| Journal of Smart Sensors and Computing
Received: 20 April 2026; Revised: 22 May 2026; Accepted: 25 May 2026; Published Online: 03 June 2026.
J. Smart Sens. Comput., 2026, 2(2), 26206 | Volume 2 Issue 2 (June 2026) | DOI: https://doi.org/10.64189/ssc.26206
© The Author(s) 2026
This article is licensed under Creative Commons Attribution NonCommercial 4.0 International (CC-BY-NC 4.0)
An Intelligent Low-Cost Digital Stethoscope Using
Arduino, Signal Processing, and Machine Learning for
Real-Time Heart Sound Analysis
Nafisa M. Mapari, Alefiya J. Managori,
*
Hajra A. Naik, Monika G. Chaurasiya and Tasmiya I. Kandhal
Department of Computer Engineering, Anjuman-I-Islam M. H. Saboo Siddik College of Engineering, Mumbai, Maharashtra, 400008,
India
*Email: alefiya.231233.co@mhssce.ac.in (Alefiya J. Managori)
Abstract
Cardiovascular diseases remain among the leading causes of mortality worldwide, creating a growing need for
affordable and accessible cardiac monitoring systems. Traditional acoustic stethoscopes rely heavily on
physician expertise and do not provide capabilities such as signal visualization, storage, or automated analysis.
To address these limitations, this paper presents an intelligent low-cost digital stethoscope system using
Arduino, signal processing, and machine learning for real-time heart sound analysis. The proposed system uses
an Arduino UNO microcontroller integrated with a modified KY-037 heart sound acquisition module and an
external electret microphone coupled with a stethoscope chest piece for phonocardiogram acquisition. The
acquired signals are transmitted through USB serial communication and processed using a multistage signal
preprocessing pipeline consisting of median filtering, Butterworth bandpass filtering, moving average
smoothing, and outlier suppression to improve signal quality and reduce environmental noise. The processed
heart sound signals are analyzed for real-time heart rate estimation using adaptive peak detection techniques.
In addition, a machine learning-based cardiac sound classification module is implemented using a random
forest classifier trained on the PhysioNet CirCor DigiScope phonocardiogram dataset. Feature extraction is
performed using Mel Frequency Cepstral Coefficients (MFCCs), spectral features, chroma features, and signal
energy parameters. The developed system also includes a Streamlit-based monitoring dashboard for live
waveform visualization, BPM tracking, patient record management, data storage, CSV export, and PDF report
generation. Experimental evaluation demonstrated successful real-time heart sound acquisition and an overall
classification accuracy of 82%, with effective waveform visualization and stable BPM estimation. The proposed
system provides a compact, economical, and scalable solution for intelligent cardiac monitoring and has
potential applications in low-resource healthcare environments, telemedicine, and biomedical education.
Keywords: Digital stethoscope; Arduino UNO, Phonocardiogram; Signal processing; Machine learning; Random Forest
classifier; Cardiac monitoring.
1. Introduction
Cardiovascular disease (CVD) is currently the number one cause of morbidity and mortality worldwide and is
thus a significant concern for all global healthcare systems. Early and ongoing detection of cardiovascular
conditions can help prevent morbidity and mortality due to cardiovascular complications. The initial diagnosis
of CVD has historically been achieved through an auscultative approach involving the use of noninvasive
techniques (i.e., using a stethoscope). The primary challenge with the use of a stethoscope is the need for an
appropriate skill level and experience of the clinician to provide an accurate diagnosis and the lack of ability to
visually display (display), record/store (store) and automate (automate) analyses of heart sounds from the
stethoscope.
[1]
If a clinician is unable to accurately diagnose cardiovascular disease using a stethoscope, other
diagnostic testing methods (other than a stethoscope) may lead to additional problems with respect to detecting
and diagnosing accurate cardiovascular disease.
[2]
New developments in both biomedical instrumentation and Intel-based systems have led to new types of digital
stethoscopes that can provide a greater ability to acquire and analyze heart sounds.
[2.3]
A digital stethoscope
converts the acoustic cardiac signal produced by the heart into an electrical signal that is capable of being
processed, displayed, and stored using standard computer-based techniques. Many studies have used
microphone modules, embedded digital computers, and other signal conditioning methods to acquire
phonocardiograms and monitor heart patients.
[3,4]
Additionally, the collaboration between universities and
private/nonprofit organizations has generated significant recognition concerning new types of low-cost
wearable medical devices that are based on open-source hardware platforms because of their low cost, small
size/weight, and ease of use.
[2,5]
Therefore, advancements in digital medical technology have created new
opportunities for new types of intelligent and affordable healthcare systems.
In recent years, many advancements have been made in how we investigate and analyze certain biological
signals, such as heart sounds, using signal processing and machine learning methods. Many different methods
have been used to filter and remove noise from recordings of heart sounds (phonocardiograms) so that better
quality recordings can be obtained and that noise introduced by the external environment can be reduced.
