Conference PaperPDF Available

Live Demonstration: An AIoT Wearable ECG Patch with Decision Tree for Arrhythmia Analysis

October 2019

DOI:10.1109/BIOCAS.2019.8919138

Conference: 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS)

Authors:

Show all 6 authorsHide

Content uploaded by Shuenn-Yuh Lee

Content may be subject to copyright.

Live Demonstration: An AIoT Wearable ECG Patch

with Decision Tree for Arrhythmia Analysis

Yu-Jin Lin

, Chen-Wei Chuang

, Chun-Yueh Yen

, Sheng-Hsin Huang

, Ju-Yi Chen

and Shuenn-Yuh Lee

Department of Electronic Engineering, National Cheng Kung University, Tainan, Taiwan

Department of Internal Medicine, National Cheng Kung University Hospital, Tainan, Taiwan

Email: jerry71520@gmail.com, ieesyl@mail.ncku.edu.tw

Abstract—This live demonstration presents a novel

electrocardiogram (ECG) monitoring system with artificial

intelligence of things (AIoT) design, which is based on decision

tree (DT). The proposed system includes a front-end device

and a software system. The front-end device includes a solar

charging circuit, a wireless charging circuit and an analog

front-end circuit. First, the solar charging block takes a dye-

sensitized sorlar cell from National Chung Hsing University,

which is responsible for energy harvesting under indoor

illuminance. Second, the wireless charging block gives users an

additional charging method to meet the demand of long-term

monitoring. Third, the analog front-end block is composed of

the ECG-sensing circuit, the microcontroller unit (MCU) and

the Bluetooth Low Energy (BLE) module. The ECG-sensing

circuit is based on single lead measurement, and it includes

level shifter units, differential amplifiers and filters. The

circuits are implemented by the commercial components and

realized by self-designed print circuit boards (PCB). On the

other hand, this paper takes ARM Cortex M4 and BLE 5.0 as

the solution for data transmitting and encoding. All the above

circuits are integrated into one PCB, and the prototype is

mm* 50 mm* 20 mm. The software system includes an

application (APP) with DT algorithms, a cloud server is

available to execute DT training and to provide a user interface

for supporting telemedicine. This paper proposes a simplified

DT model, which can be realized in APP based on iOS system.

The APP classifies real-time ECG data into different

arrhythmias, and the delay latency is 500 ms in average.

Meanwhile, according to 4G or Wi-Fi, the collected ECG data

are uploaded to the cloud server for training DT. Then, the

coefficients of the pre-trained DT will be sent back to the APP

for updating. The accuracy is 98.7%. By the proposed AIoT

system, doctors and users can realize the task of long-term

ECG monitoring, which is valuable for cardiovascular disease

diagnosis. Also, doctors can assist users instantly by the web

user interface, to meet the demands of telemedicine. The

proposed AIoT system has been conducted human trials in

National Cheng Kung University Hospital. The power

consumption of the proposed front-end device is 8.25 mW, and

it can be continuously used up to 32 hours with a 120 mAh

lithium-ion battery. If it turns on solar charging, the device can

continually operate, until the solar cell is dead.

I. D

EMONSTRATION

ETUP

For the demonstration, our team will bring the proposed

ECG patch prototypes, smartphone, tablet and laptop, as

shown in Fig. 1. The ECG patch will be wirelessly connected

with smartphones or tablets by Bluetooth to display the real-

time electrocardiogram (ECG) signals, simultaneously,

which can help physician to diagnose the arrhythmia on

health examination. Furthermore, to meet the scenario of

telemedicine, the web user interface is also available on this

demonstration to display subject’s ECG signals, which can

be accessed by health-care workers by anytime and

anywhere to monitor the cardiovascular disease of patients.

II. V

ISITOR

XPERIENCE

The live demonstration is based on submitted 2019

BIOCAS concurrent full paper [1]. Our team will demo how

to finish a full measurement of ECG signals by the proposed

ECG patch. Participants who may interest can have the

measurement by themselves using the devices of ECG patch.

