11.1 Introduction

11.2 Related Work

In the medical domain, monitoring is the observation of a disease, condition or one/several medical parameters over time. In hospitals, patients’ ECG must be monitored constantly, which is usually done by doctors or paramedical staff. They observe ECG of patients continuously and maintain a record of it. This process is quiet slow and bit expensive [12] for which there is a great need of developing advanced monitoring system to monitor ECG continuously.

Rahman et al. [13] discussed about a smart patient monitoring system which will display the health information of patients automatically using multiple sensors as shown in Figure 11.1. The collected information is stored in IoT cloud after it is being processed through Raspberry Pi. The system extracts the primary data as ECG by using ECG sensor (single lead heart rate monitor (AD8232)) and secondary data as temperature by using temperature sensor (DS18B20) and by continuous monitoring the patient’s health conditions can be tracked by doctors, nurses and relatives remotely. Based on the criticality of the condition, a warning is conveyed to the user, and the doctors/nurses/relatives can start a video call. This helps in constant remote caregiving facilities for monitoring condition of patients, where the physicians received the text notification if there are any changes in the patient’s condition. The system contains a push switch that can be pressed by the patient during emergencies like when they get uncomfortable. This will send an SMS notification and make a video call to the people who are responsible for attending like doctors/relatives in order to enable access to the patient’s condition through multimedia such as video. Additionally, there is also an option of live stream to monitor condition of the patient through the website. This can be attained using the camera module of Raspberry Pi positioned near the patient.

Budida et al. [14] suggested an idea of smart healthcare system utilizing the concepts of IoT as shown in Figure 11.2. Patients’ data are collected using various IoT sensors and are transferred to the microcontroller ATMEL 89s52, which stores the data in the MySQL database server. The MySQL database server is employed to control the collected information and provide accessibility, and it can then be viewed using available android application by patients. If any abnormality is detected in the collected data, the patient is warned through a notification and a message is dispatched to their respective caregiver. The abnormality in the trend in data is detected by employing various decision-making algorithms. People can then have access to the database in order to check their medical records. Hence, this system can provide a better health monitoring framework. Thus, the system in [14] can be exploited to detect the condition of the patient and hence provide a rapid and effective mechanism. The attributes of the patient required for patient monitoring are collected using different sensors. Then, the collected data are analyzed and stored, and the results are presented on the server to be accessed by doctors and patients. The data is also available on mobile applications and webpage to be viewed by doctors and patients in addition to booking appointments, emergency button, patient’s review, alert, single and family registration and so forth.

Riaz et al. [15] proposed an IoT-based system to capture various vital signs such as ECG signals, blood pressure, body temperature and heart rate of patients. Their proposed system also utilizes a Wireless Sensor Network (WSN) with IoT to build a continuous monitoring system. The data are transferred through wireless medium, for instance Internet cloud computing and Zigbee, so that doctors can view the data. The proposed system can be implemented in areas lacking proper healthcare facilities where it is difficult to get reasonable healthcare. Using wireless transmission to acquire data, the work showed that such a continuous monitoring system can be built to monitor and report patient’s condition over Internet. The data collected from the sensor are processed and analyzed in the microcontroller and then sent to doctors and patients’ receiver-end node through wireless module of Zigbee. The data then are dispatched to the mobile application as well as the caregiver’s web repository. The flowchart of the whole process is represented in Figures. 11.3 and 11.4.

Bhoyar et al. [16] indicated the progress in multiparameter health tracking system using Arduino to detect illness. A disease identifying algorithm was formulated to identify certain illnesses such as Hyperthermia, Dysautonomia and so on, using various parameters, e.g., metabolic conditions, weight, oxygen level, temperature, heartbeat pulse, stress level and blood pressure. Their introduced system acts as a monitoring and early detection system utilizing the Arduino environment. The collected sensor data are sent to a gateway over Bluetooth. Thus, this research focused on the WSN; different sensor devices are present in a wireless network that are used to collect users’ data in order to monitor their health conditions by taking into consideration two aspects, namely sensor deployment (Figure 11.5) and disease detection algorithm.

