Due to the increased risk factors for cardiac diseases and a huge burden on the existing medical system and staff, preliminary heart health prediction is essential. The high cost of currently available home-based health monitors, along with the fact that ECG- based health predictions are available in only high-end or expensive smartwatches, there is a need for a low-cost, less resource-consuming heart health predictor model. Thus, we propose an Intelligent Heart Health Predictor Model (IHHPM) in this study. This model was based on supervised learning, where a support vector machine algorithm [with linear one-versus-one (OVO) classifier] was trained using the “ECG-ID” database from PhysioNet and tested with real-time data collected from a self-designed heart rate monitor. The performance metrics of the model were evaluated and then compared with other machine learning algorithms. Results indicate that the IHHPM is highly reliable for predicting ‘Risky’ and ‘Good’ heart conditions with high accuracy [area under the curve (AUC) ≈ 0.91 and 0.94, respectively]. IHHPM was trained with only three attributes (resting heart rate, age, and gender), making it one of the best heart health predictors compared to existing ones. Its efficiency (accuracy per unit attribute) was 30.35%. This heart health prediction can also be communicated to a remote physician once daily using IoT protocol. IHHPM can be used for preliminary heart health screening using low-cost and accessible hardware, thereby improving the health of the entire population.
Keywords
ECG-ID database
heart health prediction
machine learning
multi-class classification
support vector machine
Author information
Introduction
Globally, cardiovascular diseases (CVDs) are the leading cause of deaths in a year. CVDs include diseases like atherosclerosis, heart attack, stroke, heart failure, arrhythmia, ischemic heart disease, hypertensive heart disease, valvular heart disease, pulmonary heart disease and rheumatic heart disease [1, 2]. 31% of all the global deaths in the year 2016, i.e., approximately 17.9 million deaths were due to CVDs [2] (Figure 1). In 2015, 17 million people under 70 years of age died (premature deaths), 37% of which were due to CVDs [2]. 82% of these deaths were from low-income or middle-income countries [2], such as countries from the Indian subcontinent, Africa and Latin America.
Figure 1.
Graph showing CVDs as the main cause of death among top 10 reasons, globally [3].
CVDs are expensive to manage when diagnosed at a late stage. Early diagnosis can substantially reduce the mortality of CVDs through awareness and regular monitoring. However, due to the low doctor-to-patient ratio [4] and the cost of regular health check-ups, most people from low-income or middle-income groups avoid visiting doctors. Hence, they are diagnosed with CVDs at a much later stage. The doctor density in several countries including India is not uniformly distributed. In India, the doctor density in rural to urban regions is nearly 1:4 [5]. So, it becomes extremely difficult for the elderly and rural people to regularly monitor their heart health in a clinic.
Although there have been a lot of state-of-the-art studies in this domain mainly in disease prediction from ECG signals, where the focus was on ‘Arrhythmia’, overall heart condition has been hugely neglected. Diker et al. altered the traditional Extreme Learning Machine using a Differential Evolution Algorithm and proposed a novel classification method for ECG signals [6]. Nandhini Abirami et al. used the ensemble prediction method to improve the prediction accuracy of cardiac arrhythmias from ECG signals [7]. Based on R–R interval characteristics, Li et al. developed a multi-class classifier to predict or identify a few cardiac issues like left bundle branch block (LBBB), right bundle branch block (RBBB), atrial premature complex (APC) and ventricular premature beat (VPB) from ECG signals [8]. Alarsan and Younes developed a 4-class classifier for normal, premature ventricular contractions (PVC), premature atrial contractions (PAC) and others using the Gradient-Boosted Tree and Random Forest Algorithms [9]. Mustaqeem et al. used feature selection along with one-against-one (OAO), one-against-all (OAA) and error- correction code (ECC) of the support vector machine (SVM) algorithm for classifying arrhythmias [10].
Mehanović et al. used ensemble learning techniques based on artificial neural networks (ANN), SVM and k-nearest neighbor (kNN) algorithms and based on the majority vote, predicted heart diseases [11]. Akgül et al. developed an intelligent method for diagnosing heart diseases using a hybrid combination of ANN and genetic algorithm (GA) [12]. Another similar predictive system was proposed by Panda and Dash. They also compared their proposed system with other algorithms for validation [13].
