Detecting Atrial Fibrillation Using an Electrocardiogram Signal on a Low-Power Microcontroller

Abstract

In this work, we suggested using an enhanced Variable Step Dynamic Threshold Local Binary Pattern algorithm to build a straightforward and lightweight method for detecting atrial fibrillation. We can cut the input feature size to just 44 features without noticeably lowering classification accuracies by using feature selection based on statistical significance and association. In conjunction with a support vector machine classifier, the suggested approach can attain 99.14% sensitivity, 99.12% specificity, and 99.13% accuracy when tested on 15-second signal segments from the MIT-BIH Atrial Fibrillation Database. The sensitivity, specificity, and accuracy are 99.49%, 99.46%, and 99.47%, respectively, at a 60-second input signal length. When the input signal length is 60 seconds, the machine learning model size is as tiny as 132.86kB due to the lower input feature size. With an average current usage of only 27mA, the suggested approach can perform comparably to a PC-based system when used on an Arm Cortex M4-based STM32F413ZHT3 CPU with a 100MHz clock frequency. The embedded C software may finish one SVM inference in as little as 11.28 ms and fit into a flash memory as little as 114.46 kB. The findings in this work demonstrate that a low-power, resource-constrained microcontroller may perform highly accurate machine learning classification that can identify atrial fibrillation from ECG signals. Our findings will greatly simplify the process of creating a wearable smart medical electronic device that is high-quality, inexpensive, and low-power for identifying atrial fibrillation from an ECG signal.

Specification

Description:


Abstract

In this work, we suggested using an enhanced Variable Step Dynamic Threshold Local Binary Pattern algorithm to build a straightforward and lightweight method for detecting atrial fibrillation. We can cut the input feature size to just 44 features without noticeably lowering classification accuracies by using feature selection based on statistical significance and association. In conjunction with a support vector machine classifier, the suggested approach can attain 99.14% sensitivity, 99.12% specificity, and 99.13% accuracy when tested on 15-second signal segments from the MIT-BIH Atrial Fibrillation Database. The sensitivity, specificity, and accuracy are 99.49%, 99.46%, and 99.47%, respectively, at a 60-second input signal length. When the input signal length is 60 seconds, the machine learning model size is as tiny as 132.86kB due to the lower input feature size. With an average current usage of only 27mA, the suggested approach can perform comparably to a PC-based system when used on an Arm Cortex M4-based STM32F413ZHT3 CPU with a 100MHz clock frequency. The embedded C software may finish one SVM inference in as little as 11.28 ms and fit into a flash memory as little as 114.46 kB. The findings in this work demonstrate that a low-power, resource-constrained microcontroller may perform highly accurate machine learning classification that can identify atrial fibrillation from ECG signals. Our findings will greatly simplify the process of creating a wearable smart medical electronic device that is high-quality, inexpensive, and low-power for identifying atrial fibrillation from an ECG signal.

CLAIMS:

CLAIM 1:
The proposed method enhances the detection of Atrial Fibrillation (AF) from ECG signals by improving classification accuracy and simultaneously optimizing the model size, making it more efficient without compromising performance.


CLAIM 2:
Implemented on an Arm Cortex M4-based microcontroller, the proposed method can run an SVM inference within 11.28ms using as low as 27mA average current, enabling near real-time atrial fibrillation detection on a low-power and low-cost device.
CLAIM 3:
The proposed approach reduces the cost and enhances the quality of smart wearable ECG monitor devices, thereby alleviating the burden on medical professionals, patients, and their families by providing an accessible and efficient solution for monitoring Atrial Fibrillation.

1.Introduction

One of the most prevalent forms of arrhythmia worldwide is atrial fibrillation (AF). It has been linked to an increased chance of developing other illnesses, such as non-cardiovascular conditions like dementia or Alzheimer's disease or potentially fatal cardiovascular conditions like stroke.
AF burden has been found to be a significant determinant of the risk of stroke associated with AF. The total amount of time spent in AF over a certain time period is known as the "AF burden," and it is anticipated to fluctuate dynamically between a minimum and a maximum. Therefore, it is crucial to continuously monitor the occurrence of AF in order to diagnose and characterize AF in a patient.
Traditionally, a wearable Holter device that can record electrocardiogram (ECG) signals for an extended amount of time is used for long-term monitoring of AF. The lengthy ECG recording will next be examined by a qualified medical professional for irregular beats, including AF beats. The medical team will be under a great deal of strain due to this time-consuming and difficult procedure. Additionally, there aren't enough qualified medical professionals in some nations who can use ECG signals to identify cardiovascular disorders.