The common shortcoming of the aforementioned related work is their lack of consideration for developing a portable secured edge device capable of locally processing and analyzing the data for localized decision-making. We consider this to be the contribution in this chapter by showing how to employ distinct technologies to build a simple yet accurate enough ECG monitoring device that can be leveraged for patient-monitoring applications and further leveraged for bio-signal authentication in specific use-case scenarios.

11.3 Proposed AI-Based System Architecture

11.3.1 Block Diagram

The main motivation of the model includes monitoring the patient’s health conditions periodically and developing a secured and cost-efficient remote patient monitoring edge-based ECG system using sensors that predict results using machine learning algorithm which decides the patient’s abnormalities. Our proposed system will be using sensors that detects health problem with high accuracy by using some AI technology that tracks patients’ health on edge devices. The patient dataset obtained is preprocessed before any algorithms or analysis is performed. The dataset used for training the system is MIT-BIH Supraventricular dataset and the test data is the live data which is obtained from the Raspberry Pi module. The training dataset is classified and trained with multiple machine learning algorithms as mentioned earlier, and the best algorithm based on the performance accuracy is selected to process the test data. The pre-trained data model is loaded on the SoC, and the result and evaluation are displayed on the screen. The entire proposed procedure is summarized in Figure 11.6 below.

To connect the heart rate sensor to SoC, we need to have an idea about the GPIO pin diagram of SoC as shown in Figure 11.7, which helps us in identifying which pins are to be used to connect the device and the sensor together.

Figure 11.8 shows how the connection is to be made between the sensor and the board, and we can also see how analog to digital converter acts as an intermediate device between the sensor and the board. Here, the pins of ADC, such as voltage supply and ground, are connected to the respective supply and ground pins in raspberry pi. Other pins such as clock, input and output are connected to the GPIO pins of the Raspberry Pi module.

11.3.2 Data Collection and Pre-Processing Steps

MIT-BIH Supraventricular dataset from PhysioNet is used for the evaluation of the adopted early warning system involving ECG-based arrhythmia classification. Altogether, the data was divided in two cycles which includes futures of ECG such as PQRST wave as depicted in Figure 11.9. The main feature is the QRS complex which identifies heart abnormalities. The dataset contains 184428 total records and 33 total attributes. It is divided into five main target classes which are – Normal (N), Supraventricular ectopic (SVEB), Ventricular ectopic (VEB), Fusion (F) and unknown (Q) beats.

Figure 11.10 depicts a representation of the raw ECG signals collected from the heart rate sensors which are later preprocessed to use as test set for the system. This preprocessing step of the signals consists of the following steps: data cleaning, matching, combining and removing noise and irrelevant data. Initially, the collected raw ECG signals were cleaned using three filters which were Band reject, High pass and Low pass filter. The purpose of these filters is to eliminate various noises from the raw signals and clean the data. The cleaned data is then passed to the heartbeat segmentation phase where various peaks of the heartbeat such as Q peak, S peak, R peak and so forth are identified to further extract features from the peaks. Finally, to extract meaningful features from the peak, the distance between the peaks is calculated. The distance between the peaks gives us an idea about the situation of the peak, which can help us assume various crucial features of the ECG signal. Therefore, various features regarding the heartbeat amplitude, heartbeat interval, morphological features and so forth are comprehended in this way. These extracted features are considered as the final feature set for feeding to the machine learning algorithms.

11.3.3 Detecting Heart Abnormalities Using AI-Aided Techniques

All the preprocessed data obtained from the previous step is then processed on a high-end computing device and the results in terms of accuracy rate are determined. Therefore, the pre-trained model having the best performance is then loaded on the Raspberry Pi.

The following steps are the intermediate steps for training the dataset and saving it into a pickle file.

