A Systematic Review of Machine Learning and IoT Applied to the Prediction and Monitoring of Cardiovascular Diseases

According to the Pan American Health Organization, cardiovascular disease is the leading cause of death worldwide, claiming an estimated 17.9 million lives each year. This paper presents a systematic review to highlight the use of IoT, IoMT, and machine learning to detect, predict, or monitor cardiovascular disease. We had a final sample of 164 high-impact journal papers, focusing on two categories: cardiovascular disease detection using IoT/IoMT technologies and cardiovascular disease using machine learning techniques. For the first category, we found 82 proposals, while for the second, we found 85 proposals. The research highlights list of IoT/IoMT technologies, machine learning techniques, datasets, and the most discussed cardiovascular diseases. Neural networks have been popularly used, achieving an accuracy of over 90%, followed by random forest, XGBoost, k-NN, and SVM. Based on the results, we conclude that IoT/IoMT technologies can predict cardiovascular diseases in real time, ensemble techniques obtained one of the best performances in the accuracy metric, and hypertension and arrhythmia were the most discussed diseases. Finally, we identified the lack of public data as one of the main obstacles for machine learning approaches for cardiovascular disease prediction.

everyday objects connected to the Internet without human interaction [4]. The application of the tools, principles, concepts, and techniques used in the accepted Internet of Things approach within the medical and healthcare sectors is called the Internet of Medical Things (IoMT) [5]. Machine learning, on the other hand, is defined as a branch of artificial intelligence techniques that extract knowledge from data, also known as predictive analytics or statistical learning [6]. In particular, these techniques derive models from data (i.e., information such as documents, audio, and images), where the resulting model is the final product of machine learning [7]. Machine learning applied to medicine can transform existing modes of healthcare delivery [8].
Technological solutions based on IoMT and machine learning can improve the quality of life of patients diagnosed with CVD by preventing risk conditions, as well as helping those without easy access to healthcare services. The use of IoMT through smart devices enables the real-time detection, monitoring, and prediction of CVD, as well as allowing for emergency communication (i.e., alerts sent to caregivers or hospitals), all integrated into a portable device. Over the past few years, IoT/IoMT has evolved rapidly, with sensors or devices that are more powerful and capable of monitoring the vital signs of patients with chronic diseases, making it one of the most important technologies. The integration of smart wearables through IoT/IoMT has a significant impact on modern healthcare systems, providing value to health-seekers, delivering high-quality and cost-effective services, and facilitating effective remote care. However, the amount of data generated by these devices in the cloud environment is a major concern, which has led to several challenges, including determining the best machine learning techniques to mine this data. Many applications and frameworks have been developed using machine learning and deep learning for CVD prediction and monitoring, improving the quality of healthcare and providing accurate results. Thus, it is important to be aware of and analyze the state-of-the-art of the techniques and technologies being implemented to predict, monitor, or classify CVD.
We conducted a systematic review to highlight the use of IoT/IoMT and machine learning in the detection, prediction, or monitoring of CVD. We adopted the PRISMA [9] guidelines (Preferred Reporting Items for Systematic Reviews) and the research scope was based on the application of the PICOC framework [10]. This systematic review aims to identify state-of-the-art CVD and machine learning approaches based on four main contributions: (i) IoT/IoMT devices for monitoring or predicting cardiovascular disease; (ii) different types of CVD in the population; (iii) machine learning applications for detecting, predicting, or monitoring CVD; and (iv) data sets used with such machine learning techniques.
The main goal of this paper is to identify the current state of IoT/IoMT in CVD detection or monitoring, machine learning techniques, and data sets used to predict or classify these conditions. Despite the existence of literature reviews on the use of IoT/IoMT technologies to detect, predict, or monitor CVD, a compilation of research proposals on the use of these technologies in combination with machine learning techniques and the data sets used remains lacking. We aim to collectively analyze data sets, IoT/IoMT wearable devices/smart devices/medical devices, machine learning approaches, and disease types. In this paper, we present the best-practice technological devices, relevant machine learning approaches, evaluation metrics, and their results over the past seven years.