A. Previous Works on Detection of Atrial Fibrillation by Machine Learning
An automatic technique for identifying atrial fibrillation beats from ECG signals is required in order to address the issues with traditional approaches. This machine learning-based program can identify similar ECG beats in new signals by using patterns it has learnt from expertly annotated signals. Prior studies have demonstrated the importance and promise of analysing ECG signals using machine learning. A number of earlier studies have suggested deep learning-based techniques for identifying anomalies, such as atrial fibrillation and other arrhythmias, using ECG signals because of its capacity to automatically identify important aspects. By using these techniques, researchers can take use of developments in electronic engineering technology, such as the newest CPU or GPU with huge computing powers.

A two-dimensional deep convolutional neural network (2D CNN) is adapted for usage with one-dimensional (1 D) signals in the work of Mahmud et al. In order to reduce computation time and model complexity, the authors adopted a technique known as depth wise separable 2D convolution, in which the spatial and inter-channel convolution operations are carried out independently. The MIT-BIH Arrhythmia Dataset's 5-class classification of ECG data demonstrated a high accuracy of 99.28% for the method suggested in. However, the accuracy of the work by depends on the quality of the beat detection and segmentation method because it presupposes a beat-by-beat segmented input ECG data.

Convolutional neural networks (CNNs) and long short-term memory (LTSM) were used by Petmezas et al. to classify ECG beats into four classes (Normal, Atrial Fibrillation, Atrial Flutter, and AV Junctional Rhythms) on the MIT-BIH Atrial Fibrillation Database (AFDB) with 97.87% and 99.29% sensitivity, respectively. Similar to the work of , this system can operate with a very small input signal length (one beat), but its accuracy is highly dependent on the precision of the beat detection technique being employed.

B. Previous Works on Resource-Constrained Device-Based Automatic Atrial Fibrillation Detection

Small, battery-operated electronic biomedical devices that can be worn by a person and used continuously with little disruption to daily activities have been made possible by recent developments in semiconductor and microelectronic technologies. The quality of life for patients and their families will be greatly enhanced, and the workload for medical personnel would be lessened, if these gadgets are able to monitor and identify patients' ailments.

Relatively high classification accuracy can be attained by deep learning-based techniques, albeit at the expense of a significant memory and processing load. Implementing a deep learning-based detection system on a low-cost and low-power wearable microcontroller device is challenging due to these trade-offs. Cloud-based detection or edge devices with comparatively high processing capability, like an NVIDIA Jetson Nano or Raspberry Pi board, have been used in earlier studies to get over this issue.
A sliding big kernel-based deep learning technique is proposed by Tseng et al.for the processing and classification of ECG signals obtained by mobile devices. ECG data is transferred from a mobile device to a network server using the technique suggested by, which makes use of the widely available wireless high-speed internet connection. The server uses a deep learning technique to process the ECG signals after converting them into images.

In order to create a spectrogram from the input ECG signal, Farag et al. suggested employing a one-dimensional convolutional layer based on the short-time Fourier transform (STFT). Two-dimensional convolutional neural networks are then used to classify these one-dimensional features once they have been transformed into two-dimensional heat maps. With up to 12MB of random access memory and an edge device based on the Raspberry-Pi 3 model B+,was able to achieve 99.1% classification accuracy using the MIT-BIH Arrhythmia Database.
Two-dimensional convolutional neural networks (2D-CNN) were employed by Seitanidis et al.to categorize two-dimensional picture representations of ECG signals. The input ECG signals are divided into beats, and a 128 x 128 pixel grayscale image is created for each beat. classified ECG data from the MIT-BIH Arrhythmia Database with 95.3% accuracy using the NVIDIA Jetson Nano as the edge device.