However, these models were based on features extracted from ECG signals or require a detailed clinical history to work properly. Thus, these proposed models are not viable for a home-based self-heart monitoring system. There are three major problems with this approach:
Incorrect placement of ECG leads/electrodes by non-medical personnel lead to erroneous readings and hence faulty diagnosis [14].
These systems do not have the provisions for the incorporation of multiple vital signs monitoring sensors and are not suitable for regular basic health monitoring.
Moreover, they are not user-friendly enough to be used by non-medical personnel and to understand the implications of the result.
To overcome these problems, we have proposed an Intelligent Heart Health Predictor Model (IHHPM) which uses a simple photoplethysmography-based sensor that can detect and count a person’s heart rate along with simple associated data, like age and gender, which even non-medical personnel can accurately provide, to predict heart health. IHHPM has provisions to include other vital parameter monitoring sensors that can give a more detailed prediction about a person’s overall health. Moreover, the model is user-friendly as it collects data non-invasively and provides the prediction in non-medical terms for any person to understand its implications.
Primarily, heart health can be determined by monitoring the heart rate [15, 16] and with basic information, such as person’s age and gender. Hence, there is an urgent necessity for an Intelligent Heart Health Predictor Model (IHHPM). An ideal heart health predictor model must possess the following characteristics:
Low cost and minimal power consumption.
It should be user-friendly and easy to use, even without prior medical knowledge.
It should give fairly accurate predictions.
In this study, we addressed the issue of heart condition prediction and have proposed an IHHPM trained with data from the ECG-ID Database [17] of PhysioNet [18].
As depicted by an ECG signal in Figure 2, the QRS complex denotes ventricular depolarization. The R peaks of an ECG signal occur due to an electrical activity (depolarization) of the main mass of the ventricles indicating ventricular contraction, which ultimately pumps blood to the whole body [19]. This is the same as the pulse rate measured at the radial artery or any other pulse point of the human body. Thus, the time interval between two consecutive R peaks is the heart rate (in beats per minute or bpm). The heart rate data extracted from the filtered ECG signals of the training database was used to train the IHHPM.
Figure 2.
Graphical representation of a normal ECG signal.
The proposed model has been verified using data collected from our self-designed system [20]. The novel model predicts the heart health condition as “Good”, “Average”, and “Risky”. Any person whose heart condition is predicted as “Risky” must immediately consult with a doctor. Thus, it reduces the pressure on the doctors to provide preliminary regular heart check-ups. The Linear one-versus-one (OVO) SVM algorithm of supervised machine learning was used to predict a person’s heart condition. The performance of the IHHPM was assessed by different performance metrics.
Novelty of the proposed IHHPM
The novelties of the proposed IHHPM are as follows:
The model works on basic data that can be collected from cheap wearable devices, so heavy resources are not required to test the model and to predict.
The proposed model works with only 3 attributes and provides the best efficiency (30.35% accuracy per attribute) compared to existing models.
IHHPM is suitable for incorporation with smartwatches and wearables capable of reading heart rate data.
In our lab prototype, IHHPM worked perfectly with a simple self-designed sensor [20] and the data was used to validate the present model.
The model showed more than 90% accuracy while identifying both “Good” and “Risky” heart health conditions which is expected for a home-based screening device.
Even though the model’s SVM-OVO algorithm requires a slightly longer execution time than a few other algorithms while training, IHHPM predicts a person’s heart health condition within 4s.
Even if a person doesn’t use the model to check heart health manually, every day at a specific time set by the user, the system can send an email to the person’s physician via ThingSpeak IFTTT’s (If This Then That) IoT framework.
The paper is divided into four sections. Section 2 describes the methodology of the proposed model. The results are stated and analysed in Section 3 and the conclusions are presented in Section 4.
Methodology
In our proposed model, we are dealing with a classification problem to determine if an individual has a healthy heart as opposed to one who is at risk of having cardiac problems. The prediction of unhealthy metabolism is based on a certain range of values of biomarkers. Hence, to categorize healthy and unhealthy conditions accurately, the training dataset must be properly labelled. Therefore, the use of supervised learning in our proposed model serves such a training purpose.