The remainder of this paper is organized as follows: Section 2 describes the related surveys. Section 3 presents the methods, such as the research question, the scope of the study, the literature review, and the inclusion and exclusion criteria. Section 4 describes the selected papers using IoT/IoMT technology. Section 5 presents the selected papers applying machine learning techniques. Section 6 discusses the papers considered for further analysis. Finally, Section 7 concludes the paper with future directions. The structure of the paper, as a comprehensive study roadmap, is shown in Figure 1.

Related Surveys
Several articles have been published on IoT/IoMT technologies and machine learning techniques in healthcare. This section briefly reviews the studies detailing systematic reviews on IoT/IoMT technologies applied to CVD along with machine learning methods and their results. We found that some of these articles focused only on the machine learning approaches and their results in a specific disease, while others focused on the technologies applied for disease monitoring.
For example, along with some recommendations, Friedrich et al. [11] focused on applications of machine learning approaches related to cardiovascular drugs. They identified 215 studies in their systematic review using PubMed and Embase as research databases. They concluded that 87% of the methods used belong to the supervised learning context (tree-based methods being the most common, followed by network and regression analysis, and boosting approaches). Similarly, Hazra et al. [12] provided brief descriptions of 35 research papers published between 2006 and 2016 which examined computational methods for predicting heart disease. They concluded that, among the classification techniques, Decision Tree, Naïve Bayes, Artificial Neural Networks, Association Rule Mining, and Fuzzy Logic were the most commonly used. The data mining tools with better resultsin terms of practical execution-were Java, WEKA, Tanagra, and Matlab. On the other hand, Shameer et al. [13] discussed machine learning algorithms and potential data sources. They summarized the open-access biomedical and healthcare ontologies and big data resources in cardiovascular medicine for the development of machine learning resources. Furthermore, they assessed the potential limitations and challenges associated to implementing AI in medicine. Bolhasani et al. [14] explored deep learning techniques for healthcare IoT applications. They presented how deep learning can address telemedicine and ambient assisted living systems, machine health monitoring systems, human activity recognition, patient vital signs collection, and data fusion. Their survey divided the research studies into four categories: Medical diagnosis and differentiation applications, home and personal health applications, disease prediction applications, and human behavior recognition applications. They identified the number of studies by deep learning techniques used (convolutional neural networks being the most-used) and the proposed evaluation environments (data set, implementation, and simulation).
Other papers have focused on IoT/IoMT technologies applied in healthcare systems or applications. Huang et al. [15] focused on IoT technologies for health management systemincluding clinical device management, medication management, clinical data management, remote medicine, mobile medical care, and individual health management-with the purpose of serving as a starting point for future IoT/IoMT security management and design. Lin et al. [16] presented recent developments in monitoring various physiological signals using flexible sensors for CVD through pulse wave technology (ECG, PCG, PPG, and SCG/BCG); in particular, they focused on five types of signals that can reflect CVD using flexible sensing technology. Rahaman et al. [17] presented IoT-based smart health monitoring systems with their advantages and disadvantages, highlighting the design and implementation of these monitoring devices with respect to the patients. They summarized 13 studies from 2015 to 2019, including the year, feedback device, key hardware components, use, and cost. Panicker and P. Gayathri [18] classified their work into various categories, such as feature selection, heart sounds, heart images, heart rate variability, IoT/wearable technology, fuzzy systems, and predictive models. They presented a tabulated summary in their literature review highlighting the different machine learning techniques and data sets used, results, and research gaps. However, this summary included only 13 works published between the years 2015 and 2018. The remaining works were cited and described in the document.