Meanwhile, the cardiologist expert in our team will describe

the relationship between ECG signals and arrhythmia. It is

expected that participants can have a practical measurement

and an overview of those bio-signals, as shown in the

attached video [2]. In addition, participants will not only

understand the difference between the traditional ECG Holter

and the proposed ECG patch, but also know how to use this

ECG patch on health examination. In this demonstration, we

hope to bring participants the knowledge includes the

biomedicine, the cardiovascular disease, internet of things,

wearable devices and the cloud server for telemedicine.

EFERENCES

[1] An AIoT Wearable ECG Patch with Decision Tree for Arrhythmia

Analysis, submitted to BIOCAS 2019, Paper ID:8164.

[2] Video link : https://youtu.be/DHMtppW0xTY

Fig 1. The proposed demonstration setup

Fig 2. The actual ECG patch measurement

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on June 14,2021 at 02:06:39 UTC from IEEE Xplore. Restrictions apply.

Edge2Analysis: A Novel AIoT Platform for Atrial Fibrillation Recognition and Detection

Article

May 2022

Atrial fibrillation (AF) is a serious medical condition of the heart potentially leading to stroke, which can be diagnosed by analyzing electrocardiograms (ECG). Technologies of Artificial Intelligence of Things (AIoT) enable smart abnormality detection by analyzing streaming healthcare data from the sensor end of users. Analyzing streaming data in the cloud leads to challenges of response latency and privacy issues, and local inference by a model deployed on the user end brings difficulties in model update and customization. Therefore, we propose an AIoT Platform with AF recognition neural networks on the sensor edge with model retraining ability on a resource-constrained embedded system. To this aim, we proposed to combine simple but effective neural networks and an ECG feature selection strategy to reduce computing complexity while maintaining recognition performance. Based on the platform, we evaluated and discussed the performance, response time, and requirements for model retraining in the scenario of AF detection from ECG recordings. The proposed lightweight solution was validated with two public datasets and an ECG data stream simulation on an ATmega2560 processor, proving the feasibility of analysis and training on edge.

Edge2Analysis: A Novel AIoT Platform for Atrial Fibrillation Recognition and Detection

Preprint

Full-text available

Feb 2022

Processing signals on edge nodes has become a popular trend, which requires data to be stored in terminals and processed, so it has the advantages of low latency, low power consumption and privacy protection. However, due to the limited resources of embedded devices, few studies have made relevant attempts. Our task was to implement a task for atrial fibrillation based on edge devices, using data from the open dataset CinC2017 and CPSC. The classification accuracy is about 90% and the model training can be completed on the edge platform. It is a feasible method to retrain the model by changing the training dataset of equipment.

Classification of arrhythmia disease through electrocardiogram signals using sampling vector random forest classifier

Article

Full-text available

Dec 2022
MULTIMED TOOLS APPL

An electrocardiogram (ECG) is an electrical signal produced by ECG sensors to examine and visualize the heart’s functionality, quick identification of arrhythmia aids in proper care and reduction of the risk factor. This paper uses machine learning algorithms to detect and identify arrhythmia types by extracting simple dynamic features from the ECG signals. The workflow will undergo three stages: pre-processing, feature extraction, and classification. A Low Pass Filter (LPF) and a two-stage median filter are proposed to eliminate all the signaling artifacts for successful feature extraction. The pre-processed signal is then fed into extracts of dynamic ECG features. With the aid of extracted features, the proposed Sampling Vector Random Forest Classifier (SVRF) predicts the form of arrhythmia. The novelty of the proposed SVRF is the ensemble-based approach for multi-class classification has overcome the noise resistance and gained inaccuracy when compared to another state-of-the-art classifier. The proposed method was tested with all the possible eleven types of arrhythmia from the MIT-BIH arrhythmia database, as stated in the association for the advancement of medical instrumentations recommendations. With the help of the proposed (SVRF), 98.22% of accuracy is achieved, which is better than other existing techniques. These aid medical professionals in the early detection of arrhythmia stages. The performance and comparative analysis are presented in this paper to prove the efficiency of the proposed method with the performance metrics of the confusion matrix and receiver operating characteristics.