A model can be trained using various algorithms depending on the purpose and the criteria requirements of the training model and dataset used. The popular machine learning algorithms are as follows [21]:
Linear Regression (LinR)
Logistic Regression (LR)
Decision Tree (DT) or Classification and Regression Tree (CART)
Support Vector Machine (SVM)
Linear, using
One-versus-All (OVA) or One-versus-Rest (OVR) multiclass classifier
One-versus-One (OVO) multiclass classifier
Polynomial
Sigmoid
Radial Basis Function (RBF)
Naive Bayes (NB), including Gaussian NB
k-Nearest Neighbour (kNN)
Linear Discriminant Analysis (LDA)
Principal Component Analysis (PCA)
Random Forest (RF)
For our proposed system, we required a supervised machine learning multi-class classifier to predict the heart health of a person in terms of three independent classes— GOOD, AVERAGE and RISKY. Most of these algorithms are used to train and test the IHHPM to judge their performances and based on the result, the most suitable algorithm is chosen to develop the proposed model. The Gaussian NB and RF algorithms were not tested since our dataset had correlated data points (unsuitable for Gaussian NB algorithm) and there were no distinct class boundaries (unsuitable for RF algorithm).
The entire predictive model is elucidated in the following subsections—System Architecture, Model Training and Model Validation. The complete workflow for the IHHPM development is given in Section 2.4, in the form of an algorithm. The entire code has been executed in the Python language using the Google Colab Integrated Development Environment (IDE).
System architecture of the IHHPM
In the proposed system, the ECG-ID database [17] was downloaded from the PhysioNet [18] server. This database has filtered ECG signals and the R–R intervals provide the heart rate of the volunteer. The database is also annotated with the age and gender of each volunteer. The information provided by the American Heart Association (AHA) was used to determine the threshold values and designate the class labels for our multi-class problem. The collected data were classified into the designated class labels to prepare the training database. Based on this training database, the proposed IHHPM was developed and trained. The validity of this model was then verified by taking into account the performance metrics of the model when executed on the testing database, prepared by collecting the heart rate of several volunteers using a self-designed heart rate sensor [20] along with their age and gender, and labelled using the AHA’s heart rate threshold. The proposed IHHPM can then send the data to a doctor at a remote-end, using an Internet of Things (IoT) communication link, for the required management, at least once daily as a precautionary measure. The entire system architecture of the proposed model is shown in Figure 3.
Figure 3.
System architecture of the proposed model, IHHPM.
Training procedure of IHHPM
To train our model, we followed a three-step procedure—training data collection, preparing the training dataset and training the proposed model.
Training data collection
To train our proposed model, we used the ECG-ID dataset [17] from PhysioNet [18]. The attributes of the ECG-ID dataset are tabulated in Table 1.
Attribute
Description
Total ECG recordings
310 ECG recordings
Number of persons
90 persons
ECG lead
Lead I
Recording duration
20 seconds
Sampling rate
500 Hz
Resolution
12-bit
Nominal range
±10 mV
Annotations
10 annotated beats (R- and T-wave peaks) from an automated detector
Information in ‘.hea’ File
Contains age, gender, and recording date
Number of men
44 men
Number of women
46 women
Age range
13 to 75 years
Source of records
Volunteers (students, colleagues, and friends of the author)
Number of records per person
Varies from 2 to 20 (depending on periodicity of collection)
Noise in raw signal
Contains both high and low-frequency noise components
In order to prepare the training database, based on the information presented by the American Heart Association (AHA), the following criteria were used to select the valid training signals.
Age > 18 years
Heart rate > 49 bpm for gender =
“Male”
Heart rate > 54 bpm for gender =
“Female”
After the signal selection process, 282 ECG recordings were finalized for preprocessing. Signal 1 of each valid recording was considered as it provided the filtered ECG signal, devoid of noise.
The average RR interval of each ECG signal was determined by calculating the mean interval between the various consecutive RR intervals of the signal. This average RR interval was then used to calculate the heart rate (in bpm) for each valid record, using Equation (1).
According to AHA, depending on age and gender, the resting heart rate of a person can be a preliminary deciding factor of overall heart health. Table 2 shows the resting heart rate (in bpm) for different age groups along with their respective fitness conditions of both male and female volunteers [22]. This table was used to define the proposed three-class Heart Health Predictor Model.