Other papers have discussed relevant studies focused on chronic diseases such as diabetes, cancer, CVD, hypertension, and glaucoma. Dadkhah et al. [19] summarized the use of IoT for chronic disease management, and concluded that CVD is one of the highest priorities for the use of IoT in the context of developing countries. Their results included 92 studies classified into 47 focused on CVD, 37 on hypertension, 5 on cerebrovascular, 1 on rheumatic, 1 on rheumatism, and 1 on ischemic. Lamonaca et al. [20] focused on monitoring blood pressure from a metrological point of view, aiming to address the lack of traceability and reliability of BP measurements. They analyzed the vulnerabilities and opportunities of smart devices and wearables, including medical devices, in terms of accuracy and reliability. They focused on smart metering devices, Internet-connected devices, and devices that enable the implementation of the Internet of Medical Things (IoMT).
Since 2020, reviews have focused on artificial intelligence for disease diagnosis, highlighting the analysis of cardiovascular disease. Argha, Celler, and Lovell [21] reviewed AI-based blood pressure estimation approaches with a focus on recent advances in deep learning-based techniques. They concluded that deep learning methods make it possible to develop reliable and accurate blood pressure estimation algorithms/devices. They also noted the lack of adequate data sets on invasive and non-invasive blood pressure as standard references. Huang et al. [22] identified and described recent developments in the application of digital health to CVD, focusing on AI models driven by data collected from wearables. They reported the type of disease detected, the algorithms applied, the application, and the performance for machine learning and deep learning approaches. Hinai et al. [23] identified articles on the end-to-end deep learning analysis of resting ECG signals for the detection of structural cardiac pathology. They identified 12 articles, 3 of which detected left ventricular systolic dysfunction, 1 of which detected left ventricular hypertrophy, 6 of which detected acute myocardial infarction, and 2 of which detected stable ischemic heart disease. The performance measures used were AUC and accuracy. On the other hand, Faizal et al. [24] outlined various conventional models for assessing and predicting risk and compared them with AI-based approaches. They briefly reported some deep learning and machine learning algorithms, with their respective performance, focusing on the country, study area, risk factors, and performance.
Chen et al. [25] reported on the use of deep learning algorithms in medical technology applications, focusing on challenges and recommendations for ECG detection and classification. The highlights of this paper were as follows: algorithms for CVD detec-tion and classification, smart wearable devices and hardware based on deep learning and ECG, and recognition using ECG biological signs. Qureshi et al. [26] presented ambient assisted living solutions to reduce morbidity and mortality in patients with cardiovascular conditions. They focused on the application, devices, testing platform, monitored signals, features, and limitations for ambient assisted solutions and monitoring/clinical management. They also focused on the purpose, architecture, accuracy, deep learning methods, databases, and pre-processing of deep learning-based solutions for ambient assisted living. However, they only reviewed articles published between 2015 and 2019, and selected only 40 as relevant to their research. Bhushan, Pandit, and Garg [27] discussed how machine learning and deep learning approaches have been used for the analysis of various heart diseases. They classified the articles by summary, technique/tool, advantages, disadvantages, and performance measure. The existing works related to ensemble models using machine learning and deep learning, as well as a description of relevant data sets, were reported as a separate section.
The review of Rath et al. [28] included methods for feature extraction, selection, and reduction, as well as machine learning-/deep learning-based classification schemes, CVD data sets, and types of heart disease. They also listed some heart disease attributes identified in the 60 collected papers. Maurya et al. [29] reviewed studies on the early prediction of heart failure, determination of its severity, prediction of adverse outcomes, and improving patient adherence to medication. They focused on the parameters measured, endpoint, impact on heart failure, algorithms, evaluation measures, and the data (e.g., how patients were monitored, data sets). However, they did not explain the study selection, whether they used a methodology or guidelines for the review, or how many papers they found relevant. Kumar et al. [30] conducted a survey based on artificial intelligence to diagnose chronic disease including heart disease, stroke, and hypertension. They focused on healthcare applications, type of disease, data set, technique, reported outcomes, feature extraction, and the classification process for prediction.