[6,7]
Recent work has demonstrated that machine learning and deep learning techniques show promise for the
automatic classification of heart sounds, as well as for the detection of abnormalities.
[3,8,9]
The availability of
numerous publicly available biological datasets, such as PhysioNet and the CirCor DigiScope database, has
allowed researchers to obtain standardized recordings of these heart sounds to develop and test algorithms for
intelligent cardiac monitoring systems.
[10,11,12]
Additionally, researchers are using advanced software
frameworks and visualization tools to create applications to monitor patients in real time, manage patient data
remotely and develop interactive dashboards for healthcare providers to enhance healthcare services.
[2,13]
As cost-effective smart medical devices have been developed to meet the growing demand for such devices, this
project presents a new low-cost digital stethoscope system that utilizes an Arduino microcontroller and Digital
Signal Processing (DSP) and Machine Learning (ML) algorithms to analyze heart sounds. The heart sound
acquisition system is based on the modified KY-037, which includes an external electret microphone for
collecting a PCG (phonocardiograms) using an Arduino UNO microcontroller. The retrieved PCG signals undergo
a detailed, multistage preprocessing pipeline that includes a median filter (MF), Butterworth bandpass filter
(BPF), moving average filter (MAF), and outlier removal to create quality heart sounds. Heart sounds are
classified by using the random forest algorithm through a previously developed ML classifier, which uses the
PhysioNet CirCor DigiScope database for training data collection.
[11]
The real-time PCG data collected are
displayed as a visual waveform on a Streamlit-based dashboard, along with additional functionality such as BPM
monitoring, patient record management, data storage and automated PDF report generation. Overall, this
proposed architecture provides a low-cost, compact, and portable solution for academic, research, and initial
patient screening purposes.
This study contributes greatly to the development of a low-cost intelligent digital stethoscope through the use
of open-source hardware/software technologies, developing a multistage heart sound preprocessing pipeline
that minimizes noise and improves the quality of a signal, providing a real-time estimate of the BPM through an
adaptive peak detection method, and classifying heart sounds using machine learning and the PhysioNet
database.
[11,12]
Additionally, the inclusion of a Streamlit monitoring Dashboard, which includes features such as
waveform visualization, patient record management, and automated PDF report generation, creates additional
capabilities for future smart healthcare solutions and allows for further scaling.
[2,13]
2. Methodology
A new intelligent digital stethoscope system has been produced using a combination of embedded hardware,
signal processing methods, machine learning algorithms, and visualization tools based on software for
monitoring and analyzing sounds produced by the heart. The design approach is based on using low-cost
methods for recording heart sounds, processing these data to make BPM estimates, using intelligent techniques
for classifying sounds, and providing interactive methods for monitoring patients through the use of open-
source solutions.
The architecture of the system consists of an acquisition module made from an Arduino UNO board connected
to a modified KY-037 microphone and an external electret microphone for recording phonocardiograms. The
recorded signals are processed using multiple filtering techniques and analyzed with machine learning
techniques that have been trained with the PhysioNet CirCor DigiScope dataset. A Streamlit-based dashboard
has also been developed to provide real-time visualization of the waveforms and to monitor the BPM, manage
patient information, and create reports automatically. The detailed methodology and stages of the
implementation of the proposed system are discussed in the subsequent subsections.
2.1 System overview
The intended intelligent digital stethoscope system was developed to enable heart sound acquisition in real
time, signal preprocessing, BPM (beats per minute) estimation, machine-learning-based classification, and
interactive monitoring via a software interface or dashboard. Initially, the heart sounds are captured using an
external electret microphone attached to the stethoscope chest piece and connected to a modified KY-037 sound
acquisition module. The Arduino UNO microcontroller acquires the analog signals via its analog input pins, and
these analog signals are converted to digital signals using the built-in (on-board) ADC (analog-to-digital
converter).
Once the signals have been acquired, they are transmitted via USB serial communication to a PC for further
signal processing and analysis. A multistage preprocessing pipeline consisting of median filtering, Butterworth
bandpass filtering, moving average smoothing, and outlier suppression is implemented to improve signal
quality and reduce environmental noise levels. After processing, the heart rate is estimated in the BPM via
adaptive peak detection methods.
The classification of processed heart sound signals is accomplished through the implementation of a machine
learning model trained on PhysioNet CirCor DigiScope phonocardiogram data. The model predicts whether the
acquired heart sound can be classified as either a normal or murmur. The processed heart sound signals, BPM,
prediction results, and patient information will be made available through an application built on the Streamlit
web framework, which will allow for visual representation of the waveforms, storage in a database, export to
CSV format, and report generation in PDF format. The complete workflow of this proposed intelligent digital
stethoscope is illustrated in the graphical abstract.
