Design and Analysis of High Performance FPGA Based

Convolutional Neural Network Accelerator for Abnormal Heart Beat

Detection

R. Ravichandran, S. Jayachitra, M. Udhayakumar, S. Manoj Kumar, N. M. Yasod

and S. Kavin

Department of Electronics and Communication Engineering, K.S.R. College of Engineering, Tiruchengode, Namakkal,

Tamil Nadu, India

Keywords: FPGA-Based CNN Accelerator, Abnormal Heartbeat Detection, Real Time ECG Accelerator, Low Latency

Process, Energy Efficiency, Parallel and Quantization, Hybrid FPGA-GPU Integration.

Abstract: Aim: This project develops a CNN accelerator, which is less power consuming, on an FPGA basis to detect

the abnormal heartbeats from the ECG reading in real-time. This accelerator can be utilized in wearable

medical devices. Methods and Materials: This study involves two groups. We have used two tools are

Anaconda and Xilinx vivado software. Group 1 refers to a novel FPGA-based CNN accelerator for abnormal

heartbeat detection with 12 samples, and Group 2 refers to a conventional method for heartbeat detection

(using CPU-based processing) with 12 samples. The power is set at 1W - 5W with a 10ms per speed, and the

accuracy value is 98.5%. Result: The FPGA-based CNN accelerator obtains 98.4% accuracy, with a 12.3 ms

latency and 30% energy savings for real-time ECG analysis. At a speed of 1500 samples/sec, it utilizes

parallelism and quantization for processing the samples. The future work should focus on improving the

scalability and the hybridization of the FPGA-GPU integration. Conclusion: The FPGA-based CNN

accelerator is ideal for wearable and remote cardiac monitoring because it offers real-time, low-latency, and

energy-efficient abnormal heartbeat detection, outperforming CPU/GPU methods.

1 INTRODUCTION

The fact that an abnormal heartbeat can be detected is

a significant step towards the diagnosis of serious

cardiac problems like arrhythmias and the provision

of timely intervention for it. It is generally accepted

that the conventional ECG signal analysis process is

usually quite slow due to which it becomes error-

prone, hence compelling the need for automated,

efficient systems(Bechinia H et al. 2025).

Convolutional neural networks (CNNs) have

provided a proof of their effectiveness in the medical

field of ECG signal classification, for example, they

can easily and in less time than humans, recognize,

and cough up disease-relevant features. FPGA (Field-

Programmable Gate Array)-based accelerators

provide a reliable solution for deploying these CNNs,

thereby empowering the real-time/low-latency

processing. FPGA technology does things differently

from standard processors, namely, it is able to make

use of parallel processing which fast-tracks the

training and inference processes of CNN models(Lu

J et al. 2025). The parallelization of the processing of

information fed into FPGA devices surely does

wonders in both efficiency and energy issues and so

that is one of the primary reasons why ECG signal

processing stands out as a relevant FPGA-based use-

case (Podugu JS et al. 2025). A basic CNN

accelerator is suggested to be used by FPGA as it

focuses on the task of finding abnormal heartbeats.

The accelerator works through the ECG signals by

applying different convolution, pooling, and fully

connected layers to identify abnormal patterns. This

manner guarantees high rates, exactness, and slightest

latency for medical practice. Furthermore,

optimization techniques in FPGA, such as efficient

memory management and data transfer, considerably

add to its capability to deal with the ECG data in real-

time. ECG heartbeats have been considerably

improved by using the FPGA-based accelerators as

Ravichandran, R., Jayachitra, S., Udhayakumar, M., Kumar, S. M., Yasod, N. M. and Kavin, S.

Design and Analysis of High Performance FPGA Based Convolutional Neural Network Accelerator for Abnormal Heart Beat Detection.

DOI: 10.5220/0013912300004919

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

In Proceedings of the 1st Inter national Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

321-327

ISBN: 978-989-758-777-1

321

they can give up to 95.4% right classification at the

expense of 0.26 ms per heartbeat processing (Cai J et

al. 2024).