Blockchain-Envisioned Secure Authentication Approach in AIoT: Applications, Challenges, and Future Research

Article

Full-text available

Oct 2021
WIREL COMMUN MOB COM

The Artificial Intelligence of Things (AIoT) is the amalgamation of Artificial Intelligence (AI) methods and the Internet of Things (IoT) infrastructure, which are deployed there to improve the overall performance of the system. AIoT can be deployed to achieve more efficient IoT operations; thereby can improve human-machine interactions and provide better data analysis. AI methods can be used to transform IoT data into useful information for the better decision-making processes, and it further increases the overall usability of the system. AIoT frameworks are very useful and applicable in a variety of applications, like security and surveillance system, smart home, intelligent transportation system, smart farming, secure and safe healthcare monitoring, industrial automation and control, eCommerce, logistics operations and control, and many more. However, AIoT frameworks may have issues related to data security and privacy as they are vulnerable to various types of information security-related attacks. These issues further cause the serious consequences, like the unauthorized data leakage and data update. Blockchain is a specific type of database. It is a digital ledger of transactions, which is duplicated and distributed across the entire network of computer systems. It stores data in the form of some blocks, which are then chained together. Blockchain is tamper proof and provides more security as compared to the traditional security mechanisms. Hence, blockchain can be integrated in various AIoT applications to provide more security. A generalized blockchain-envisioned secure authentication framework for AIoT has been proposed. The adversary model of blockchain-envisioned secure authentication framework for AIoT is also highlighted that covers most of the potential threats of a kind of communication environment. Various applications of the proposed framework are also discussed. Furthermore, different issues and challenges of the proposed framework are highlighted. In the end, we also provide some future research directions relevant to the proposed framework. 1. Introduction AIoT is the combination of Artificial Intelligence (AI) methods and Internet of Things (IoT) infrastructure. As we know, IoT is about different “Things” (i.e., smart IoT devices), which are connected to various users through the Internet [1–3]. AI provides methods to train the devices so that they can understand the novel data based on the training procedure that they have completed. AIoT works for a common goal, which is the generation of useful data about the world through IoT devices and drawing of useful insights from the collected data through some AI methods [4–9]. Artificial Intelligence (AI) refers to any human-like intelligence manifested by a machine, i.e., a computer and a robot. It is the ability of a machine, which mimics like the learning capabilities of the human, i.e., learning from the experience, object recognition, decision-making, and problem solving. It is an interdisciplinary science with multiple methods and tools, especially the advancements, which happen due to machine learning and deep learning creating a paradigm shift in various sectors of the tech industry [10–14]. 1.1. Common AI Technologies The following are the common AI technologies [1–3, 15–19]: (i)Speech recognition: speech recognition mechanism is used to convert and transform human speech in some useful format so that a computer application can process it. The “transcription and transformation of human language” into some useful format is in demand these days.(ii)Natural language processing (NLP): it focuses on the interactions between computers and human languages. Text analysis methods are used to analyse the structure of sentences and their interpretation through the ML algorithms. NLP is also helpful for the fraud detection systems. Automated assistants and applications derive unstructured data through NLP.(iii)Image recognition: it is the process of identification and detection of features in a video or an image file. It further facilitates the process of image searches for example, detecting license plates and diagnosis of diseases.(iv)Machine learning platforms: it is a subdiscipline of computer science and the important part of AI. The motive is to develop new mechanisms to enable the learning of computer systems to make them more intelligent. With the deployment of various algorithms, application programming interface (APIs), training tools, big data analytics, and machine learning platforms become popular these days, which are used for the purpose of categorization and predictions. The organizations like Amazon, Fractal Analytics, Google, and Microsoft provide various ML platforms as per the requirements of the customers.(v)Decision management: due to the AI logic and capabilities, machines can be used for training and maintenance. For adding value to the business and to make it more profitable, decision management system is being used. With the deployment of ML-based mechanisms, these systems execute automated decision.(vi)Deep learning platforms: the deep learning (DL) techniques use artificial neural networks. DL is another form of ML, which duplicates the neural network of human brain to process the data and draw patterns from this. These patterns are further used in the decision-making process. Some of the applications of DL are speech recognition system, image recognition system, and prediction system, which can predict about any phenomena of the digital sphere. Some of the DL platforms providers are Deep Instinct, Ersatz Labs, Fluid AI, and MathWorks.