2.2 Working principle
The intelligent digital stethoscope system operates as follows: the heart sound sample is collected in real time
and then preprocessed, the BPM is estimated, and the data are classified via machine learning and then
visualized on a dashboard. The heart sounds collected are converted to a digital signal via an Arduino UNO
microcontroller and then passed through a series of filters to reduce noise and enhance the waveform. Next, the
processed signals are used for estimating BPM and for classifying the cardiac sounds using machine learning.
Finally, the results are visualized on a Streamlit-based dashboard with data storage and reporting capabilities.
2.2.1 Hardware
The proposed hardware architecture consists of an Arduino UNO microcontroller, a modified KY-037 sound
acquisition module, and one or more of the following peripherals: an external electret microphone(s) and a
stethoscope chest piece(s). The prototype hardware is connected to a compatible computer via USB
communication. A block diagram of the proposed prototype is shown in Fig. 1, while a picture of the
experimental hardware setup is shown in Fig. 2.
Fig. 1: Block diagram of the proposed Arduino-based digital stethoscope prototype.
The Arduino UNO microcontroller serves as both the central processing unit and the data acquisition unit for
the proposed system. The architecture of the Arduino UNO uses ATmega328P and has multiple analog input
channels to acquire real-time biomedical signals. The analog heart sound signal is converted to a digital value
ranging from 0 to 1023 by the integrated 10-bit analog-to-digital converter (ADC) within the Arduino.
[14]
After
acquiring the data, Arduino processes it (smoothing) and communicates with the computer over a serial
communication link using a baud rate of 9600 bps.
Sound will be captured with a customized KY-037 sound sensor, with the original microphone located onboard
KY-037 removed and an external electret microphone incorporated into a pair of stereo earphones. To provide
better acoustic coupling between the heart and microphone and increase the sensitivity to low-amplitude heart
sound signal events, the new microphone was added internally to the stethoscope chest piece. The stethoscope
chest piece will create a concentrated area in which acoustic vibrations produced by cardiac activity will have
an outlet to transmit sound, resulting in a better quality of sounds being received.
[4]
The signals received from the heart are transmitted to a laptop system via a USB connection with no
requirement for additional power, and the laptop system is used as the main processing and visualization
platform where signal preprocessing, BPM estimation, machine learning inference and dashboard visualization
are accomplished using Python-based software frameworks. The overall architecture is aligned with the vision
of an inexpensive intelligent digital stethoscope as well as recently proposed IoMT-based healthcare monitoring
systems in the surgical realm.
[13]
Fig. 2: Experimental hardware setup of the developed digital stethoscope system.
2.2.2 Signal acquisition and preprocessing
Heart sounds were recorded using both a modified KY-037 sound sensor module and an external electret
microphone through a stethoscope chest piece. The heart-generated acoustic vibrations were picked up by the
external microphone and then converted to analog electrical signals. The analog signals were sent to the A0
analog input pin on the Arduino UNO for digitization and subsequent processing. Since heart sounds are low-
amplitude biomedical signals, careful acquisition and preprocessing are required to reduce the amount of
environmental interference, motion artifacts, and signal instability. Studies involving phonocardiogram (PCG)
analysis have suggested that signal enhancement and denoising techniques should be used to improve the
quality and reliability of acquired cardiac sounds.
[6,7]
The Arduino UNO uses its onboard 10-bit analog-to-digital converter (ADC) to convert the analog heart sound
signal into digital values ranging from 0 to 1023. The ADC conversion process can be represented as follows:
where
represents the input analog voltage, denotes the digital ADC value, and
corresponds to the
reference voltage of the microcontroller.
The sampling frequency of the system was maintained at approximately 100 Hz using a 10 ms delay between
consecutive samples. The sampling frequency is calculated as follows:
where
represents the sampling frequency and
denotes the sampling interval.
The Arduino microcontroller carried out an 8-sample moving average smoothing (M.A.S.) directly on the device
to smooth out fluctuations and make the signal more stable at the point of acquisition prior to transmitting it to
the computer via a serial connection. The smoothed signal and BPM values were sent to the computer through
a USB-based serial connection in a signal/BPM data format. The smoothed signal and BPM values were then
continuously received with PySerial using a separate thread to allow for uninterrupted real-time signal
processing and visualization.
2.2.3 Signal processing pipeline
Environmental noise, motion artifacts, electrical interference, and sudden spikes in heart sound signals may all
impact acquired heart sound recordings. A multistage pipeline for preprocessing signals was developed to
enhance the basic characteristics of low-cost electret microphones and the effects of external acoustic
disturbances to provide reliable measurements of BPMs and accurate classifications based on machine learning.
Similar signal preprocessing techniques, denoising, and segmentation techniques have been used in the analysis
of phonocardiograms to enhance the quality of heart sound signals and improve the ability to interpret heart
sounds for diagnostic purposes.
[6,4]
The workflow of the signal-processing pipeline is shown in Fig. 3.