Figure 11.11 demonstrates a chunk of the dataset that was extracted from the ECG signals. Here, the columns such as 0_pre_RR, 1_qrs_morph2, 1_qrs_morph3 etc. are the features of the dataset and the type column is our class label to detect supraventricular arrhythmia. The dataset was scaled using standard scaling method before passing it to the machine learning algorithms. The scaling was performed to normalize the range of the features

The problem was converted to a binary classification case where the Normal and SVEB classes in the dataset were converted as Type 0 and 1. The total number of records in the two classes are 162240 and 12480, respectively.

During the evaluation, a ‘pkl’ file is created for all the models. This file is used to load the trained ML model on to Raspberry Pi. The pkl file of the model having best performance is copied to Raspberry Pi for displaying the results on the device. The pkl file has a pre-trained model, and this model can be later used for testing the live data which will be obtained from the ECG sensor connected to the raspberry pi device. Once the file is downloaded it is ready to be run on the raspberry pi device provided all the requirements mentioned in 3.1 are already satisfied. The program can be executed by writing a simple python program and it displays the current health status of the person.

11.4 Considered Smart ECG Monitoring System

In this section, we present the hardware (sensing) and AI-logic components of our considered smart ECG monitoring system.

11.4.1 Edge Hardware Components

The centerpiece of the hardware setup for the considered smart ECG monitoring system is a microcontroller. Among various off-the-shelf microcontroller devices, Raspberry Pi 3 Model B+ is considered due to its cost-efficiency and performance for deploying pre-trained AI model for ECG classification with reasonable accuracy. The microcontroller and other hardware devices needed to construct the ECG monitoring system are described next.

11.4.1.1 System-on-a-Chip (SoC) Model

After rigorous study of different microprocessors which are readily available in the market, the Raspberry Pi 3 Model B+ is selected as the SOC model for constructing our considered smart ECG monitoring system. Figure 11.12 illustrates an image of the raspberry pi 3B+ model. This hardware model is based on an ARM Cortex-A53 1.4GHz processor operating on the Raspbian Operating System, a dual-band 2.4GHz and 5GHz Wireless Local Area Network (WLAN) interface, Bluetooth 4.2/BLE, and Ethernet with 300Mbps. A 700mA power supply is used for the operation of the SoC and the IoT peripherals.

11.4.1.2 IoT Sensor for Heart Rate Data Acquisition

An off-the-shelf optical pulse sensor is used to measure the heart rate of the user as shown is Figure 11.13. This has the capability of collecting reliable pulse readings from the fingertip or earlobe of the user by using the state-of-the-art on-sensor noise-cancellation and signal amplification circuits. The sensor does not require soldering to a system board since it consists of a three-pin cable that is terminated with standard male headers. Furthermore, the sensor consists of a luminous Light Emitting Diode (LED) and a light detector that allow light to pass through the finger and detected, respectively. As the heart pumps a pulse of blood through the blood vessels, the finger turns opaque to such an extent that a less amount of light reaches the detector. Furthermore, the light signal captured at the detector varies with each heart pulse that is translated into an electrical pulse and amplified so that a +5 V logic level output signal is generated. The LED blinks on the generation of each output signal that corresponds to each heartbeat. This analog output is connected to the microcontroller for converting the electronic signal to number of heart beats per minute.

11.4.1.3 Microprocessor and Analog to Digital Converter

Arduino microprocessor is selected to connect the analog pulse sensor to the SoC with the support of a cost-efficient, eight-channel and 10-bit analog to digital converter (MCP3008) which is illustrated in Figure 11.14. This converter uses serial interface employing successive approximation algorithm to sample the signal at 75 ksps and 200 ksps with the operating voltages of 2.7 and 5 V, respectively.

The above-mentioned set of hardware was integrated to develop our desired secured ECG monitoring system for collecting ECG signals. Figure 11.15 shows a demonstration of our developed system. Here, the ECG signal data are collected using the pulse heart rate sensor, which with the help of MCP3008 is converted to digital data. The digital data can then be accessed through the raspberry pi SoC.

