. Kato, and J. A. Behar, ‘‘SHDB-
AF: A Japanese Holter ECG database of atrial fibrillation,’’ 2024, SALSABILA AURELLIA received the bachelor’s
arXiv:2406.16974. degree in biomedical engineering from Airlangga
[47] X. Gu, J. Hu, L. Zhang, J. Ding, and F. Yan, ‘‘An improved method with University and the master’s degree in biomedical
high anti-interference ability for R peak detection in wearable devices,’’ engineering from the University of Indonesia. She
IRBM, vol. 41, no. 3, pp. 172–183, Jun. 2020. is currently a Research Assistant with the Faculty
[48] H. Li, G. Ditzler, J. Roveda, and A. Li, ‘‘DeScoD-ECG: Deep score- of Computer Science, University of Indonesia. Her
based diffusion model for ECG baseline wander and noise removal,’’ IEEE research interests include artificial intelligence and
J. Biomed. Health Informat., vol. 28, no. 9, pp. 5081–5091, Sep. 2024. biomedical signal and image processing.
[49] F. P. Romero, D. C. Piñol, and C. R. Vázquez-Seisdedos, ‘‘DeepFilter:
An ECG baseline wander removal filter using deep learning techniques,’’
Biomed. Signal Process. Control, vol. 70, Sep. 2021, Art. no. 102992.
[50] S. Banerjee and G. K. Singh, ‘‘Monte Carlo filter-based motion artifact
removal from electrocardiogram signal for real-time telecardiology
system,’’ IEEE Trans. Instrum. Meas., vol. 70, pp. 1–10, 2021.
[51] W. G. Hawkins, ‘‘Fourier transform resampling: Theory and application,’’
IEEE Trans. Nucl. Sci., vol. 44, no. 4, pp. 1543–1551, Aug. 1997. MGS M. LUTHFI RAMADHAN received the
[52] P. Kligfield, L. S. Gettes, J. J. Bailey, R. Childers, B. J. Deal, bachelor’s degree from the Faculty of Computer
E. W. Hancock, G. van Herpen, J. A. Kors, P. Macfarlane, D. M. Mirvis, Science, Sriwijaya University, and the master’s
O. Pahlm, P. Rautaharju, and G. S. Wagner, ‘‘Recommendations for the degree from the Faculty of Computer Science,
standardization and interpretation of the electrocardiogram,’’ Circulation, University of Indonesia. He is currently a Research
vol. 115, no. 10, pp. 1306–1324, Mar. 2007. Assistant with the Faculty of Computer Science,
[53] J. Wang, R. Li, R. Li, K. Li, H. Zeng, G. Xie, and L. Liu, ‘‘Adversarial de- University of Indonesia. His research interests
noising of electrocardiogram,’’ Neurocomputing, vol. 349, pp. 212–224, include deep learning, pattern recognition, and
Jul. 2019. computer vision.
[54] R. Desislavov, F. Martínez-Plumed, and J. Hernández-Orallo, ‘‘Trends in
AI inference energy consumption: Beyond the performance-vs-parameter
laws of deep learning,’’ Sustain. Computing: Informat. Syst., vol. 38,
Apr. 2023, Art. no. 100857.
[55] C. Scheerlinck, H. Rebecq, D. Gehrig, N. Barnes, R. E. Mahony,
and D. Scaramuzza, ‘‘Fast image reconstruction with an event camera,’’
in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV), Mar. 2020,
MUHAMMAD HANNAN HUNAFA received the
pp. 156–163.
bachelor’s degree from the Faculty of Informat-
[56] J. L. Devore, Probability and Statistics for Engineering and the Sciences,
9th ed., Boston, MA, USA: Cengage Learning, 2016. ics, Telkom Institute of Technology Purwokerto.
He is currently pursuing the master’s degree with
the Faculty of Computer Science, University of
Indonesia. His research interests include machine
learning, deep learning, and computer vision.

MUHAMMAD FEBRIAN RACHMADI is cur-

rently a Faculty Member with the Faculty of
ALIF WICAKSANA RAMADHAN received the Computer Science, University of Indonesia. He is
bachelor’s and master’s degrees from the Depart- also affiliated as a Postdoctoral Research Scientist
ment of Electronics Engineering, Electronic Engi- with the Brain Image Analysis (BIA) Unit, RIKEN
neering Polytechnic Institute of Surabaya. He is Center of Brain Science (RIKEN CBS), Wako,
currently a Research Assistant with the Faculty Japan, working together with Dr. Henrik Skibbe,
of Computer Science, University of Indonesia. and the Team Leader of the BIA Unit, to develop
His research interests include computer vision, machine/deep learning techniques for processing
robotics, and machine learning. and analyzing marmoset brain image data. His
main research interests include medical image analysis and computation
using data-driven methods, such as deep learning algorithms.

VOLUME 13, 2025 54421

 A. W. Ramadhan et al.: CAE-SCA for Robust ECG Signal Denoising With Edge Device Implementation

APRINALDI JASA MANTAU (Member, IEEE) SATRIA MANDALA (Member, IEEE) received
received the Ph.D. degree in computer science the Bachelor of Science degree in electrical
and systems engineering from Kyushu Institute engineering from Brawijaya University, Indonesia,
of Technology, Japan, in 2024. He is currently a in 1997, and the master’s degree in computer
Lecturer with the Faculty of Computer Science, science with a specialization in network security
University of Indonesia. He is a Researcher and and the Ph.D. degree in computer science, with a
a Lecturer in the field of computer science. focus on wireless network security from Universiti
His research interests include machine learning, Teknologi Malaysia (UTM), in 2006 and 2011,
computer vision, robotics, data mining, swarm respectively. He has also held the position of the
intelligence, and swarm robotics. Dean of the Faculty of Information and Industrial
Technology, Telkom Institute of Technology Surabaya. He is currently the
Director of the Human-Centric Engineering, Telkom University, Indonesia.
He is also a Lecturer with the School of Computing, Telkom University.

WISNU JATMIKO (Senior Member, IEEE)

received the B.S. degree in electrical engineering
and the M.Sc. degree in computer science from
SITI NURMAINI (Member, IEEE) received the the University of Indonesia, Depok, Indonesia,
master’s degree in control systems from Institut in 1997 and 2000, respectively, and the Dr.Eng.
Teknologi Bandung (ITB), Indonesia, in 1998, degree from Nagoya University, Japan, in 2007.
and the Ph.D. degree in computer science from He is currently a Full Professor with the Faculty
Universiti Teknologi Malaysia (UTM), in 2011. of Computer Science, University of Indonesia.
She is currently a Professor with the Faculty His research interests include autonomous robots,
of Computer Science, Sriwijaya University. Her optimization, real-time traffic monitoring systems,
research interests include biomedical engineering, machine learning, and artificial intelligence. From 2019 to 2020, he was the
deep learning, machine learning, image process- Chairman of the IEEE Indonesia Section.
ing, control systems, and robotics.

54422 VOLUME 13, 2025