Target class
18–25 years
26–35 years
36–45 years
46–55 years
56–65 years
65+ years
M*
Fˆ
M*
Fˆ
M*
Fˆ
M*
Fˆ
M*
Fˆ
M*
Fˆ
Good
49–65
54–69
49–65
54–68
50–66
54–69
50–67
54–69
51–67
54–68
50–65
54–68
Average
66–73
70–78
66–74
69–76
67–75
70–78
68–76
70–77
68–75
69–77
66–73
69–76
Risky
74+
79+
75+
77+
76+
79+
77+
78+
76+
78+
74+
77+
Table 2
Resting heart rate chart.
*M = Male & ˆF = Female. All the heart rate values are given in beats per minute (bpm).
After labelling, the prepared training database was found to have 148 ‘Risky’, 77 ‘Average’, and 57 ‘Good’ data points. Synthetic minority over-sampling technique (SMOTE) is utilized to oversample the ‘Average’ and ‘Good’ classes to balance and normalize the training database.
Training the proposed model
The proposed prediction model was designed for a multi-class classification problem and needs to be fairly accurate, as it is related to human health. SVM algorithm of supervised learning is the most suitable one for training a model addressing such a multi-class problem. Table 2 shows that all the classes are separable from one another and no overlapping of classes is seen. Hence, the linear type of classifier for learning and the OVO approach for the multi-class classification was considered [23].
Validation of the IHHPM
Validation of the proposed model is a critical aspect since the device is intended for precautionary medical benefit. Validation also ensures a high-quality reliable device and minimizes the chance of mistakes.
Testing data collection
Heart rate information in bpm from 55 volunteers in the age group of 20–52 years was collected, using our self-designed sensor [20]. The group comprised 35 male and 20 female volunteers. The designed sensor was based on reflective photoplethysmography (rPPG) and collected the reflected infrared (IR) signals from the fingertip of a person, which were then processed to give the heart rate in bpm. During this measurement, the corresponding person’s age and gender were also noted, similar to the training dataset.
All these measurements were performed under resting conditions (relaxed, sitting position for 10 minutes) of the person with an ambient temperature of 32 °C–34 °C.
Testing the proposed model
The collected heart rate, gender and age information from multiple volunteers was considered to be the testing dataset, which was then used to test/validate the proposed model.
Algorithm of the IHHPM
The entire prediction procedure of the proposed IHHPM is depicted as an algorithm.
Remote data transmission
After the heart health prediction by the proposed model, the predicted information in the form of an email was sent once daily to the remote physician via ThingSpeak IFTTT’s IoT protocol as a precautionary measure [24]. The frequency and the time of the sent email can be adjusted by the user. Additional trigger criteria could have been added, like sending the email only during ‘Risky’ predictions. Since this model’s purpose was just a preventive screening model and not for emergency medical use, the real-time data transmission interval and rate did not play an important role.
The predicted data was uploaded to the cloud server using Application Specific Interface (API) keys. These API keys were not hardcoded to avoid security risks, to enable hassle-free code updation, and consistent program execution across all environments and hardware platforms. To improve the security, flexibility, and portability of API keys, they were stored in separate environmental variables. The uploaded information could be reviewed by by the user and remote physician by using their designated account’s username and password.
Results and analysis
The trained classifier with the balanced training dataset using the SVM-OVO algorithm provided >90% accuracy while identifying the “Risky” and “Good” heart health cases. The linear OVO hyperplanes separating the three classes in the trained IHHPM are shown in Figure 4.
Figure 4.
Trained IHHPM using the SVM-OVO algorithm, showing the balanced training data points and the hyperplanes.
Performance of proposed classifier model
The performance of the proposed classifier model executed on the testing dataset was analysed using the normalized Confusion Matrix (Figure 5) which showed that 100% of the ‘Risky’ cases of the testing database were identified correctly.
Figure 5.
Normalized Confusion Matrix of the proposed model on the testing dataset.
From the given confusion matrix (Figure 5), the binary normalized confusion matrices of all three classes were designed (Figure 6) using the general format of a binary classification confusion matrix as shown in Table 3.
Figure 6.
Normalized confusion matrices showing TP, TN, FP and FN values for all the three classes—(a) Risky, (b) Average, and (c) Good.