2.2.3.1 Median filtering
The first step of signal processing involved applying a median filter to the input signal to reduce the influence
of sudden impact noise and transient spikes. A median filter with a kernel size of nine points was used to smooth
out the acquired waveform. As median filtering is a nonlinear filtering operation, it was able to remove singular
outlier points while maintaining the basic morphology of the heart sound waveform.
[4,6]
The application of
median filtering increased the stability of the signal and reduced sudden amplitude variations caused by
microphone handling and environmental disturbances.
2.2.3.2 Butterworth bandpass filtering
After the median filtering step was complete, the signal was filtered using a fourth-order Butterworth bandpass
filter with cutoff frequencies of 0.5 Hz (low) and 15.0 Hz (high). Low-frequency completion of the Butterworth
filter removed baseline drift and low-frequency motion artifacts, whereas high-frequency completion of the
Butterworth filter removed high-frequency electrical noise and unwanted acoustic interference from the signal.
The use of the scipy.signal.butter and filtfilt functions allowed zero-phase filtering of the signal without
distorting its waveform. Butterworth filtering techniques are very common methods for enhancing the
phonocardiogram signal and removing noise from biomedical signals to enhance the clarity and reliability of
the signal.
[7]
2.2.3.3 Moving average smoothing
Once the signal was bandpass filtered, a smoothing technique that uses a moving average filter was applied to
lessen the minor fluctuations and any residual ripples still present in the signal. The moving average filter has
made the cardiac waveform smoother, thus resulting in better consistency between cardiac cycles that have
occurred sequentially. This means that after being smoothed, the stability of the processed data is improved
when the beats per minute are estimated, thus increasing the reliability of the extracted features and providing
a better basis for machine learning analysis of the data collected.
2.2.3.4 Outlier clamping
During the last stage of preprocessing, outlier clamping was applied to the signal data to eliminate any
remaining extreme peaks or spikes prior to performing the waveform analysis. By clamping any values that
exceed ±3 standard deviations from the original mean amplitude, this operation suppresses transitory artifacts
and sudden peak conditions. This will also increase the consistency of the cardiac waveform and significantly
reduce false positive detections of peak values when the beats per minute are estimated. Other types of acoustic
signal processing noise reduction techniques have also been reported to provide a more reliable means of
processing phonocardiograms through various methods of signal stabilization.
[6,7]
Fig. 3: Schematic overview of the proposed heart sound signal processing pipeline.
2.2.4 Heart rate (BPM) estimation
Techniques for adaptive peak detection were applied to analyze heart sounds that were filtered to determine
the heart rate of the patient (BPM). To estimate the BPM, the envelope characteristics of the waveform were
filtered, and each peak point was identified as a corresponding point of a cardiac cycle. Methods for the
segmentation of heart sounds and passing through the peaks as described above have routinely been used
within the analysis of phonocardiograms (PCGs) to reliably detect cardiac events.
[4,6]
The peaks in the processed
heart sound signal were automatically found using the functionality available through the SciPy library to find
peaks within the dataset.
To improve the detection accuracy, an adaptive prominence threshold equal to 0.5 times the standard deviation
of the processed signal was employed. In addition, a minimum peak distance of 0.35 seconds was maintained to
avoid false detections and support heart rate measurements up to approximately 170 BPM. The BPM value was
calculated using the mean interpeak interval as follows:
The estimated BPM values were validated within a physiological range of 40180 BPM to eliminate unrealistic
measurements caused by noise or motion artifacts. The BPM results were continuously updated and visualized
in real time through the Streamlit monitoring dashboard.
2.2.5 Machine learning-based classification
2.2.5.1 Dataset description
A machine learning classification model was developed and trained using the PhysioNet CirCor DigiScope
phonocardiogram dataset.
[11,12]
The CirCor DigiScope database consists of 1507 heart sounds from 474 patients;
there are 2 major classes made up of recorded heart sounds, such as heart sounds and
heart sounds. Recordings have been validated clinically and annotated by physician
specialists, who have created a viable database for melodic cardiac sound analysis and machine learning-based
verification of heart sound categories.
2.2.5.2 Data balancing
In the original database, there is a significant imbalance between the number of normal and murmur records
present in the dataset, with a predominance of the normal category of recordings. To mitigate bias during
training toward the majority class, random oversampling was performed on the murmur class of recordings
until an equal number of records from both classes were represented in both the training and evaluation phases
of machine learning production, thus increasing classification fairness and reducing sensitivity to abnormal
heart sounds.
2.2.5.3 Feature extraction
Using the Librosa audio analysis library,
[15]
clinicians were able to extract Mel Frequency Cepstral Coefficients
(MFCCs) from their heart sound recordings. MFCCs are widely accepted for their ability to represent both
temporal and spectral aspects of the phonocardiogram signal. They are therefore commonly used in biomedical
audio analysis and in heart sound classification systems.