Jasinska-Piadlo et al. [31] offered a comprehensive examination of the utilization of data science and machine learning in heart failure data sets. They summarized significant discoveries while critically assessing the effectiveness, suitability, and precision of various approaches. In different sections, they also reported the dimensionality of the used data sets, missing data, the performance of the algorithms (for the most-used machine learning methods) and, most importantly, how machine learning and data analysis impact heart failure problem-solving. Chakrabarti et al. [32] reported the diagnostic applications of wrist-worn devices in detecting multiple diseases, including cardiovascular conditions. They also provided a brief discussion of machine learning algorithms for wearable data analysis and addressed the current challenges associated with wearables and medical data. Finally, Guo et al. [33] reviewed the advancements in wearable devices-specifically, unobtrusive sensing technologies-that provide support and tools for the management of chronic disease. They not only focused on cardiovascular diseases, but also on chronic diseases for long-term health monitoring and patient management.

Methods
According to Xiao and Watson [34], literature reviews are essential for academic research. MacMillan et al. [35] stated that a systematic review provides a broad overview of a particular research topic. For this reason, literature reviews should have clearly defined inclusion and exclusion criteria. They should also present a comprehensive search that identifies all of the relevant literature, uses explicit and reproducible selection criteria for included studies, rigorously assess potential bias in the included studies, and systematically summarize the results of the included studies [35]. The purpose of a systematic review should be to answer an important, answerable question or to identify areas of high importance [36]. This review aims to map the landscape of IoT/IoMT technologies and machine learning techniques used for the detection, prediction, and monitoring of CVD.
This systematic review was guided by three research questions based on detecting, predicting, or monitoring CVD, as presented in Table 1. These research questions are based on the main contributions previously described in Section 1: IoT/IoMT devices for monitoring and predicting CVD; different types of CVD in the population; and machine learning applications for detecting, predicting, or monitoring CVD. This review also follows the PICO(C) (population, intervention, comparison, and outcome) template to develop answerable, researchable questions by considering the elements listed in Table 2, always in accordance with the three research questions defined in Table 1.

Data Sources
For accessibility, we focused on reviewed articles from databases such as PubMed, IEEE Library, Springer Link, and Science Direct, in order to retrieve relevant JCR proposals using IoT or IoMT for CVD detection, prognosis, or surveillance. Our time-frame was between 1 January 2016 and 9 May 2023, in order to include technologies that are new and have not yet been discontinued. The search terms consisted of the keywords listed in Table 3, derived from the template elements defined in Table 2. P Cardiovascular disease, heart disease, cardiovascular events, heart illness, heart condition I IoT, IoMT C Machine learning, deep learning, data mining O Early detection, detect, predict, monitor C Wearable devices, devices With the keywords described above, we proceed to construct the generic search string, which included some words to be searched only in the abstract (ab): (ab "cardiovascular diseases" or ab "heart disease" or ab "cardiovascular events" or ab "heart illness" or ab "heart condition") and ("IoT" or "IoMT" or "machine learning" or "deep learning" or "data mining") and (ab "early detection" or ab "detect" or ab "predict" or ab "monitor") and ("wearable devices" or "devices").
After retrieving the potential papers from the databases, we implemented the following inclusion criteria: (i) JCR studies only, (ii) articles written in English, and (iii) studies published between January 2016 and May 2023, based on compatibility and technological evolution. On the other hand, we applied the following exclusion criteria to the resulting set of papers: (i) Publication types other than JCR journal articles (posters, conferences, proceedings), (ii) articles whose full text was not in English, (iii) studies published before January 2016 and after May 2023, and (iv) articles detecting or predicting CVD by heart sounds. Due to accessibility issues, 24 articles were excluded from the PubMed, IEEE, Springer Link, and Science Direct databases, as they could not be downloaded.

Study Selection
We obtained 1500 articles in total after applying the inclusion/exclusion criteria, of which 39 were from PubMed, 555 from IEEE, 416 from Springer Link, and 490 from Science Direct. Figure 2 shows the number of studies found in each digital library consulted.