Predicted values
Positive
Negative
True values
Positive
True positive (TP)
False negative (FN)
Negative
False positive (FP)
True negative (TN)
Table 3
General format of a binary classification confusion matrix.
This classifier was designed to predict “Risky” heart conditions more accurately compared to average heart health. The main reason for this choice is that for “Risky” cases, immediate medical attention was deemed necessary. Considering the parameters of Table 3, the FN cases must be minimal for the proposed IHHPM. This ensured that for our proposed model, no ‘Risky’ heart condition is predicted as another condition.
Various performance metrics were determined as shown in Figure 6 and Equations (2)–(6).
The results obtained are shown in Table 4. Figure 7 shows the required Receiver Operating Characteristics (ROC) curve for the multi-class proposed classifier. The ROC curve is very significant for predicting diseases. It measures the performance accuracy of a classifier at various thresholds. From the ROC plot, the Area Under the Curve (AUC) was calculated for all these individual classes.
Figure 7.
ROC plot of the proposed model on the testing dataset for all three predictive classes.
Class
Accuracy
Precision
Recall
F1 score
Specificity
AUC
Risky
0.909
0.844
1.000
0.915
0.821
0.911
Average
0.873
0.917
0.647
0.759
0.974
0.811
Good
0.964
0.909
0.909
0.909
0.977
0.943
Table 4
Performance metrics of the proposed model.
The prediction accuracy of each of the classes can be easily obtained from the AUC of the ROC plot, as shown in Figure 7. The Risky class was predicted correctly in 91.05% of cases, the Average class was correctly identified in 81.05% of cases, and the Good class was identified accurately in 94.3% of cases. The high accuracy in Risky and Good class prediction is very significant as a medical benefit. However, the Average class was misinterpreted as Good or Risky classes in many cases, but it posed no risk to the person. The effectiveness of the proposed system lies in high accuracy during Risky and Good heart condition prediction. The high accuracy in Risky condition prediction is highly desirable as a person can immediately seek hospital care. Since our system is attached to an IoT communication link, the person can immediately contact a doctor (remotely) even if he/she does not have a nearby medical centre. This technology makes our system highly beneficial for rural areas.
The overall accuracy of the proposed system was calculated to be 0.873 due to uneven class distribution in the collected testing dataset. This accuracy is fairly good for a heart health screening model that can nudge the user to visit a doctor for a Risky prediction, effectively serving as a precautionary health device.
Comparison of different algorithms
The performance of the proposed model using SVM (linear OVO) classifier algorithm was compared with various other algorithms and classifiers.
The performance of different supervised learning algorithms (mentioned in Section 2) were compared using different performance metrics, such as Precision, Recall and F1 Score [calculated using Equations (3), (4), and (6)] and are displayed as a spider plot in Figure 8. The plot clearly identifies SVM, LR, and LDA as the best-performing algorithms.
Figure 8.
Comparison of different algorithms for the proposed IHHPM, in terms of their performance metrics. The green highlighted cases show our chosen algorithm.
Figure 9 depicts the performance of the algorithms in terms of the overall accuracy of the trained IHHPM. It is evident from the spider plot that SVM, LR, and LDA have the highest accuracy.
Figure 9.
Comparison of different algorithms in terms of their overall accuracy. The orange highlighted cases show our chosen algorithm.
It is evident that the SVM (OVO) classifier outperformed almost all algorithms. The performance of LR and LDA algorithms was similar to our proposed model (as shown in Figure 8 and Figure 9). However, the SVM (OVO) classifier algorithm was preferred over them for the following reasons:
SVM determines the hyperplane that maximizes the margin between classes. This typically leads to better generalization of unseen data—especially important in high-dimensional spaces.
SVM’s decision boundary depends only on support vectors, not all data points. This makes it less sensitive to outliers, unlike LR and LDA, which consider all data points.
LDA assumes that features are normally distributed with equal covariance matrices across classes. On the other hand, LR assumes a specific functional form with linear boundaries and independent errors, which maximizes log loss in this probabilistic approach. However, SVM makes no assumptions about data distribution—making it more versatile in real-world scenarios.
The SVM (OVO) algorithm builds binary classifiers for n classes. This leads to simpler problems per classifier, as each sub-problem involves only two classes. These individual binary problems are often more linearly separable, leading to better local decision boundaries.