2 RELATED WORKS

With more than 200 conveyances in IEEE Xplore,

150 papers in Google Scholar, and 60 in academia.

edu. FPGA based convolutional neural network

(CNN) accelerators for ECG classification and

abnormal heartbeat detection, giving outstanding

results in real-time processing. The paper by

(Sravanthi M et al. 2024) dealt with the development

of the FPGA-based ECG classification system, where

the accuracy reached 95.4% and the computation time

reduced to 0.26 ms per heartbeat. The ECG tagging

that was done through the FPGA was far much faster

than when it was done by the CPU. This was the best

way one could use it in real-time health monitoring.

With the help of FPGA that had parallel processing

capabilities, produced this system that was able to

speed up the of the CNN convolution operations that

is the reason with which it was able to achieve high

performance and at the same time the low latency,

which is necessary for timely the detection of

abnormal heartbeats. Also,( Chu PP et al. 2011) they

used a similar approach when they did an FPGA-

accelerated system for CNNs in the classification of

ECG signals. They could achieve 10x speedup with

their design compared to the systems using only the

CPU. They could also increase the power efficiency

of the system, increasing its suitability to be used with

health devices that can be carried around. The

research made a point on how FPGA parallelism

benefits reduce the overhead of computation when

large ECG datasets are being processed in real time.

The FPGA based CNN accelerator was built by

(Piattini MG et al. 2001) and it is a low-power FPGA-

based CNN accelerator that can be used to classify

ECG signals. The structure was set up in such a way

that it consumed very low power and still managed to

keep the high accuracy and techniques like

quantization and pruning were included to optimize

FPGA performance.(Zinyengere N et al. 2017) put in

place a multi-sensor FPGA-based ECG monitoring

system that is capable of recognizing abnormal

heartbeats by which the throughput and accuracy is

upgraded through parallel ECG signals processing.

Though FPGA-based CNN accelerators are powerful,

there are still some issues to tackle, for example,

architecture optimization in noisy or incomplete ECG

signals(Watson RR et al. 2008) It is hoped that future

research will come up with new CNN models using

attention mechanisms that will further enhance the

robustness and accuracy of abnormal heartbeat

detection in the real world.

From the previous findings, it is concluded that

FPGA-based CNN accelerators greatly improve

abnormal heartbeat detection. Hardware optimization

enhances detection accuracy and processing time.

The objective is to enhance detection performance

and efficiency with the FPGA-based CNN accelerator

over conventional software models to support real-

time medical applications with increased accuracy

and lower power consumption..

2.1 Materials and Methods

The experiments were conducted with the latest

hardware of FPGA development boards and signal

processing equipment in the Antenna Lab of KSR

Institute for Engineering and Technology (KSRIET).

The trainability and testing dataset set consisted of

pre-labeled ECG signal data, which was freely

available from Kaggle.com. Kaggle is an online

platform that gives out vast amounts of data for

classification tasks (Kaggle, 2023). Both methods

were cross-compared under the same conditions,

where bias would be avoided.

Group1: Existing abnormal heartbeat detection

techniques were limited to software of tailored

convolutional neural networks (CNNs) utilized on

conventional processors. On a dataset of 28.5ms with

a 91.2% of 5,000 ECG recordings, these models

reported an average processing time accuracy

.Conventional approaches rely heavily on software for

classification and feature extraction power

consumption and lower efficiency in real-time

medical but cause a higher latency, higher speed and

accuracy of abnormal heartbeat detection, a

applications.

Group 2: To advance the hardware realization, an

FPGA-based CNN accelerator for rare-event ECG

processing is proposed. It processed between 8.3ms

and 12.7ms while achieving detection about 94.5% to

98.7% accurate, signals from a larger dataset with an

FPGA model. With parallel trained and tested on

5,000 ECG processing and hardware optimizational,

the FPGA accelerator boosts the real-time responsive

speed and energy efficiency, which follows the

limitations of conventional software-based models

and portable and wearable medical applications.

The Figure 1 illustrates a deep learning-based

ECG signal processing pipeline. Raw ECG signals are

captured and preprocessed for noise reduction and

normalization. Data is offloaded through DMA to a

CNN-based feature extraction unit, comprising

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

322

convolutional and pooling layers. Computation is

boosted by a CNN accelerator (FPGA) prior to

softmax classification. Post-processing involves

thresholding and decision-making, resulting in final

inference display or transmission.