Studies were selected for this review using the four-step process (identification, screening, eligibility, and inclusion) according to the PRISMA flowchart shown in Figure 3. In the first stage of identification, we collected 1500 articles, of which 340 were reduced by title. Duplicates were removed at this stage, resulting in 1144 articles. In the screening stage, articles were thoroughly screened by reading the abstract to determine whether the article was focused on detecting, predicting, or monitoring CVD, and 700 papers that did not meet the inclusion criteria were excluded. At the eligibility stage, we read the full text of 444 papers to determine whether they were eligible for inclusion in the systematic review.  As shown in Figure 3, articles were excluded for several reasons (exclusion criteria) at the screening and eligibility stages. Each article was analyzed and classified by the authors, according to the following criteria: (i) area of study (monitoring, prediction, detection), (ii) disease (stroke, arrhythmia, atrial fibrillation, hypertension), (iii) data set used (public or private), (iv) approach (machine learning techniques), and (v) results (evaluation metrics: accuracy, precision, F1 score).
We conducted a peer review of the 162 potential studies answering the above three research questions to identify those that addressed the detection or prediction of cardiovascular disease and met the inclusion criteria. Appendix A shows these results, sorted by date. The 162 articles were divided into two main categories: CVD detection using IoT/IoMT and CVD detection using machine learning with public/private data sets. For the first category, we retrieved 78 papers, and 84 were retrieved for the second category. Figure 4 shows the JCR papers selected for this systematic review, divided into the two categories described above (i.e., papers that used IoT/IoMT technologies and machine learning-based CVD detection with public/private data sets). However, 32 of the papers in these two categories considered both IoT/IoMT technologies and machine learning-based CVD detection with public/private data sets.

Bibliometric Analysis
The 162 papers included in the final analysis were published between January 2016 and May 2023 ( Figure 5). We observed an increase in research output starting around 2019, with an even stronger increase in the following years. In 2022, we retrieved more papers related to the diagnosis, detection, prediction, and monitoring of CVD. The percentages of articles retrieved from the various databases are depicted in Figure 6. Science Direct had the major publication percentage (with 44%), followed by IEEE (with 35%). Springer Link had a 16% publication percentage, while PubMed contributed only 5% of the 164 papers selected for this review. We conducted a bibliometric analysis regarding authors, keywords, and journals. For each article, we obtained the database, journal, title, keywords or index terms, authors, number of citations, number of pages, and publication year. A word cloud was generated, according to the frequencies of keywords. The most frequently used keywords are highlighted in larger and bolder fonts, while the less frequently used keywords are highlighted in a smaller font in Figure 7. The keywords were grouped by similar words, and the most frequently used keyword was machine learning (43), followed by electrocardiogram (34), deep learning (30), and convolutional neural network (25).  However, some authors seemed to have the same surname (e.g., Li) and a similar first name beginning with Y, such as Li, Ye (3); Li, Ya; Li, Yuanlu; Li, Yaowei (2); and Li, Yixuan. Acharya, U.R. was the only author with seven written articles related to CVD diseases. Similarly, Ciaccio, Edward J., have written four articles related to CVD disease. In total, Acharya, U.R. had 1074 citations in the seven papers, while Cicaccio, Edward J. had 433. The top 10 most-cited articles of the 164 articles that were selected are shown in Table 4.  The most-cited articles were from the year 2019 and the IEEE database: Mohan, S., Thirumalai, G., and Srivastava, G. Science Direct had seven of the ten most-cited CVDrelated articles: three from 2018, two from 2019, and two from 2020. The most-cited article had 888 citations, whereas the second and third had 579 and 432, respectively. The journal word cloud is shown in Figure 9.
The journal with the most articles related to CVD was Biomed. Signal Process. Control (28) from the Science Direct database. In second place was IEEE Access (20), followed by Comput. Biol. Med.

Research on CVD Detection Using IoT/IoMT
The JCR papers selected as relevant for this review were classified according to the disease under study: Abnormality detection, arrhythmia, atrial fibrillation, blood pressure and hypertension, cardiovascular disease and heart disease, and another type of disease (e.g., aortic stenosis, arterial stiffness, chronic disease, chronic heart failure, myocardial infarction, and ischemic heart disease). Proposals may not address more than one disease, data set, or IoT/IoMT technology.