The hyperplane (used to separate the different classes and acts as the class boundary) in an SVM algorithm can vary depending on the type of classifier used. Based on the nature of these classifier boundaries, different SVM classifier algorithms show different responses and hence classify differently. Figure 10, Figure 11, and Figure 12 show the performance comparison of different SVM classifiers in terms of their performance metrics (Precision, Recall and F1 Score) and their overall accuracy, and training time respectively. The spider plots of Figures 10 and 11 demonstrate that the SVM (OVO) classifier is better than other SVM classifiers based on performance metrics as well as its classification accuracy. Figure 12 depicts the comparison of code execution time taken for each of the SVM classifiers while training the model.
Figure 10.
Comparison of different SVM algorithms in terms of their overall accuracy. The green highlighted cases show our chosen algorithm.
Figure 11.
Comparison of different SVM algorithms in terms of their overall accuracy. The orange highlighted cases show our chosen algorithm.
Figure 12.
Comparison of different SVM algorithms in terms of their training time.
This, it can be inferred that the Linear [One-versus-One (OVO)] classifier was best suited for our purpose. OVO performed better than OVR classifier due to the following reasons:
It creates multiple balanced and easier binary sub-problems,
It can handle complex multi-class boundaries better and more efficiently,
It works better with linear kernels,
It is less sensitive to class imbalance.
Comparison with models in literature
The proposed IHHPM was compared with a few other predictive models, as shown in Table 5. This clearly validates our hypothesis that the proposed model can effectively identify “Risky” heart health conditions in a person utilizing only 3 parameters (resting heart rate, gender, and age) with a commendable accuracy of 91.05%. The primary reasons for choosing these three attributes are as follows:
Resting heart rate helps identify basic heart diseases like tachycardia, bradycardia, etc.
The gender of a person is important as “Female” persons generally have a smaller physical stature than “Male” persons and hence a higher resting heart rate.
Age is also a significant factor in understanding the heart health of a person as stated by AHA.
The model also showed the best efficiency (30.35%), compared to existing models, and expressed with Equation (7).
Comparative study of the proposed model with those in existing literature.
The computed efficiency shown in Table 5 is plotted as a bubble scatter plot in Figure 13. The size of the bubble is directly proportional to the efficiency of the proposed model.
Figure 13.
Comparison of efficiency (accuracy per unit attribute) of our proposed model with similar works from existing literature.
Conclusion
The proposed Intelligent Heart Health Predictor Model (IHHPM) has a significant societal impact. In developing or under-developed countries, where there are fewer doctors to cater to the needs of the entire population, the proposed IHHPM offers a great support. People can get an initial assessment from this predictor model and doctors can then prioritize the persons having risk factors for cardiac issues. This can significantly reduce the burden on the existing healthcare system and also improve the quality of available medical services.
The proposed model provided more than 90% correct predictions for patients with extreme cases of heart health, namely, the “Risky” and “Good” cases. It also identifies the patients at risk of having cardiac problems with very basic information like heart rate, gender and age of the person. This enables the general population with access to basic wearables and smartwatches to undergo a preliminary heart health screening along with their daily activities. This fast, effective, and less resource-consuming model, with a testing or predicting time of less than 4 s, can quickly run on low- power wearable devices and be the “first line of heart health diagnosis” for people in developing and under-developed countries. With 91.05% accuracy for “Risky” category, the proposed IHHPM serves as a good screening predictor model.
Despite its significant benefits, the proposed model is intended for preventive purpose only, and is considered suitable for preliminary heart health condition detection, utilizing just the basic data obtained from wearable devices, for personal and remote or home-based use. The model in its current form is not suitable for thorough medical diagnosis and is definitely not a replacement for medical staff or equipment.
With more datasets for upgrading the present model, the proposed model can become the perfect backing system for the medical community. Steps can be taken in the future to improve the IHHPM’s accuracy by fine-tuning its parameters. Incorporation of ECG sensing in low cost wearables can explore these signals and gain further insights into a person’s heart health at an affordable price. IoT communication of the IHHPM enables remote access to doctors for further medical advice in case of a “Risky” prediction. Additionally, the IoT framework can be improved with better security protocols.
This research did not receive external funding from any agencies.