Wearable medical devices are severely limited by the need to use hardware that is cloud-based or has a lot of processing power. There will be significant data transmission costs because all captured signals must be sent to the cloud for classification. Significant energy will also be needed for wireless data transfer, which will shorten the device's battery life or necessitate the use of larger batteries, so compromising comfort and usability. Utilizing a system with a lot of processing power may result in increased heat dissipation and energy consumption, which are both undesirable for a wearable medical device. Furthermore, there will be serious security and privacy issues when sending and storing a patient's raw biological signal to the cloud. Platforms with greater memory and processing capability are also usually more costly, which raises production costs and puts additional financial strain on patients. It is crucial to provide an ECG classification algorithm that can operate on an edge device with constrained processing power for the reasons listed above.
By employing manually chosen features-based machine learning, the detection system's computational load can be decreased. Significant aspects of the input signals are chosen beforehand by a human expert rather than being discovered by the machine learning model on its own. A machine learning classifier, like a support vector machine (SVM), uses the chosen input features as its input vector. The input dimension and search space for the machine learning model training are decreased when extracted features are used as classifier input. This strategy will lower the hardware's need for processing power and memory capacity.
Five characteristics made up of the statistical properties of the RR intervals were chosen by Chen et al. to be used with a basic artificial neural network (ANN) classifier that has only one hidden layer and three nodes. Although a personal computer (PC) was used for the machine learning model's training phase, the trained model was integrated into a low-power microcontroller for on-the-edge inference. The China Physiological Signal Challenge 2018 (15-second ECG signals) yielded 94.5% classification accuracy and a 2.22 ms processing time for this approach.

A technique based on dynamically assigned symbolic representations of an ECG signal's RR intervals as the input feature to traditional classifiers was proposed by Ganapathy et al. . Threshold values that are dynamically derived from the average value of the RR intervals of the entire signal are used to partition each RR interval into many classes according to its length. A co-occurrence matrix, which will serve as the input features for machine learning, is then created from the generated time series of codes. The approach suggested by demonstrated 94.0% and 99.8% classification accuracies, respectively, when tested on the Atrial Fibrillation Prediction Challenge Database (AFPDB) and the Atrial Fibrillation Termination Challenge Database (AFTDB).

Variable Step Dynamic Threshold Local Binary Pattern (VSDTLBP) is a novel feature extraction approach that we previously introduced in the work by Yazid and Mahrus for the detection of atrial fibrillation from ECG signals. Our suggested feature extraction strategy, which only requires basic arithmetic and logic operations, can produce a 99.11% sensitivity and a 99.29% specificity when applied to 60-second ECG data from the MIT-BIH Atrial Fibrillation Database in conjunction with an SVM classifier.

The VSDTLBP algorithm is a good option for developing a wearable AF detection device on a low-power microcontroller because of its comparatively simple calculation and noticeably higher classification accuracy compared to other published research on atrial fibrillation detection algorithms.


C. Contribution of This Work
The algorithm suggested in is improved upon in this study, and its effective application to low-power microcontrollers is discussed. It will be demonstrated that the approach put forth in this study can achieve excellent classification accuracy even when it is implemented as an embedded application on a microcontroller with low power and low cost. The suggested method is appropriate for an at-the-edge classification of atrial fibrillation and has a comparatively modest memory footprint and calculation time.

This work's contributions include:
1.Modifications to the VSDTLBP algorithm to increase its suitability for low-power microcontroller applications.
2.Effective application of the VSDTLBP algorithm in a microcontroller or device with limited resources.
3.The suggested method's classification accuracy is demonstrated to be on par with earlier research based on powerful computers like PCs or GPUs.
4.The suggested work can produce a greater classification accuracy with reduced production cost and power consumption when compared to earlier studies on edge device-based systems.
The organization of the paper is as follows. First, in section II, we will explain the dataset, data segmentation, hardware, and analysis methods used in this paper. In section III, we will elaborate on our proposed method. Section IV will show and discuss the experiment results, which will then be analysed and compared with results of previous works in section V. Finally, section VI will be the conclusion.
Section-II
Materials and Methods
A. ECG Signal Dataset

The quality of the created approach is verified and compared with earlier research using ECG signals from the MIT-BIH Arrhythmia Database (MITDB) and the MIT-BIH Atrial Fibrillation Database (AFDB). Prior research on AF identification has frequently used these publicly accessible ECG signal datasets. This will make it easier to compare our suggested approach objectively with those studies.