Abnormality Detection and Arrhythmia
An abnormality or arrhythmia refers to an abnormal heartbeat rhythm; this means that the heart may beat too fast, too slow, or with an irregular rhythm. Arrhythmia is caused by changes in heart tissue, activity, and the electrical signals that control the heartbeat, which may be caused by damage from disease, injury, or genetics. Although there are usually no symptoms, some people experience an irregular heartbeat. Symptoms may include disorientation, difficulty breathing, fainting, or dizziness. The most common test used to diagnose arrhythmia is an electrocardiogram (EKG or ECG) [122]. Atrial fibrillation is the most common kind of treated heart arrhythmia. In atrial fibrillation, the atria-the upper, smaller chambers of the heart-do not generate normal electrical impulses and, therefore, do not contract. This causes the ventricles-the main pumping chambers of the heart-to beat rapidly and irregularly. Although atrial fibrillation is the most commonly sustained arrhythmia, it is not common. For example, in a 22-year study of 5191 adult men and women, only 2% developed chronic atrial fibrillation [123]. Table 6 shows the proposals classified by abnormality and arrhythmia detection, as well as atrial fibrillation. There were 26 proposals, but only Sannino and De Pietro [40] used a medical device (unspecified Holter) for condition monitoring. Keyanfar et al. [52] used a cardioverter-defibrillator device, Raheja and Manocha [60] used an ECG machine without specifying the model, and Fayyazifar et al. [62] used a MAC 550HD and a MUSE V9 (GE Healthcare, Chicago, Illinois, USA). Venkataramanaiah and Meenakshi [51] used biomedical sensors for the detection of arrhythmia without specifying the device. Two proposals used a smartphone (specifically, the Sony Xperia Z series model) (Sony, Tokyo, Japan): Mehrang et al. [70] and Lahdenoja et al. [66]. Mehrang et al. [70] also used a continuous fivelead telemetry ECG (Philips IntelliVue MX40). Cai et al. [68] and Hill et al. [69] used portable devices (KardiaMobile and a Mason linear ECG lead system, respectively) (AliveCor Inc., CA, USA), (CardioCloud Medical Technology, Beijing, China). Yang et al. [67] used an integrated analog front-end for heart rate monitoring, while Rawal, Prajapati, and Darji [71] used the device ZYNQ Ultrascale ZCU106 FPGA (Advanced Micro Devices, Inc., Santa Clara, CA, USA).  [65] used an unspecified set of sensors nodes.
The best accuracy achieved was 99.94% by Al et al. [53] with AD8232 EKG Sensor (SparkFun Electronics, Niwot, CO, USA), Arduido board (Arduino, Scarmagno, Italy), and a Jetson Nano microcomputer (Nvidia Corporate, Santa Clara, CA, USA) using the R location correction (RLC) algorithm. The lowest accuracy was achieved by Yang et al. [67], with an integrated analog front-end using a linear kernel SVM. Other proposals have used metrics such as recall, precision, specificity, sensitivity, AUC, true positive rate, true negative rate, positive predictive, negative predictive, and ROC.

Aortic Stenosis
Aortic stenosis is an obstruction of blood flow from the left ventricular outflow tract which can occur at various levels, including at the aortic valve (valvular aortic stenosis), above (supravalvular aortic stenosis), or below the semilunar valve (subvalvular aortic stenosis). However, the clinical presentation of shortness of breath, syncope, and/or chest pain may be identical. Patients may have a systolic ejection murmur that is constant or changes with certain maneuvers (as in hypertrophic obstructive cardiomyopathy), as well as a variable intensity of the second heart sound, depending on the severity of the obstruction [125].