(vii)Robotic process automation: robotic process automation depicts the functioning of corporate processes, which automate the process through the mimicking of human activities and tasks. However, it is essential to mention that AI is not there to replace the humans, but to support and complement their skills and associated tasks. The organizations like automation anywhere, blue prism, and WorkFusion are working in this domain(viii)Cyber security: it is a computer defense mechanism, which detects and defends the various information security-related attacks happening in the cyber space. Neural networks, which have the ability to process sequences, can be deployed with ML techniques to create learning technologies for the mitigation of cyber attacks.(ix)Marketing automation: these days, AI is also going famous for marketing automation, especially to predict about the market trends, offers, and customers’ mood. That happens because of the advantages it put into the domain.(x)Virtual agents: a virtual agent can be a computer agent or a program, which has the ability of interaction with humans. It is used in customer service’s system via chat bots. The organizations like Apple, Google, Amazon, and Microsoft provide support through virtual agents. 1.2. Categories of Machine Learning Algorithms The different categories of machine learning algorithms are given below. 1.2.1. Supervised Learning Algorithms This class of algorithm uses labeled data to learn (training). There is a mapping function that turns input variables into the output variable . In other words, it solves function as per the following equation: Supervised learning algorithms are further divided into three categories, i.e., classification, regression, and ensembling. (1)Classification: in this method, variables are in the form of certain categories, and then, there is a prediction about the outcome of the given sample. A classification model might look at the input data and try to predict labels for example, in case of cyber attack detection, threat, or normal flow.(2)Regression: there are another categories of algorithms, which come under regression, which is used to predict the outcome of the given sample, in case when output variables are in the form of real values. A regression method may be used to process input data for the prediction of amount of rainfall and temperature in a specific week of a month.(3)Ensembling: for this particular category, a combination of different algorithms is used to produce better results, i.e., “Bagging with Random Forests” and “Boosting with AdaBoost.” “Linear Regression, Logistic Regression, CART, Naive-Bayes, and -Nearest Neighbors (KNN)” are examples of supervised learning algorithms. 1.2.2. Unsupervised Learning Algorithms These algorithms are used when only one input variable () is there and no corresponding output variables. The unlabeled training data is used to model the underlying structure of the data. Unsupervised learning algorithms are further divide into three categories, i.e., association, clustering, and dimensionality reduction. Their details are as follows: (1)Association: this method is used to find out the “probability of the cooccurrence of items in a collection.” Most of the time this technique is used in market-basket analysis. For example, it can be used to find out if a customer purchases shirt, he/she is 80% likely to also purchase trouser.(2)Clustering: this method is used to group samples such that objects in a cluster are more similar to each other than to the objects in the other clusters, i.e., K-means and clustering algorithm.(3)Dimensionality reduction: in unsupervised learning, there is another approach called as dimensionality reduction, which is used to reduce the number of variables of a dataset without omitting the important information. This can be done through “feature extraction methods” and “feature selection methods.” Feature selection mechanism selects a subset of the original variables whereas feature extraction does data transformation from a “high-dimensional space” to a “low-dimensional space.” Algorithm Principal Component Analysis (PCA) comes under feature extraction approach. Algorithms like Apriori, -means, and Principal Component Analysis (PCA) are examples of unsupervised learning. 1.2.3. Reinforcement Learning It is the another category of ML algorithms, which allows an agent to decide the next best action to be performed on the basis of its current state via learning behaviors, which will maximize a reward. These methods learn optimal actions in a trial and error way, for example, a video game in which the player has to move to certain places at certain times to earn more points. The player may correct his/her move on the basis of previous loses attempts. Examples of reinforcement learning algorithms are Q-Learning and State-Action-Reward-State-Action (SARSA). 1.2.4. Deep Learning Deep learning is another category of ML algorithms, which are inspired by the structure and function of the human brain. They are based on artificial neural networks and representation learning. They are concerned with building much larger and more complex neural networks. Deep learning algorithms are concerned with very large datasets of labelled analog data, i.e., image, text, audio, and video. Some of the popular deep learning algorithms are Generative Adversarial Network (GAN), Convolutional Neural Network (CNN), Recurrent Neural Networks (RNNs), and Long Short-Term Memory Networks (LSTMs). Here, learning can be performed in a supervised or unsupervised way. The different categories of machine learning along with algorithms are also depicted in Figure 1.