11.5 Bio-Authentication Application of the Considered ECG Monitoring System for Specific Use-Cases

By implementing this IoT-based system in a single chip of logic, improved and economical health services can be provided for the users. The system can authenticate users based on their bio-signals such as ECG signals and make the procedure safe and secure. Since user can access the system on the edge, the entire process of heart monitoring becomes easily accessible. Thus, it is able to offer improved and reliable services related to healthcare in situations where the patient is far away from the doctor. Thus, this allows doctors to acquire data of the patients from far away and prescribe medicine accordingly more quickly and efficiently. It also enables patients and doctors across a country to connect for superior treatments by tracking and recording patient data in real-time. The proposed system can save hospital bill, valuable time and in some cases might avoid long queues in the hospitals. The proposed design provides a cost-efficient, lightweight alternative to the current technologies present in the medical centers which are less efficient and more expensive.

Next, the propositioned system can have a multitude of applications. Several continuous user monitoring applications such as car insurance monitoring apps observe user data to study the user's vehicle-driving behavior and conditions to decide when claims are made to the companies. These apps usually monitor the car speed, road conditions and weather conditions using various features of the user mobile phone. However, to activate and use the application, there is only one-time authentication, and it does not guarantee whether the registered user or any other unauthenticated person is driving the car. We introduce an ECG-based continuous biometric authentication scheme for all such applications where continuous user monitoring is required. We aim to identify the change from authenticated users to the unauthorized user for user monitoring apps. The proposed AI-aided model will analyze and find patterns or trends in the user’s heart activity by analyzing the ECG wave of each individual. The ECG data will be collected from the user and will be processed locally on edge devices adopting AI-aided techniques to also monitor if the person’s health condition is normal or not.

In order to demonstrate our bio-authentication system, we have constructed a machine learning model that will be able to differentiate authorized user from unauthorized user. For implementing it using ECG signals, we have utilized the same dataset containing the extracted features that were used for classifying supraventricular arrhythmia. The classification was carried out on only the normal cases of the dataset to show the efficacy of our proposed concept. In this case, we have classified the users, i.e., the record column as our goal is to authenticate the user. Therefore, we have considered two classes: authorized and unauthorized. We have run the experiment five times, where at each run, only one of the records (i.e., person) will be considered authorized and the rest of the users will be unauthorized. The ECG recordings of both authorized and unauthorized were shuffled and considered in both train and test set. We have applied Random Forest algorithm to classify the records because initially Random Forest performed better for supraventricular arrhythmiaclassification.

11.6 Performance Evaluation

Evaluation on performance of a classification algorithm is usually done on its accuracy. The estimations of an instance to that instance’s genuine value to the level of familiarity is known as accuracy. The dataset was split into train and test set with the ratio of 70 and 30, respectively. The models were validated using stratified five-fold cross-validation.

11.6.1 Supraventricular Arrhythmia Classification

This research used four machine learning models; decision tree, Random Forest, CNN and ANN to predict the chances of heart failure in a patient from the MIT-BIH Supraventricular dataset. The comparative study is performed on the results (Accuracy) of all the ML models as shown in Figure Figure 11.16. In Figure Figure 11.16 it is seen that the model with the best performance is Random Forest with 97.38% accuracy. Thus, the Random Forest model is used as the prediction model for integrating with raspberry pi.

The performance also depends on the sensor manufacturer as the capacity of accurate sensing varies from sensor to sensor and also depends on the manufacturer. The system requires high electronic circuit’s knowledge to reduce the noise from the sensor and get accurate results. The experimental results for the patient’s health status is 100% accurate for normal patients according to the live data which was tested against the training data using the machine learning algorithms.

The result achieved while classifying supraventricular arrhythmia data was compared with the performance of previous studies where similar analysis is performed. Table 11.1 shows a comparative analysis of our model with three studies, using the MIT-BIH Supraventricular dataset, in terms of their performance.