1) MIT-BIH Atrial Fibrillation Database
Twenty-five long-term ECG signal records from AF patients make up the MIT-BIH Atrial Fibrillation Database. Two ECG signals recorded at a sampling frequency of 250 samples/second and manually annotated information about each beat make up each 10-hour recording.
In the MIT-BIH Atrial Fibrillation Database, only the first of the two ECG channels is utilized. The ECG signals are divided into non-overlapping segments with lengths of 10, 15, 20, 30, 40, 50, and 60 seconds after. The beat annotations that are part of the database are used to label the ECG segments. Only wholly AF or fully non-AF segments are utilized; parts with both AF and non-AF beats are eliminated. Random under samplings are applied to the larger class in order to balance it with the other class because the input data classes are not balanced.
The number of segments produced for every input length from the MIT-BIH Atrial Fibrillation Database is displayed in Table 1. Following random under sampling to balance the classes, the figures under "Balanced" represent the number of signal segments utilized in the 10-fold stratified cross-validation test.


2) MIT-BIH Arrhythmia Database
Of the 47 human participants, 48 ECG data are included in the MIT-BIH Arrhythmia Database. Every recording has a sampling frequency of 360 samples per second and lasts roughly 0.5 hours. Each beat's personally annotated information is likewise included in this database.

There are two types of signals in the MIT-BIH Arrhythmia Database: 100 series signals and 200 series signals. Only 200 of these series signals have episodes of atrial fibrillation. Because of this, this work solely uses the 200 series signals. The signals undergo resampling at a sampling frequency of 250 samples per second. The signals from the MIT-BIH Atrial Fibrillation Database undergo the identical pre-processing, segmentation, and labelling procedures. The number of segments produced for every input length from the MIT-BIH Arrhythmia Database 200 series is displayed in Table 2.

B. Hardware Implementation
Within a Nucleo-144 development board manufactured by STMicroelectronics the suggested approach is implemented as an embedded application in the STM32F413ZH microcontroller. The Arm Cortex M4 core serves as the basis for the microcontroller, which has 320kB of random access memory (RAM) and 1.5MB of flash memory. The STM32F413ZH microcontroller can operate at a clock frequency of up to 100MHz and has a hardware floating point unit (FPU) as well. The manufacturer's integrated development environment (IDE), the STM32CubeIDE, is used to program the microcontroller. A number of other STMicroelectronics microcontrollers, such as the STM32F207ZG, STM32L4R5ZIP, and STM32H7A3ZIQ, will also be used to evaluate the suggested approach in order to compare its performance on various microcontroller types. The microcontroller specifications utilized in this project are displayed in Table 3.

In every experiment described in this paper, a hardware floating point unit (FPU), unless otherwise noted, is enabled if an MCU has one.

C. Analysis Method
Using the widely used criteria of sensitivity, specificity, and accuracy, the performance of the suggested algorithm is evaluated and contrasted with earlier research. Two experimental settings are used in this work:
1.Experiment Setting 1:
10-fold stratified cross-validation against the entire balanced portions of the MIT-BIH Arrhythmia Dataset and the MIT-BIH Atrial Fibrillation Dataset is carried out in order to evaluate the performance of the suggested method with earlier efforts. This report presents the average findings of five investigations.
2. Experiment environment 2:
Research pertaining to the integration of taught machine-learning models into a microcontroller is mostly conducted in this environment. After balancing, the available data is initially divided into train and test segments, with 20% going to testing and 80% going to training. The training segments in this experiment undergo 10-fold stratified cross-validation, and the model with the highest accuracy is retained. In order to compare performance, this model is then utilized to execute inference on the test segments as an embedded C program on the microcontroller and as a Python program running on a personal computer (PC). The test data is stored onto an SD card so that the microcontroller may access it for inference. This experiment is repeated five times and the average results are shown in this paper. Experiment Setting 2 is done only with the MIT-BIH Atrial Fibrillation Database signals.
The size of the extracted support vector array and the created machine learning model are both displayed and compared with earlier work to demonstrate the appropriateness of the suggested approach for use in an embedded system.

Additionally, the embedded implementation of our suggested method's computation time, average current consumption, and embedded memory utilization are reported in this work. The inbuilt microcontroller timer is used to measure computation time, and the STM32CubeIDE integrated development environment (IDE) provides information on memory consumption. The manufacturer's Nucleo-144 board datasheets specify the current measurement position where MCU average current consumption is measured.