[7]
In addition to MFCCs, clinicians extracted additional
complementary acoustic features that included spectral centroid, spectral bandwidth, spectral roll off, zero-
crossing rate (ZCR), root mean square (RMS) amplitude, and chroma features. A total of 77 features were
extracted per recording for subsequent machine learning analysis, as shown in Table 1.
Table 1: Feature extraction summary used for heart sound classification.
Feature Type
Number of
Features
Purpose
MFCC Mean
20
Captures tonal characteristics of heart
sounds
MFCC Standard
Deviation
20
Represents variability in heart sound
patterns
MFCC Delta Mean
20
Captures temporal variations in cardiac
signals
Spectral Centroid
1
Represents center frequency distribution
Spectral Bandwidth
1
Measures frequency spread
Spectral Rolloff
1
Indicates high-frequency energy distribution
Zero Crossing Rate
1
Measures signal irregularity
RMS Energy
1
Represents signal amplitude and energy
Chroma Features
12
Captures harmonic and tonal information
Total Features
77
Combined feature vector used for
classification
2.2.5.4 Random forest classification
Random Forest classifier was chosen to classify heart sounds because it is adaptable, strong, simple to apply,
and works well for processing high-dimensional biomedical features.
[16]
Many current studies have shown how
machine learning and deep learning methods can be used for automatic murmur detection and intelligent
analysis of heart sounds.
[3,8]
Our model utilized 300 decision tree estimators with balanced weightings for all the
associated classes to increase the sensitivity for detecting murmurs. Additionally, recent research in the field of
evaluating biomedical signals has researched alternative methods for classifying heart sounds using intelligent
phonocardiograms.
[9]
Model validation involved the use of fivefold stratified cross-validation to ensure that performance was
evaluated on the basis of different partitions of data. The prediction threshold was decreased from its default
value of 0.50 to a lower value of 0.30 to increase the sensitivity of murmur detection. During cross-validation,
the proposed model achieved an overall classification accuracy of 82%, with a corresponding F1 score of 95.9%.
Owing to the nature of our model, it could be integrated into existing real-time monitoring systems and thus
provide for both the live extraction of heart sound data and the automatic detection and intelligent analysis of
such sounds.
2.2.6 Streamlit-based monitoring dashboard
To facilitate real-time analysis of heart sounds and efficiently manage monitoring data for patients, an
interactive dashboard has been developed using the Streamlit framework, as shown in Figs. 4 and 5. This
dashboard includes a number of features, including but is not limited to, live scrolling waveform visualization,
trends of beats per minute (BPM), confidence scores for machine learning predictions, and organized display of
patient information. Through continuous and uninterrupted acquisition and display of heart sounds, this system
offers the opportunity for healthcare providers to monitor their patients in real time. Other intelligent
monitoring systems, such as IoMT-based healthcare applications, that offer real-time visualization of biomedical
signals have been widely investigated for the purpose of physiological signal monitoring.
[2,13]
The waveform visualization component was developed using Matplotlib and functions at a refresh rate of
approximately 6 frames per second to provide smooth rendering of real-time signals. The BPM and machine
learning prediction outputs were also updated continuously on dedicated metric panels to provide real-time
feedback to the user. Prior to acquiring and analyzing the signals, patient name, age, identification (ID), and
clinical notes can be entered by the user into the integrated patient information forms and displayed on the
dashboard.
Additionally, the system allows for the continuous storage of executive records and historical monitoring data
through an SQLite database. The functions of this component allow for long-term storage of historical data to
provide for future clinical assessment. Usability and reporting were also enhanced by adding a means to export
data to a CSV file and formatting automated PDF reports on the monitoring dashboard. The reports generated
in Fig. 6 include waveform plots, BPM trends, heart sound classification results, confidence scores, and patient
identification information. Integrating intelligent monitoring, data management, and reporting capabilities
increases the potential use of the proposed system for biomedical research, educational purposes, and the smart
healthcare monitoring industry.
[2]
2.3 Implementation
2.3.1 Real-time serial communication
The real-time communication between the Arduino UNO and the PC was via USB serial communication at 9600
Bps. The PySerial Library was used to receive continuous heart sound data sent from the microcontroller, and a
background thread was created to provide nonblocking serial communication and uninterrupted real-time
processing of the heart sound signals.
2.3.2 Live-signal visualization
Matplotlib integrated into the Streamlit Framework was used to visualize the acquired and processed heart
sound signals in real time. The waveform plots were continuously updated with such updated waveforms; thus,
live monitoring of cardiac activity could be achieved. In addition, the visualization module also provided the
current BPM trends and prediction confidence values for interactive cardiovascular analysis.
2.3.3 Data storage and reporting
The patient records and monitoring output from the SQLite database were integrated into a Python application
that stored the information in an SQLite database. The data were exportable to the CSV format (using the Pandas
library) for further analysis and documentation. Automated PDF reports were generated via Matplotlib
PdfPages, which included waveform plots, BPM trend analysis, patient information and machine learning
predictions.