Problem	Classification model	Performance
Our model for classifying supraventricular arrhythmia	Random Forest	97.38%
Arrhythmiaclassification using ECG signals [21]	Convolutional Neural Network	93.31%
Classification of arrhythmia in selected feature set [22]	K Nearest Neighbor	92.8%
ECG-based arrhythmia classification [23]	Deep belief networks	96.94%

11.6.2 Authorized User Classification for Bio-Authentication System

The performance was evaluated based on Accuracy, Precision, Recall, F1 Score and their respective confusion matrix. For Random Forest algorithm, the number of trees was considered to be 150. As mentioned earlier, the experimentation was repeated for five users. The performance evaluation on the proposed bio-authentication system is illustrated in Table 11.2. Here, record represents the authorized user. It can be seen from the table that Random Forest is performing remarkably while classifying users based on their ECG signal features. The accuracy, precision, recall and F1-score are giving near-perfect values denoting that the ECG signals of each person is distinct from the other person and hence can be thoroughly identified.

Record	Accuracy	Precision	Recall	F1-score
800	0.999918	0.999918	0.999918	0.999918
801	0.999938	0.999938	0.999938	0.999938
802	0.990094	0.995021	0.990094	0.991723
803	0.992396	0.995277	0.992396	0.993286
804	0.998993	0.999038	0.998993	0.999005

The confusion matrix of the results was also plotted for the five experiments. The results are stated in Figure 11.17. It can be seen that the results of false positives and false negatives are almost insignificant and most of the instances are classified accurately.

Thus, we can conclude that implementing the bio-authentication system using Random Forest classification can be a useful method for authenticating the authorized users based on their ECG signals, thus making the ECG system more secure. One of the advantages of this system is that it employs the same dataset for arrhythmia classification and bio-authentication, thus saving valuable processing time along the way.

11.7 Challenges Involved with the Proposed System

The automated treatment employing various ECG techniques is a popular field; however, it still lacks various aspects in the implementation domain in terms of algorithms and so forth. Therefore, it is important to study more in this field in order to improve it. The necessity of automated ECG analysis system is required with the popularity of remote monitoring technology and long-term ECG recording devices. In order to facilitate a real-time processing of dynamic ECG signals, parameters like complicated, time-consuming algorithms, delayed restriction and the search time of distinguishing wave position window should be kept minimal. The noise of ECG signals should be removed in pre-processing steps and the QRS waves should be detected accurately. Additionally, various cases such as the false positives and missed detection rates can also be determined by the algorithms.

The above-mentioned challenges are faced during the initial phase, but there is a list of challenges that were faced during the implementation part of the system. We have listed the following challenges and how we overcame that during implementation.

• Connecting Analog Sensor to Raspberry Pi-based SoC without using Arduino Module: One of the important parts of the project was to connect the ECG sensor to the Raspberry Pi device. It would not be possible to use Arduino module as the portability is an important aspect for the system; therefore, we used a MCP3008 ADC to connect the sensor.

• Libraries to Access ADC: After we were able to connect the sensor with the board, it was still hard to communicate with the device because the libraries for the communication to ADC are not present in python which was a major programming language used for implementation. Therefore, the connection was designed with processing since it has libraries to communicate directly with the ADC device.

• Understanding the Sensor Signal and Plotting ECG Signal: One of the major challenges involved with our implementation was to understand the signal which was sent by the sensor and also plot it as ECG graph. The signal contains output only in bits and the sensor documentation helped in finding out the formulas for converting the signal to ECG Graph.

• Collecting Live Data for Processing Health Information: The live data collected was to be studied and the QRS component of the data was to be separated from the signal to generate the live dataset from the sensor device. Dividing the signal and saving it into a csv file ignoring the values was the solution we came up with to solve this problem.

• Machine Learning on Raspberry Pi Device: Raspberry Pi is known as a simple compatible device and known for its portability, but we wanted to include machine learning algorithms on our hardware and machine learning to the contrast requires very high-end computing device. Therefore, we used a pre-trained model which was later dumped on the device using ‘pickle’ to overcome the problem of training the data on the device and saved our times and effort in analyzing the real-time data.