SECTION III.
Improving Variable Step Dynamic Threshold Local Binary Pattern for Microcontroller-Based Machine Learning

A. Previously Proposed Variable Step Dynamic Threshold Local Binary Pattern
Yazid and Mahrus suggested the Variable Step Dynamic Threshold Local Binary Pattern (VSDTLBP) as a feature extraction technique for AF detection from ECG signals. In ECG segments were classified into AF or non-AF classes using histograms of LBP codes as the input to the SVM classifier.
A specific target data point's 8-bit VSDTLBP code is determined by taking the values of its eight neighbouring data points (4 before and 4 after). In contrast to the traditional LBP, the VSDTLBP establishes a step a defined interval between these data points that can be modified based on the particular application goal and signal characteristics. Additionally, unlike the traditional LBP algorithm, which uses the value of the centre data point as the threshold, the VSDTLBP algorithm uses a dynamically determined threshold value. Because of these, the VSDTLBP method was able to be readily modified to fit various time series signal types and classification issues.
An illustration of the operation of the VSDTLBP as described
LBP (x[i]) =∑_(r=0)^(p/2-1)▒〖X {S(x[i+(r-p/2)X step]-f(x,i,P,step)) 2^r+S(x[i+(r+1)X step]-f(x,i,P,step)) 2^(r+p/2)}〗
Where
S(a)={ 1, a≥0
{ 0,a<0

f(x,i,P,step)=1/(P+1) ∑_(r=i-(p/2)Xstep)^(r=i+(p/2)Xstep)▒〖x[r]〗
In these equations, f(x,i,P,step) is the dynamic threshold function, P is the binary bit length of the LBP code, and step is the selected step value between data points being considered for generating an LBP code. From the 256 possible 8-bit codes generated by Equation 1 and 2, only 58 types of local binary pattern (LBP) codes which are defined as the "uniform" LBP codes by Ojala et al. are used. Uniform LBP codes are LBP codes that have at most two transitions between "0" to "1" or "1" to "0" in its binary form representation. More detailed information about VSDTLBP and LBP can be found in previous works respectively.





B. Addition of Low Pass Filter to Remove Aliasing Effect
Any signal component with wavelength less than n times the sample period will produce an aliasing effect, a well-known adverse side effect of down sampling, because VSDTLBP with step value n will neglect n−1 data points between the data points utilized in the calculation of each LBP value. Pre processing the input signal with a low pass filter will lessen the aliasing effect and enhance the LBP codes' capacity to more precisely reflect signal features, hence improving classification accuracy.
With a cut-off frequency determined by Equation where fc is the LPF cut-off frequency, fs is the signal sampling frequency, and step is the LBP step value, a Butterworth low pass filter (LPF) was employed in the suggested work. In this study, the LBP step value is 6, which was proven to be ideal for 250 S/s ECG signals in earlier research and the signal sampling frequency is 250 samples/second.
fc=fs/(2X step)
C. Feature Size Reduction Based on Correlation and p-Values
Even though uniform LBP feature vectors are smaller than those of other approaches, a smaller machine-learning model can be produced by further shrinking the feature vector size, which will save computation time and memory. For a microcontroller implementation, the size of the machine learning model is especially crucial because internal flash memory, which must also be shared with other program components like the graphical user interface, is usually limited to a few hundred kilobytes to a few megabytes.
Eliminating superfluous features is one method of reducing the feature vector size. It makes sense that only one feature is required if many features are highly associated with one another, as adding the other connected features will only provide a small amount of information to the machine learning model.
Prior research has suggested the Pearson correlation coefficient as a measure of feature redundancy. In this work, Pearson correlation coefficient between input features is utilized as a criterion of feature selection. Only one feature is chosen and the others are eliminated when there is a strong correlation between two or more features. The feature selection procedure uses just 10,000 randomly chosen 10-second ECG segments (5000 AF and 5000 non-AF) from the MIT-BIH Atrial Fibrillation database ECG signals in order to avoid overfitting.
In feature selection, the P-value has also been widely employed, especially to eliminate features that aren't relevant .To ascertain each feature's significance in differentiating between the two classes (AF and non-AF), we run statistical tests on it. Large p-value features are eliminated because they are deemed inconsequential.
D. Efficient Circular Buffer Implementation
A popular data structure in systems with limited resources, including low-power microcontrollers, is the circular buffer. A simple circular buffer construction with length L is shown in Figure 2. The most recent data entered into the circular buffer in Figure 2 is denoted by d[n], the data that came right before the most recent by d[n−1], and so forth. The oldest data in the circular buffer is found in d[n−(L−1)] if its length is L. The oldest data must be removed when new data is received, and the new data must be added to the buffer (d[n]) to become the most recent data.
The circular buffer architecture is a very effective way to implement the VSDTLBP and its derivation as proposed in this research.
The AF detection system in our implementation, which is given in this work, continuously receives fresh ECG data points, creates the associated LBP codes, updates the LBP histogram, and then uses the LBP histogram as input characteristics to categorize the ECG signal using a trained machine learning model.
Only the detected AF ECG segment and the data points surrounding it need to be saved or sent to the cloud in order to reduce the expenses associated with data transmission and storage. A circular buffer with a length equal to the number of data points to be sent or stored whenever AF is detected can be used to do this with ease