Fig. 4: Streamlit dashboard showing normal heart sound analysis and BPM monitoring.
Fig. 5: Streamlit dashboard showing abnormal heart sound classification with murmur prediction.
3. Results and analysis
3.1 Heart sound acquisition results
The validation experiment for the digital stethoscope system was completed by using the Arduino IDE (serial
plotter) to confirm the acquisition of sounds from the heart (heart sounds), as well as whether the system
created waveforms of the heart sounds. After the sound waves were acquired using a modified KY-037 sensor
module along with an external electret microphone (from the same acoustical source) and a stethoscope chest
piece, the heart sound periodic (cyclical) signals captured from the subject were transmitted using USB serial
communication and displayed in real time on the Arduino serial plotter interface.
The resultant raw waveform as output from the Arduino IDE incorporated recognizably cyclical (periodic)
cardiac cycle representations of heartbeat (heart sound) activity. The heart sound cyclic report in the waveform
output indicated proper functioning for both the proposed hardware equipment and the corresponding
acquisition circuit to enable later implementation of complex signal processing techniques and AI/machine
learning algorithms. Although noise and some interference were evident in the raw signals, the peaks of the
heart beats could be easily differentiated to allow subsequent processing and determination of heart rate in as
many heart beats per minute (BPM).
After validation in the Arduino IDE, the acquired data were imported into a Streamlit dashboard developed
in Python for further processing. The heart sound signals were pre-processed, visualized, and analyzed to
provide real-time BPM estimation and machine learning-based cardiac sound classification. The real-time raw
waveform obtained from the Arduino serial plotter is presented in Fig. 7.
3.2 Signal processing results
The multistage signal processing pipeline that was applied enabled the heart sound signals that were collected
to have improved signal quality. The median filter was successful at eliminating sudden impulse-type noise and
transient spikes caused by both motion of the microphone and other outside sources. The Butterworth
bandpass filter was successful at eliminating baseline drifts in the low-frequency domain while also removing
electrical noise from the high-frequency range, producing a more continuous, stable waveform pattern.
By using the moving average filter for additional smoothing of the waveform, waveform continuity was
enhanced while small fluctuations in the signal were reduced. A final processing step for the signal was to clamp
outliers, thus preventing large amplitude spikes that would have resulted in inaccurate strengths for future BPM
calculations. The comparison between the processed heart sounds in real time and the raw heart sounds
revealed clear differences in clarity and periodicity, indicating that the degree of filtering that was performed
on the waveform was an excellent visual indicator.
The proposed steps described in the overall preprocessing pipeline were successful at improving the overall
stability of the signal, resulting in greater confidence in the future estimation of BPM from the processed signal
and greater reliability when machine learning-based classifications of heart sounds obtained from the same
signal were performed. In addition, the filtering steps reduce the number of false peak detections because of
motion artifacts or interference from nearby electronic sources (i.e., computers and machines).
Fig. 6: Generated PDF-based heart sound analysis report produced by the proposed digital stethoscope system.
3.3 BPM estimation, machine learning performance and comparative analysis
According to an experimental evaluation, the intelligent digital stethoscope demonstrated stable real-time BPM
estimation and effective machine learning-based cardiac sound classification. The adaptive peak detection
algorithm was successful in detecting cardiac peaks from the processed phonocardiogram signals and
generating BPM values in a physiological range of 40180 beats per minute (BPM) on a continuous basis. The
use of preprocessing stages greatly reduced false peak detection and improved the stability of the waveform
during real-time monitoring. Finally, the developed system was able to monitor heart sounds continuously and
display BPM values in real time via the Streamlit dashboard.
The random forest (RF) ML model trained using the PhysioNet CirCor DigiScope dataset identified normal heart
sounds and murmur heart sounds. We successfully analyzed and classified cardiac sounds on the basis of the
MFCC and the spectral, chroma, and energy features that we extracted. The overall accuracy of the classifier was
measured at 82%, and the cross-validation F1 score was 95.9%. Normal heart sounds were classified with high
recall; however, there was lower sensitivity for murmur classification because of discrepancies between how
clinical grade systems record audio vs our low-cost electret microphone recording in the prototype setup. Real-
time testing also affected performance when murmurs were distinguished because of environmental noise,
sensitivity limitations of the microphone used, and inconsistencies in the chest piece placement path.
The KY-037-based digital stethoscope was compared with existing low-cost digital stethoscopes using the
MAX9814 chip and with commercially available digital stethoscopes across a number of parameters, such as
system cost, hardware architecture, monitoring capability, portability, and where they can be used, as shown in
Table 2. The proposed system uses a cost-effective hardware architecture while providing real-time monitoring
of cardiac signals, a software visualization program, and intelligent machine learning-based analysis. Compared
with traditional low-cost digital stethoscopes, the proposed architecture can provide features such as multistage
filtering of the signals, adaptive BPM estimation, storing of data in an SQLite database, exporting of the data in
its CSV format, and automated PDF report generation via the Streamlit monitoring dashboard.