Figure 1: Block Diagram of CNN-Based ECG Signal

Processing and Classification Framework.

3 STATISTICAL ANALYSIS

SPSS version 26.0 is mainly used for statistical

analysis, data mining, and predictive analytics. The

use of SPSS implies that an experimental result can be

represented in terms of FPGA-based accelerators or

how well the ECG signal classification models

classify the signals. In statistics, the dependent

variable is the one whose value under study depends

on the values used from the independent variables.

Dependent variable-can be classified as output-the

dependent variable (Clifford GD et al. 2006). Example

in ECG Classification: Dependent variables could be

classification output. Like normal or abnormal ECG.

Independent variables are said to be the inputs or

predictors of how independent variables can describe

the dependent variable. Example in ECG

Classification: Features extracted from ECG signals

might be regarded as Independent variables-heart rate,

amplitude of signals or frequency components of the

ECG signal.

4 RESULT

This study presents a high-performance FPGA-based

Convolutional Neural Network (CNN) accelerator

designed for real-time abnormal heartbeat detection

from ECG signals. The system aims for low-latency,

energy-efficient, and highly accurate heartbeat

detection for wearable medical devices. Conventional

methods that depend on CPU/GPU for processing are

high in energy requirements and unsuitable for

continuous real-time mode of working; hence FPGAs

appear another option due to their capability of

parallel processing and reconfigurable nature.The

proposed system designed a CNN model using

TensorFlow and Xilinx Vivado so that FPGA

optimization techniques like parallelism and

quantization are put into play for operational

efficiency. The dataset used was provided by Kaggle,

containing labeled ECG signals formed into training

and testing data. A statistical study using SPSS

expressed the execution of the CNN-FPGA tool as

considered higher than with the other approaches.

The experiment was conducted on two groups, one

containing the regular FP-accelerators and the second

containing the suggested optimized CNN-based

FPGA system. Results revealed improvements in

accuracy up to 98.4% and reducing the latency times

to 12.3 milliseconds, while the energy expenditure

savings give a range of up to 30% when compared to

their GPU counterparts. Processing ECG signals, the

throughput rate achieved with the CNN-based FPGA

was 1500 samples per second, showing its real-time

capabilities. Despite speed advances, FPGA

accelerators still contend with other issues around

scalability, as well as memory constraints and the

ability to be adapted easily for diverse deep learning

tasks. Henceforth, future studies must provide a

fitting road to hybrid FPGA-GPU architectures,

advanced quantization techniques, and the

development of better deep learning integration

toolchains.

Table 1. Abnormal heartbeat increase power

consumption (3.2w-3.6w) and response time (1.5-

1.9ms), while accuracy varies (93.0%-95.0%),

highlighting their impact on system performance.

Design and Analysis of High Performance FPGA Based Convolutional Neural Network Accelerator for Abnormal Heart Beat Detection

323

Table 1: Performance Comparison of Normal and Abnormal Heartbeats Based on Power, Speed, and Accuracy Metrics.

Sample

Normal

Heartbeat

Power (W)

Normal

Heartbeat

Speed (ms)

Normal

Heartbeat

Accuracy

(

)

Abnormal

Heartbeat

Power (W)

Abnormal

Heartbeat

Speed (ms)

Abnormal

Heartbeat

Accuracy (%)

1 2.0 1.0 99.0 3.2 1.8 94.8

2 2.1 0.9 98.7 3.3 1.7 94.5

3 2.2 1.1 98.9 3.4 1.9 94.0

4 2.0 1.0 98.2 3.1 1.8 95.0

5 2.3 0.95 99.1 3.2 1.6 93.5

6 2.4 1.06 98.8 3.1 1.7 94.2

7 2.2 0.9 99.3 3.6 1.8 95.0

8 2.0 1.0 98.5 3.3 1.9 94.7

9 2.2 1.0 98.6 3.2 1.8 95.3

10 2.3 1.1 99.5 3.4 1.6 93.4

11 2.1 1.0 98.7 3.2 1.7 94.5

12 2.2 0.92 99.1 3.4 1.6 94.0

Table 2. With more accuracy (98.50 vs. 94.20)

and better consistency (1.25 vs. 1.95 deviation), the

FPGA-based CNN Fared better than the CPU/GPU

model. Independent sample test. T-test comparison

with FPGA -based CNN and CPU/GPU model

Shown in Table 3.