In Table 7, we list the proposals classified by aortic stenosis detection using IoT/IoMT. Yang et al. [72] used a three-axis accelerometer and three-axis gyroscope employing three classifiers: Decision tree, random forest, and neural network. Petrou et al. [73] used a non-implantable mixed-flow turbodynamic blood pump (Deltastream DP2, Xenios AG, Helibronn, Germany) with a cardiac output estimation pipeline, while Cheng et al. [74] used an ultrasound with two 3D convolutional neural networks (GE Vingmed Ultrasound AS, Norway Health Tech, Horten, Norway). The best performance was obtained by Yang et al. [72] combining seismo-cardiography (SCG) and gyro-cardiography (GCG) features with a random forest classifier (98.96%). The lowest accuracy was achieved by Cheng et al. [74], with an accuracy of 83%. 1 Seismo-cardiography: measurement of the linear acceleration components of the chest wall induced by the heartbeat. 2 Gyro-cardiography: recording of heart-induced rotational vibrations of the chest wall in the form of angular speed. 3 SCG and GCG signals can be acquired by placing a microelectromechanical system (MEMS) inertial measurement unit (IMU) on the chest wall.

Arterial Stiffness
Arterial stiffness is associated with changes in the structure and function of the arteries. Increased arterial stiffness is caused by impaired smooth muscle action, resulting in altered arterial dilation and constriction and increased blood pressure. In older adults, it is associated with isolated systolic hypertension and greater CVD risk. In addition to the functional regulation of blood pressure, structural changes within the vessels contribute to arterial stiffening. These changes include thickening and re-modeling within each of the three layers of the artery [126]. Table 8 shows the proposals classified by arterial stiffness detection using IoT/IoMT. Miao et al. [75] used a medical device OMRON BP-203RPE III (OMRON Industrial Automation, Kyoto, Japan) and multi-variate linear regression to achieve an accuracy of 89% for vascular age. A back-propagation neural network was employed to achieve an accuracy of 94% for CVD risk estimation. Dami and Yahaghizadeh [76] used a different set of unspecified sensors to achieve an accuracy of 88.42% through a combination of principal component analysis, deep belief networks, and a long short-term memory model. Asorey et al. [77] proposed a diagnostic system with an unspecified bio-sensors that can issue an alert within three hours if an artery appears to be blocked, and release medication in the next three hours if the artery is really blocked. Only two proposals have reported evaluation metrics such as accuracy. The lowest reported accuracy was 88.42% by Dami and Yahaghizadeh [76], while Miao et al. [75] reported 89% and 94%. Bio-sensors [77] Diagnostic system based on implanted devices and nano-nodes circulating in the cardiovascular system Diagnosis of blocked artery in 3 h, and medication released by another 3 h

Blood Pressure and Hypertension
Hypertension-or high systemic arterial blood pressure-is a common disease that develops when blood flows through the arteries at higher than normal pressure. There are two numbers that describe blood pressure: systolic and diastolic. The pressure created when the ventricles pump blood out of the heart is called systolic pressure, while the pressure between heartbeats-when the heart is filled with blood-is called diastolic pressure. Blood pressure changes throughout the day, based on activities [127]. The 2003 Guidelines of the European Society of Hypertension/European Society of Cardiology consider a patient as hypertensive when either the systolic blood pressure or diastolic blood pressure value is ≥140/90 mmHg [128]. Table 9 provides a summary of the proposals for blood pressure and hypertension detection using IoT/IoMT. Miao et al. [85] used a vital signs monitor (Benevision N12 Mindray), while Lan et al. [81] used a ring PPG, an accelerometer, and a Zigbee device on a custom-built device. Yan et al. [84] used a blood pressure monitor Finometer MIDI Model II, (Finapres Medical Systems B.V., Amsterdam, The Netherlands) and a pulse oximeter (Contec Inc., Qinhuangdao, China). On the other hand, Mohebbian et al. [82] and Riaz et al. [83] used the Raspberry Pi (Raspberry Pi Foundation, Cambridge, UK) and Arduino (Arduino, Scarmagno, Italy) microcontroller boards. Forkan et al. [78] used an accelerometer, a GPS, an ECG, and a blood pressure monitor without specifying the devices. Zhang et al. [80] used three electronic components from TX Instruments (ADS1299EEG-FE, AFE4490SPO2, MSP430F5529IPN, TX Instruments, Dallas, TX, USA). Finally, Ghosh et al. [79] used a custom-built device based on the principle of impedance plethysmography with an autoadaptive algorithm based on impedance cardiography signals.