Fig. 7: Real-time raw heart sound waveform visualized using the Arduino IDE serial plotter.
Table 2: Comparative analysis of the proposed digital stethoscope system with existing implementations.
Parameter
Proposed System (KY-037 +
Arduino)
MAX9814 Based
System
Commercial Digital
Stethoscope
Approximate Cost
1000
1500
15,000+
Microcontroller
Arduino UNO
Arduino/MCU
Proprietary
Hardware
Sound Acquisition
Module
Modified KY-037 with External
Electret Microphone
MAX9814 (AGC
Enabled)
Integrated Clinical
Sensor
Real-Time
Monitoring
Yes (Serial Plotter + Streamlit
Dashboard)
Yes
Yes
Wireless
Connectivity
No
Optional
Yes
Machine Learning
Support
Yes
Limited
Proprietary
Data Storage and
Reporting
SQLite, CSV Export, PDF
Reports
Minimal
Integrated
Intended Use
Educational and Research
Application
Research
Clinical Application
3.4 Discussion
The digital stethoscope system created successfully shows how a high-quality open-source platform could be
created to allow for real-time acquisition of heart sound signals, to provide visual output of the acquired
waveforms, to allow for preprocessing of the heart sounds, and to utilize intelligent analytical techniques to
analyze the acquired heartbeat signal. All the experiments using the digital stethoscope produced consistent
results when the modified KY-037 sensor module was used in conjunction with the Arduino UNO as the
microcontroller to acquire heart sounds, with consistent waveform outputs and reliable BPM estimates
generated from the digital stethoscope during the experimental tests. Additionally, when all the preprocessing
techniques (e.g., filtering, denoising and segmentation) were applied to the heartbeat signals before they were
analyzed and classified according to the presence of heart murmurs, the quality of the waveform significantly
improved, and an overall reduction in the amount of noise present in the output of the analysis occurred.
[6,7]
The
development of machine learning-based murmur classification techniques has also substantially enhanced the
ability of digital stethoscope systems to provide preliminary automated classifications of heart sounds based on
a phonocardiogram, as described in some of the current literature related to the intelligent analysis of heart
sounds.
[8,9]
Other features of the digital stethoscope system, including an integrated dashboard for displaying
patient data and monitoring capabilities, integrated patient record management capabilities, CSV export
functions, and automated generation of PDF reports, all contribute to making the digital stethoscope a useful
biomedical education tool, a prototyping system for further development, and an intelligent medical monitoring
system.
[11]
Despite showing hope through experimental outcomes, limitations of the developed system exist when the
system is tested. The combination of a low-cost electret microphone and a KY-037 sound sensor resulted in
lower sensitivity to clinical-grade digital stethoscopes when the detection of low-intensity heart murmurs and
abnormal heart sounds was detected.
[12,14]
Additionally, the acquired heart sounds did not show quality during
actual monitoring because of environmental noise, improper placement of the chest piece, and improper
handling of the system by the user. Prior to clinical validation, the intended use of this system will be to educate,
research, and complete some preliminary screening rather than direct medical diagnosis. Future improvements
may involve advanced denoising algorithms, heart sound classification based on deep learning techniques,
healthcare devices that are/or may become wearable, and IoT-enabled telemedicine systems, which will allow
for remote patient monitoring and other forms of intelligent healthcare.
[2,8]
3.5 Future scope
4. Conclusion
CRediT Author Contribution Statement
Nafisa Mapari: Formal analysis, Methodology, Project administration. Alefiya Managori: Conceptualization,
Formal analysis, Data curation, Software, Visualization, Writing-Original draft preparation, Writing- Review &
editing. Hajra Naik: Data curation, Methodology, Investigation. Tasmiya Kandhal: Validation, Writing- Review
and editing. Monika Chaurasiya: Methodology, Resources, Writing-Review & editing. All authors have read
and agreed to the published version of the manuscript.
Funding Declaration
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-
profit sectors.
Data Availability Statement
The dataset used in this study is publicly available through the PhysioNet CirCor DigiScope Phonocardiogram
Dataset repository, (Ref. 11). The developed source code, preprocessing methods, and implementation details
are available from the corresponding author upon reasonable request.
Conflict of Interest
There is no conflict of interest.
Artificial Intelligence (AI) Use Disclosure
The authors declare that artificial intelligence (AI)-assisted tools were used only for language refinement,
grammar improvement, and manuscript structuring purposes during the preparation of this work. All technical
content, experimental implementation, results, and interpretations were independently developed and verified
by the authors.
Supporting Information
No Applicable.
References
[1]
J. McGee, Evidence-Based Physical Diagnosis, 4th edition. Philadelphia, PA, USA: Elsevier, 2018.
[2]
S. Abdulmalek, A. Nasir, W. A. Jabbar, M. A. M. Almuhaya, A. K. Bairagi, M. A.M. Khan, S. H. Kee, IoT-based
healthcare-monitoring system towards improving quality of life: A review, Healthcare, 2022, 10, 1993,
doi: 10.3390/healthcare10101993.