.
The suggested approach allows for the calculation of a new LBP code that corresponds to datapoint d[n−(4×step)] each time a new ECG datapoint d[n] arrives. The reason for the 4×step delay is that, assuming an eight-bit length LBP code, the suggested algorithm requires the preceding four datapoints as well as the subsequent four datapoints for each datapoint in order to calculate its associated LBP code. A circular buffer that is roughly the same length as the input segment is used to store the computed LBP codes. The LBP circular buffer will always contain the LBP codes corresponding to the most recent segment of the ECG input signal. This is because the newest LBP code in the buffer will always replace the oldest one after the initial ECG segment has been received.
Additionally, the LBP histogram will be updated following the computation of each new LBP code. This procedure entails subtracting 1 from the value stored in the histogram bin corresponding to the oldest LBP codes eliminated in the prior step and adding 1 to the value recorded in the histogram bins corresponding to the new LBP code. The LBP histogram is now prepared for use as a machine-learning model's input characteristics. Following this phase, input scaling or normalization a typical component of a machine learning inference process flow-can be introduced to increase classification accuracy. Figure 3 shows a flow chart that depicts the procedure described in the preceding paragraphs.



One major advantage of the suggested approach is that, when applied in real-time to a stream of ECG data points, only one new LBP code needs to be computed for each new data point that arrives. This is followed by a straightforward procedure that involves updating the values of two bins in a histogram and replacing the oldest value in a circular buffer. Assuming that at least one initial segment length has already been processed, a complete input vector is now prepared for the inference procedure. A machine learning inference with the most recent VSDTLBP histogram as its input can be performed at each sample period because, in general, this process is quicker than the ECG signal sampling time.
E. Machine Learning Inference Implementation in Microcontroller
In this study, we propose to use a low-resource microcontroller to perform the machine learning inference process at the edge. STMicroelectronics' Arm Cortex M-based microcontrollers are the target microcontrollers. A personal computer with a lot of memory and processing capability is used to train the support vector machine (SVM) model initially. Support vectors are then taken out of the model that was created and added to the C software that is installed on the target microcontroller as an array.
The radial basis function (RBF) kernel and the sci-kit-learn module are used in Python to train SVM models. The trained model with the highest accuracy is saved for use in the microcontroller after ten stratified cross validations. The Arm CMSIS-DSP library supplied by Arm serves as the foundation for the microcontroller embedding.

SECTION IV
Conclusion

In this study, we suggested a better way to identify atrial fibrillation using an ECG signal. The suggested approach successfully decreased the size of the training model while increasing classification accuracy. The suggested technique, when implemented on an Arm Cortex M4-based microcontroller, can do an SVM inference in 11.28 ms with an average current of about 27 mA, allowing for near real-time atrial fibrillation detection on a low-cost and low-power device. Our suggested approach can achieve as high classification accuracy as a PC-based solution when implemented in a low power microcontroller, in contrast to many previously proposed edge device-based inference systems, which typically need to implement techniques that reduce classification accuracy (like model pruning) in order to be able to run as an embedded system. The work described in this paper can help lower the cost and improve the quality of smart wearable ECG monitor devices, which will lessen the burden on patients, healthcare providers, and their families. However, additional work, such as testing with actual ambulatory ECG data recorded from a wearable device, is still required. , Claims:CLAIM 1:
The proposed method enhances the detection of Atrial Fibrillation (AF) from ECG signals by improving classification accuracy and simultaneously optimizing the model size, making it more efficient without compromising performance.


CLAIM 2:
Implemented on an Arm Cortex M4-based microcontroller, the proposed method can run an SVM inference within 11.28ms using as low as 27mA average current, enabling near real-time atrial fibrillation detection on a low-power and low-cost device.
CLAIM 3:
The proposed approach reduces the cost and enhances the quality of smart wearable ECG monitor devices, thereby alleviating the burden on medical professionals, patients, and their families by providing an accessible and efficient solution for monitoring Atrial Fibrillation.

Documents