11.8 Conclusion and Future Scope

In this chapter, we have presented a prototype of a secured ECG monitoring device with the integration of IoT technology. The proposed model will continuously monitor ECG signals to detect supraventricular arrhythmiaand provide doctors and patient with a more efficient medical service. In this research, the raspberry-pi-based health monitoring system was analyzed through IoT. Heartbeat sensors continuously sense the ECG signals, and the data is transformed through the machine learning algorithm and the prediction is done on edge to determine if the patient is suffering from heart abnormalities or not. There are multiple prediction algorithms used for doing the comparison, and based on best accuracy result the algorithms are further used for prediction. To achieve that, in this chapter, we have applied four different algorithms which are decision tree, Random Forest, ANN and CNN. Random forest performed the best with an accuracy of 97.38% with decision tree being the second best. Therefore, Random Forest was chosen as the most efficient algorithm to be deployed in our system. Some of the benefits of such systems include saving hospital bill, reduce waiting time and long lines in the hospitals. The proposed system provides a lightweight and cheaper alternative to the traditional heavy and costly devices in the hospitals. The system is easily portable, and the patients can be monitored easily from remote places. Hence, this system provides enhanced healthcare services, and is therefore more reliable in cases where the patient is far away from the doctor.

We introduced an ECG-based continuous biometric authentication scheme for all such applications where continuous user monitoring is required. We aim to identify the change from authenticated users to the unauthorized user for user monitoring apps. The proposed Random Forest model analyzes and finds patterns or trends in the user’s heart activity by analyzing the ECG wave of each individual. The ECG data will be collected from the user and will be processed locally on an edge device such as Raspberry Pi, adopting AI aided techniques to also monitor if the person’s health condition is normal or not. Currently, the supraventricular arrhythmia prediction is done based on QRS complexion from the live data and applying prediction algorithms for fast results. Further this live data that is collected can be used for live authentication purpose which can be used in user monitoring apps such as car insurance companies where ECG signal of each individual is identified and the change from authenticated users to the unauthorized user is identified.

In future, the bio-authentication system can be further developed to be implemented in other scenarios such as car insurance monitoring. For the analysis of pre-processed ECG, the PQRST waves will be used as a pivotal characteristic for the change point detection (CPD). Abrupt variations in time series ECG is considered as change points which can be employed to distinguish among different people’s heart activity. Comparing the fundamental differences in the PQRST wave of a person’s ECG and studying the behavior of the wave for a particular person we train our system and if there is a change detected in the wave, the system will report it and also find out the point where this change has occurred. Additionally, the ECG of the user will also be monitored using various machine learning algorithms for any possible abnormalities while driving which will ensure some degree of safety for the user while driving. The system model will be trained to find out whether the heart activity of the person is normal or is prone to any sudden cardiovascular abnormalities. Based on the decision, an automated alerting system can be implemented in the system architecture which will alert the user’s acquaintances about the user’s unusual heart condition. Thus, the system can help several heart patients who are unable to visit the doctor due to their critical condition and get timely treatments and avoid any kind of accidents that might worsen their situation.