[3]
A. Chorba, J. Shapiro, A. LeCun, Y. Bengio, Deep learning algorithm for automated cardiac murmur
detection via a digital stethoscope platform, Journal of the American Heart Association, 2021, 10,
e019905. doi: 10.1161/JAHA.120.019905.
[4]
H. Liang, S. Lukkarinen, I. Hartimo, Heart sound segmentation algorithm based on heart sound
envelogram, Computers in Cardiology, 1997, 24, 105108, doi: 10.1109/CIC.1997.647841.
[5]
D. Castaneda, A. Esparza, M. Ghamari, C. Soltanpur, H. Nazeran, A review on wearable
photoplethysmography sensors and their potential future applications in health care, International
Journal of Biosensors & Bioelectronics, 2018, 4, 195202, doi: 10.15406/ijbsbe.2018.04.00125.
[6]
D. Springer, L. Tarassenko, G. Clifford, Logistic regression-HSMM-based heart sound segmentation, IEEE
Transactions on Biomedical Engineering, 2016, 63, 822832, doi: 10.1109/TBME.2015.2475278.
[7]
C. Liu, D. Springer, Q. Li, B. Moody, R. A. Juan, F. J. Chorro, F. Castells, J. M. Roig, I. Silva, A. E. Johnson, Z.
Syed, An Open access database for the evaluation of heart sound algorithms, Physiological Measurement,
2016, 37, 21812213, doi: 10.1088/1361-6579/37/12/2181.
[8]
K. N. Khan, F. A. Khan, A. Abid, T. Olmez, Z. Dokur, A. Khandakar, M. E. Chowdhury, M. S. Khan, Deep
Learning based classification of unsegmented phonocardiogram spectrograms leveraging transfer
learning, Physiological measurement, 2021, 42, 095003, doi: 10.1088/1361-6579/ac1d59.
[9]
M. Kalimuthu, C. Hemanth, A comparative analysis of machine learning and deep learning approaches for
phonocardiogram classification using dataset integration, IEEE Access, 2025, 13, 170619170635, doi:
10.1109/ACCESS.2025.3612392.
[10]
A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Moody, C. K. Peng, H. E.
Stanley, PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex
physiologic signals, Circulation, 2000, 101, e215e220, doi: 10.1161/01.CIR.101.23.e215.
[11]
J. Oliveira, F. Ferreira, C. Pereira, A. Silva, R. Costa, M. Fonseca, A. Santos, H. Gamboa, The CirCor DigiScope
Phonocardiogram Dataset, In IEEE Conference, PhysioNet, 2022, doi: 10.13026/tshs-mw03.
[12]
M. A Reyna, Y. Kiarashi, A. Elola, J. Oliveira, F. Renna, A. Gu, E. A. Perez Alday, N. Sadr, A. Sharma, J. Kpodonu,
S. Mattos, Heart murmur detection from phonocardiogram recordings: The George B. Moody PhysioNet
Challenge 2022, Computing in Cardiology, 2022, 49, 14, doi: 10.1371/journal.pdig.0000324.
[13]
M. Kalimuthu, C. Hemanth, Preliminary Study on real-time phonocardiogram signal acquisition and
analysis using machine learning and IoMT for digital stethoscope, IEEE Access, 2025, 13, 6868268709.
doi: 10.1109/ACCESS.2025.3560763.
[14]
Arduino, Arduino UNO Rev3 Technical Specifications, Arduino Documentation, 2023. Accessed: May 20,
2026,
available at: https://docs.arduino.cc/hardware/uno-rev3/
[15]
B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, O. Nieto, Librosa: Audio and music signal
analysis in python, SciPy, 2015, 7, 1825. doi: 10.25080/Majora-7b98e3ed-003.
[16]
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R.
Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, É. Duchesnay, Scikit-
learn: Machine learning in python, Journal of Machine Learning Research, 2011, 12, 28252830.
Publisher Note: The views, statements, and data in all publications solely belong to the authors and
contributors. GR Scholastic is not responsible for any injury resulting from the ideas, methods, or products
mentioned. GR Scholastic remains neutral regarding jurisdictional claims in published maps and institutional
affiliations.
Open Access
This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which
permits the non-commercial use, sharing, adaptation, distribution and reproduction in any medium or format,
as long as appropriate credit to the original author(s) and the source is given by providing a link to the Creative
Commons License and changes need to be indicated if there are any. The images or other third-party material
in this article are included in the article's Creative Commons License, unless indicated otherwise in a credit line
to the material. If material is not included in the article's Creative Commons License and your intended use is
not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly
from the copyright holder. To view a copy of this License, visit: https://creativecommons.org/licenses/by-
nc/4.0/
© The Author(s) 2026