Ethical statement
All subjects included in this work provided their consent for inclusion before they participated. The work was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Ethical Committee for Bio Medical and Health Research involving Human Participants, University of Calcutta on 27-02-2020 for the project, “Development of Prototype Smart Electronic System for Remote Health Monitoring”.
Data availability statement
Data is available from the corresponding author upon request.
Conflict of interest
The authors declare no conflict of interest.
Acknowledgments
The authors thank all the volunteers who provided data for testing the proposed model.
References
1.
American Heart Association. What is cardiovascular disease? [Internet]. American Heart Association [cited 2024 Nov 13]. Available from: https://www.heart.org/en/health-topics/consumer-healthcare/what-is-cardiovascular-disease.
2.
World Health Organization. Cardiovascular Diseases (CVDs) [Internet]. World Health Organization; 2021 [cited 2025 Aug 8]. Available from: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds).
3.
World Health Organization. The top 10 causes of death [Internet]. World Health Organization; 2024 [cited 2024 Dec 2]. Available from: https://www.who.int/news-room/fact-sheets/detail/the-top-10-causes-of-death.
4.
World Health Organization. Medical Doctors (per 10 000 population) [Internet]. The Global Health Observatory. World Health Organization; 2025 [cited 2025 Feb 17]. Available from: https://www.who.int/data/gho/data/indicators/indicator-details/GHO/medical-doctors-(per-10-000-population).
5.
Health Workforce (HWF). The health workforce in India [Internet]. In: AnandS, FanV, editors. Human Resources for Health Observer. vol. 16, Geneva: World Health Organization; 2016 [cited 2025 Feb 17]. Available from: https://www.who.int/publications/i/item/9789241510523.
6.
DikerA, AvciE, TanyildiziE, GedikpinarM. A novel ECG signal classification method using DEA-ELM. Med. Hypotheses. 2020;136: 109515. doi:10.1016/j.mehy.2019.109515.
7.
AbiramiRN, VincentPDR. Cardiac arrhythmia detection using ensemble of machine learning algorithms. In: DasK, BansalJ, DeepK, NagarA, PathipooranamP, NaiduR, editors. Soft Computing for Problem Solving. Advances in Intelligent Systems and Computing. Heidelberg: Springer Nature; 2019. p. 475–487. doi:10.1007/978-981-15-0184-5_41.
8.
LiT, MaJ, PanX, ZhaiY, ManK. Classification of arrhythmia using multi-class support vector machine. In: Proceedings of the International MultiConference of Engineers and Computer Scientists. 2017. p. 710–713.
9.
AlarsanFI, YounesM. Analysis and classification of heart diseases using heartbeat features and machine learning algorithms. J Big Data. 2019 Aug 31;6: 81. doi:10.1186/s40537-019-0244-x.
10.
MustaqeemA, AnwarS, MajidM. Multiclass classification of cardiac arrhythmia using improved feature selection and SVM invariants. Comput Math Methods Med. 2018 Mar;1: 1–10. doi:10.1155/2018/7310496.
11.
MehanovićD, MašetićZ, KečoD. Prediction of heart diseases using majority voting ensemble method. In: BadnjevicA, ŠkrbićR, GurbetaPokvićL, editors. IFMBE Proceedings. Cham: Springer; 2019. p. 491–498. doi:10.1007/978-3-030-17971-7_73.
12.
AkgülM, SönmezÖ, ÖzcanT. Diagnosis of heart disease using an intelligent method: a hybrid ANN-GA approach. In: KahramanC, OnarSC, OztaysiB, SariIU, CebiS, TolgaAÇ, editors. Intelligent and Fuzzy Techniques in Big Data Analytics and Decision Making—Proceedings of the INFUS 2019 Conference. Advances in Intelligent Systems and Computing. Cham: Springer; 2020. p. 1250–1257. doi:10.1007/978-3-030-23756-1_147.
13.
PandaD, DashS. Predictive system: comparison of classification techniques for effective prediction of heart disease. In: SatapathyS, BhatejaV, MohantyJ, UdgataS, editors. Smart Intelligent Computing and Applications. Singapore: Springer; 2020. p. 203–213. doi:10.1007/978-981-13-9282-5_19.
14.