References

1. 1. S. B. Baker, W. Xiang, and I. Atkinson, “Internet of things for smart healthcare: Technologies, challenges, and Opportunities,” IEEE Access, 2017, doi:10.1109/ACCESS.2017.2775180.
2. 2. S. L. Guo, L. N. Han, H. W. Liu, Q. J. Si, D. F. Kong, and F. S. Guo, “The future of remote ECG monitoring systems,” Journal of Geriatric Cardiology, 2016, doi: 10.11909/j.issn.1671-5411.2016.06.015.
3. 3. A. Bansal et al., “Remote health monitoring system for detecting cardiac disorders,” IET Systems Biology, 2015, doi:10.1049/iet-syb.2015.0012.
4. 4. J. Son, J. Park, H. Oh, M. Z. A. Bhuiyan, J. Hur, and K. Kang, “Privacy-preserving electrocardiogram monitoring for intelligent arrhythmia detection,” Sensors (Switzerland), 2017, doi:10.3390/s17061360.
5. 5. F. J. Martinez-Tabares, Y. J. Costa-Salas, D. Cuesta-Frau, and G. Castellanos-Dominguez, “Multiobjective design of wearable sensor systems for electrocardiogram monitoring,” Journal of Sensors, 2016, doi:10.1155/2016/2418065.
6. 6. S. Tian, W. Yang, J. M. Le Grange, P. Wang, W. Huang, and Z. Ye, “Smart healthcare: Making medical care more intelligent,” Global Health Journal, 2019, doi:10.1016/j.glohj.2019.07.001.
7. 7. Y. Yin, Y. Zeng, X. Chen, and Y. Fan, “The internet of things in healthcare: An overview,” Journal of Industrial Information Integration, 2016, doi:10.1016/j.jii.2016.03.004.
8. 8. S. D. Greenwald, R. S. Patil, and R. G. Mark, “Improved detection and classification of arrhythmias in noise-corrupted electrocardiograms using contextual information within an expert system,” Biomedical Instrumentation and Technology, 1992.
9. 9. P. P. Ray, “A survey on Internet of Things architectures,” Journal of King Saud University - Computer and Information Sciences, 2018, doi:10.1016/j.jksuci.2016.10.003.
10. 10. M. S. D. Gupta, V. Patchava, and V. Menezes, “Healthcare based on IoT using Raspberry Pi,” 2016, doi:10.1109/ICGCIoT.2015.7380571.
11. 11. R. Kumar and M. Pallikonda Rajasekaran, “An IoT based patient monitoring system using raspberry Pi,” 2016, doi:10.1109/ICCTIDE.2016.7725378.
12. 12. S. Jokić, S. Krčo, V. Delić, D. Sakać, I. Jokić, and Z. Lukić, “An efficient ECG modeling for heartbeat classification,” 2010, doi:10.1109/NEUREL.2010.5644105.
13. 13. A. Rahman, T. Rahman, N. H. Ghani, S. Hossain, and J. Uddin, “IoT based patient monitoring system using ECG sensor,” 2019, doi:10.1109/ICREST.2019.8644065.
14. 14. D. A. M. Budida and R. S. Mangrulkar, “Design and implementation of smart HealthCare system using IoT,” 2018, doi:10.1109/ICIIECS.2017.8275903.
15. 15. M. H. Riaz, U. Rashid, M. Ali, and L. Li, “Internet of things based wireless patient body area monitoring network,” 2018, doi:10.1109/iThings-GreenCom-CPSCom-SmartData.2017.180.
16. 16. Priyanka Bhoyar and S. S. Sonavane, “Remote patient monitoring using disease detection algorithm,” International Journal of Engineering Research, vol. V4, no. 09, pp. 307–312, 2015, doi: 10.17577/ijertv4is090445.
17. 17.“Raspberry Pi documentation.” https://www.raspberrypi.org/documentation/usage/gpio/.
18. 18.“소리와 불빛이 나는 계단 프로토타입,” 2019. https://78project.wordpress.com/author/inkspelloo/.
19. 19.“Heart rate pulse sensor amped introduction.” https://microcontrollerslab.com/heart-rate-pulse- sensor/.
20. 20.“MCP3008 8-Channel 10-bit ADC IC.” https://components101.com/ics/mcp3008-adc-pinout- equivalent-datasheet.
21. 21. S. Sakib, M. M. Fouda, Z. M. Fadlullah, and N. Nasser, “Migrating intelligence from cloud to ultra-edge smart IoT sensor based on deep learning: An arrhythmia monitoring use-case,” 2020, doi:10.1109/IWCMC48107.2020.9148134.
22. 22. H. Leutheuser et al., “Comparison of real-time classification systems for arrhythmia detection on Android-based mobile devices,” 2014, doi:10.1109/EMBC.2014.6944177.
23. 23. S. M. Mathews, C. Kambhamettu, and K. E. Barner, “A novel application of deep learning for single-lead ECG classification,” Computers in Biology and Medicine, 2018, doi:10.1016/j.compbiomed.2018.05.013.