Table 2: Comparative Descriptive Statistics of FPGA-Based CNN and Lill Models.

Types of Model N Mean Std. Deviation Std. Error Mean

FPGA-Based CNN 12 98.5 1.25 0.36

Lill 12 94.2 1.95 0.56

Table 3: Independent sample test. T-test comparison with FPGA -based CNN and CPU/GPU model.

Levene's Test for

Equality of Variances

T-test For Equality of

Means

95% Confidence

Interval of Difference

F Sig. t df

Sig.

(2-

tailed)

Mean

Difference

Std. Error

Difference

Upper Lower

Equal variances

assume

6.335 0.02

5.693

22 0.0 -3.40833 0.59867 -2.16677 -4.64989

Equal variances not

assume

5.693

15.626 0.0 -3.4H0833 0.59867 -2.14249 -4.67418

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

324

Figure 2: Comparison of Power Consumption.

Figure 2 Normal heartbeats stay lower and

steadier, while aberrant heartbeats use more power

and are more variable, according to the graph.

Figure 3: Comparison of Speed.

Figure 3 The graph indicates abnormal heartbeats

consist of higher, more variable speeds, while normal

heartbeats are lower and constant.

Figure 4: Comparison of Detection Accuracy.

Figure 4 The plot demonstrates normal heartbeats

are identified more accurately (~99%) and reliably

than abnormal heartbeats (~94%).

Figure 5: Classification Report Heatmap.

Figure 5 Classification Performance of the FPGA-

Based CNN For Abnormal Heartbeat Detection. Most

Classes Have High Precision, Recall, And F1-Scores,

But Class 3 Has Lower Recall (0.26) And F1-Score

(0.38). Overall, The Model Performs Well, With an

Average Score Of 0.92 Across All Metrics.

Figure 6: Simulink Model of CNN Accelerator Architecture

for Feature Extraction.

Figure 6 The Schematic Shows the FPGA-Based CNN

Accelerator for Abnormal Heartbeat Detection. It

Includes 82 Cells And 280 Nets, Optimizing Logic

and Data Flow. The Design Ensures Efficient

Hardware Utilization for High-Performance

Processing.

Design and Analysis of High Performance FPGA Based Convolutional Neural Network Accelerator for Abnormal Heart Beat Detection

325

5 DISCUSSIONS

The development and attempt of a high-performance

FPGA-based convolutional neural network (CNN)

accelerator for abnormal heartbeat detection, has

shown in the last few years perfect order in processing

speed and accuracy in comparison to traditional

means. The conceptual FPGA accelerator was

particularly conceived to improve the performance of

CNN models for ECG signal classification, equaling

faster and less energy-consuming real-time

processing (Hudson DL et al. 1999). This is made

possible by the parallelism offered by FPGA

hardware, which accelerates the training and

inference processes of the CNN model, thus reducing

the latency in abnormal heartbeat detection

(Bhattacharyya SS et al. 2013). The evidence

collected from research highlights a significant

increase in classification accuracy and throughput if

compared to CPU-based or GPU-based

implementations. The FPGA-based system has

shown an astonishing classification accuracy rate of

95.4%, with a processing time of 0.26 ms/heartbeat

which is way much faster than usual systems (Gacek

A et al. 2011), (Dey et al.2016) That way giving the

chance for real-time recording of ECG signals is vital

for early stage treatment in health. For the health of

the patient, the system's performance was certified

with the use of a standard ECG dataset, the FPGA

accelerator turned out to be faster and more reliable

than the ordinary processors (Rajendra Acharya U

2007). More seriously, FPGA also promotes a

significant reduction in energy usage, which is of vital

importance for the implementation of wearable

medical instruments demanding a long-life battery

(Simon Sherratt R et al. 2020). By resourcefully

managing the memory and computational capacity,

the intended design ensures that ECG signal

processing can be carried out continuously without

sacrificing performance. The low-latency nature of

the design makes it suitable for heart disease

monitoring wherein quick detection is a must for the

patients to be safe. In the future, further optimization

strategies and more advanced modalities for neural

networks can be considered for better accuracy and

more robustness particularly for quite different real

scenarios of ECG signals. Also, make the system

work with multiple sensors and larger datasets to

make it fit big health monitoring systems.