BondR, FinlayD, NugentC, BreenC, GuldenringD, DalyM. The effects of electrode misplacement on clinicians’ interpretation of the standard 12-lead electrocardiogram. Eur J Intern Med. 2012 Oct;23(7):610–615. doi:10.1016/j.ejim.2012.03.011.
15.
American Heart Association. What’s the difference between blood pressure and heart rate? [Internet]. In: All About Heart Rate. Dallas, TX: American Heart Association; 2024 [cited 2024 Oct 23]. Available from: https://www.heart.org/en/health-topics/high-blood-pressure/the-facts-about-high-blood-pressure/all-about-heart-rate-pulse.
16.
SolanM. Your resting heart rate can reflect your current and future health [Internet]. Harvard Health Publishing. Harvard Medical School; 2024 [cited 2024 Sep 5]. Available from: https://www.health.harvard.edu/blog/resting-heart-rate-can-reflect-current-future-health-201606179806.
17.
LugovayaTS. Biometric human identification based on electrocardiogram [Internet] [Master’s Thesis]. [Faculty of Computing Technologies and Informatics, Electrotechnical University “LETI”]; 2005 [cited 2024 Jun 13]. Available from: https://www.physionet.org/physiobank/database/ecgiddb/biometric.shtml.
18.
GoldbergerAL, AmaralLAN, GlassL, HausdorffJM, IvanovPc, MarkRG, PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation. 2000 Jun 13;101(23):E215–E220. doi:10.1161/01.CIR.101.23.e215.
19.
AshleyEA, NiebauerJ. Cardiology explained. London: Remedica Medical Education and Publishing; 2004.
20.
GhoshM, BasuS, PanditS, Barman (Mandal)S. Design of a health monitoring system for heart rate and body temperature sensing including embedded processing using ARM Cortex M3. In: DasA, NayakJ, NaikB, PatiS, PelusiD, editors. Computational intelligence in pattern recognition, advances in intelligent systems and computing. Heidelberg: Springer Nature; 2020. p. 93–103. doi:10.1007/978-981-13-9042-5_9.
21.
BrownleeJ. Master machine learning algorithms. Melbourne: 2017. Available from: https://machinelearningmastery.com/master-machine-learning-algorithms.
22.
KumarK. What is a good resting heart rate by age and gender? [Internet]. In: CunhaJP, editor. MedicineNet. WebMD, LLC; 2024 [cited 2024 Jul 12]. Available from: https://www.medicinenet.com/what_is_a_good_resting_heart_rate_by_age/article.htm.
23.
Har-PeledS, RothD, ZimakD. Constraint classification: a new approach to multiclass classification. In: Cesa-BianchiN, NumaoM, ReischukR, editors. Algorithmic learning theory. Cham: Springer; 2002. p. 365–379. Avaiable from: doi:10.1007/3-540-36169-3_29.
24.
BasuS, GhoshM, Barman (Mandal)S. Raspberry PI 3B+ based smart remote health monitoring system using IoT platform. In: KunduS, AcharyaUS, DeCK, MukherjeeS, editors. Proceedings of the 2nd International Conference on Communication, Devices and Computing. ICCDC 2019. Lecture Notes in Electrical Engineering. Cham: Springer; 2020. p. 419–426. Available from: doi:10.1007/978-981-15-0829-5_46.
25.
FarooquiNA, HaleemM, AhmadM, MurugesanP. Classification and prediction of cardiovascular disease using machine learning techniques. In: MalikH, MishraS, SoodYR, García MárquezFP, UstunTS, editors. International Conference on Signal, Machines, Automation, and Algorithm. SIGMAA 2023. Advances in Intelligent Systems and Computing. vol. 1460, Cham: Springer; 2024. Available from: doi:10.1007/978-981-97-6349-8_31.
26.
AhmedR, VarshneyA, AshrafZ, FarooquiNA, PathakRS. Enhanced stroke risk prediction: a fusion of machine learning models for improved healthcare strategies. SN Comput Sci. 2024;5: 1078. doi:10.1007/s42979-024-03389-w.
27.
AkareD, GaniU, BhongadeA, MureD, ChatterjeeM, RamtekeV. Heart disease prediction system using machine learning. IJARCCE. 2024;13: 92–97. doi:10.17148/IJARCCE.2024.13315.