Live Demonstration: AI-Based Arrhythmia Detection with Wearable ECG and Cloud Platform for Dog Health Monitoring

Conference Paper

Nov 2024

A Flexible and Miniaturized Chest Patch for Real-time PPG/ECG/Bio-Z Monitoring

Conference Paper

Jul 2022

We proposed a lightweight and wearable chest patch for real-time monitoring of three vital signals, photoplethysmography (PPG), electrocardiography (ECG), and bioimpedance (Bio-Z). It comprises a flexible electrode patch and a miniaturized wireless signal acquisition module. Heart rate (HR), heart rate variability (HRV), and blood pressure (BP) can be extracted from the raw signals. The flexible electrode patch is comfortable for the user while maintaining stable contact with human skin, guaranteeing the wearability. Size of the signal acquisition module is only 17.3mm×14.5mm×9mm, and it weighs only 3.2g, including an 80mAh lithium polymer battery, which keeps the entire patch working for more than 4 hours. A host controller, involving a graphic user interface (GUI) is developed to receive and visualize the data from the chest patch. The proposed device successfully collected three vital signals with high signal quality and showed its potential in healthcare applications.

Enabling on-device classification of ECG with compressed learning for health IoT

Article

Jul 2021
MICROELECTRON J

In this paper, an on-device classification of electrocardiography (ECG) with Compressed Learning (CL) for health Internet of Things (IoT) is proposed. A CL algorithm combining with one-dimensional (1-D) Convolutional Neural Network (CNN) that directly learns on ECG signals in the compression domain without expanded normalization is proposed. Such an approach bypasses the reconstruction step and minimizes the raw input data dimension that significantly reduces the processing power. An automatic network optimization framework with Automatic Machine Learning (AutoML) tool Neural Network Intelligence (NNI) is suggested to adapt the network structure search problem introduced by input dimension reduction with Compression Ratios (CRs). That ensures the minimized model size and operation number under a guaranteed classification accuracy. To implement the resized 1-D CL classifier in hardware, which has different kernel sizes, strides, and output channels under different CRs, a flexible architecture is proposed to further lower power consumption. Evaluated on the MIT-BIH database, the specific 1-D CNN network selected under CR =0.2 achieves a Macro-F1 of 0.9214 on 5-class ECG signal classification, with a 6.4× reduction in FLOPs and a 2.6× decrease in model size with a only 0.028 loss in Macro-F1 compared with the uncompressed situation. Synthesized in UMC 40 nm Low Power process, the hardware architecture with the 1-D CL classifier achieves an energy efficiency of 0.83 μJ/Classification under a 1.1-V power supply at a frequency of 5 MHz.

ResearchGate has not been able to resolve any references for this publication.

Live Demonstration: An AIoT Wearable ECG Patch with Decision Tree for Arrhythmia Analysis

Recommended publications

A topological approach to delineation and arrhythmic beats detection in unprocessed long-term ECG si...

Learning Decision Tree for Selecting QRS Detectors for Cardiac Monitoring

Decision Tree Predictive Learner-Based Approach for False Alarm Detection in ICU

A meta heuristic model to select optimum features and classify cardiac arrhythmia ECG signals data i...