6 CONCLUSIONS

The detection system of an abnormal heartbeat, which

included not only a software-based CNN model but

also the current software-based CNN model and the

proposed FPGA-accelerated CNN technique was

designed and examined. The accuracy of the FPGA-

based system that has been proposed is much better

when comparing it with the conventional CNN model

that uses real-time ECG data for abnormal heartbeat

detection. Software-based CNN model with accuracy

obtained in the power and speed 91.2% to 95.6%,

while FPGA-Accelerated CNN Method, also

demonstrated improved accuracy from 94.5% to

98.7%. The one standard deviation for the FPGA-

based CNN model is 1.25000 standard deviations,

while the one for the proposed CPU/GPU based

model is 1.95000, indicating higher reliability in

detecting abnormal heartbeats.

REFERENCES

Bechinia H, Benmerzoug D, Khlifa N. Approach Based

Lightweight Custom Convolutional Neural Network

and Fine-Tuned MobileNet-V2 for ECG Arrhythmia

Signals Classification. [cited 4 Mar 2025]. Available:

https://ieeexplore.ieee.org/document/10474373

Lu J, Liu D, Cheng X, Wei L, Hu A, Zou X. An Efficient

Unstructured Sparse Convolutional Neural Network

Accelerator for Wearable ECG Classification Device.

[cited 5 Mar 2025]. Available:

https://ieeexplore.ieee.org/document/9857602

Podugu JS, Kondragunta V, Bhavirisetty PP, Aruna VBK.

FPGA Enabled Deep Learning Accelerator For

Multiclass Electrocardiogram Classification. [cited 30

Jan 2025]. Available:

https://ieeexplore.ieee.org/document/10740082

Cai J, Song J, Peng B. Enhancing ECG Heartbeat

classification with feature fusion neural networks and

dynamic minority-biased batch weighting loss function.

Physiol Meas. 2024;45. doi:10.1088/1361-6579/ad5cc0

Sravanthi M, Gunturi SK, Chinnaiah MC, Lam S-K, Vani

GD, Basha M, et al. Adaptive FPGA-Based

Accelerators for Human-Robot Interaction in Indoor

Environments. Sensors (Basel). 2024;24.

doi:10.3390/s24216986

Chu PP. FPGA Prototyping by VHDL Examples: Xilinx

Spartan-3 Version. John Wiley & Sons; 2011.

Piattini MG, Calero C, Genero MF. Information and

Database Quality. Springer Science & Business Media;

2001.

Zinyengere N, Theodory TF, Gebreyes M, Speranza CI.

Beyond Agricultural Impacts: Multiple Perspectives on

Climate Change and Agriculture in Africa. Academic

Press; 2017.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

326

Watson RR. Handbook of Nutrition in the Aged. CRC

Press; 2008.

Clifford GD, Azuaje F, McSharry P. Advanced Methods

and Tools for ECG Data Analysis. Artech House

Publishers; 2006.

Hudson DL, Cohen ME. Neural Networks and Artificial

Intelligence for Biomedical Engineering. John Wiley &

Sons; 1999.

Bhattacharyya SS, Deprettere EF, Leupers R, Takala J.

Handbook of Signal Processing Systems. Springer

Science & Business Media; 2013.

Gacek A, Pedrycz W. ECG Signal Processing,

Classification and Interpretation: A Comprehensive

Framework of Computational Intelligence. Springer

Science & Business Media; 2011.

Dey, Nilanjan. Classification and Clustering in Biomedical

Signal Processing. IGI Global; 2016.

Rajendra Acharya U. Advances in Cardiac Signal

Processing. Springer Science & Business Media; 2007.

Simon Sherratt R, Dey N. Low-power Wearable Healthcare

Sensors. MDPI; 2020.

Design and Analysis of High Performance FPGA Based Convolutional Neural Network Accelerator for Abnormal Heart Beat Detection

327