In terms of accuracy, the best performance was 99.78% for Patient 2 by Forkan et al. [78], achieved using a decision tree J48 as the classifier. The lowest performance was 81.52%, achieved using the radial basis function in the same proposal [78]. Other works have used the mean error, root mean square error, or mean absolute error to evaluate their results. For example, Zhang et al. [80] used mean error, mean absolute error, and root mean square error for heart rate estimation. The results were 0.8 ± 2.7 beats per minute for mean error and standard deviation, 1.8 beats per minute for mean absolute error, and 2.8 beats per minute for root mean square error. Conversely, Mohebbian et al. [82] achieved 3 ± 0.7 mmHg for systolic blood pressure by root mean square error, 2.2 ± 0.7 mmHg for diastolic blood pressure by root mean square error, 4.4 ± 1.0 mmHg for systolic blood pressure by mean absolute error, and 2.9 ± 1.2 mmHg for diastolic blood pressure by mean absolute error. Yan et al. [84] used the mean absolute error for systolic and diastolic blood pressure, obtaining 0.043 mmHg and 0.011 mmHg, respectively. The mean blood pressure obtained was 0.008 mmHg. Miao et al. [85] used the mean difference ± standard deviation accuracy, which was obtained as −0.22 ± 5.82 mmHg for systolic blood pressure. The mean arterial pressure mean difference ± standard deviation accuracy obtained was −0.57 ± 4.39 mmHg. Finally, the mean difference ± standard deviation accuracy obtained was −0.75 ± 5.62 mmHg for diastolic blood pressure. Ghosh et al. [79] predicted systolic BP, diastolic BP, and heart rate accuracies of ±2.33 mmHg, ±3.60 mmHg, and ±2.88 mmHg beats per min, respectively. Table 9. IoT/IoMT-based blood pressure and hypertension detection.

Cardiovascular Disease and Heart Disease
Some proposals have not specified the disease or disorder to be detected, instead considering cardiovascular disease or heart disease as a general disease. In Table 10, we list the proposals that correspond to this classification using IoT/IoMT. There were 31 proposals corresponding to this classification. The majority used sensors or microcontrollers such as Raspberry Pi or Arduino in combination with heart rate sensors, temperature sensors, blood pressure sensors, SpO 2 sensors, or even temperature sensors. Six proposals used medical devices such as those manufactured by OMRON, Holter, Zephyr, or even Polar. Five proposals used wearable devices, but did not provide further details. There were two proposals which used a radar and a chip in combination with a mixed-signal neural network: Reservoir-computation (RC-NN) and long short-term memory (LSTM), respectively. A total of 20 proposals used accuracy as an evaluation metric, 8 of them used specificity, 4 used sensitivity, 10 used precision, 2 did not specify the evaluation metric used, while others reported median error, root mean-squared difference, the worst performance with SI score, train loss/valid loss, and 1 reported 90% less energy consumption. Clifford et al. [97] proposed an application cloud infrastructure to provide detailed information on the physical activity, behaviors, and psycho-social and physiological status of urban African American young adults without providing further details.
The best accuracy achieved was 99.58% by Demirel, Bayoumy, and Faruque [110]. Conversely, the lowest accuracy was 60.98%, obtained by Moradkhani, Broumandnia, and Mirabedini [105] using a probabilistic neural network. In terms of sensitivity and specificity, the best performance was that of Demirel, Bayoumy, and Faruque [110], who used a convolutional neural network and obtained 99.2% and 99.4%, respectively. The lowest performance in terms of sensitivity was 45.5%, obtained by Boursalie, Samavi, and Doyle [88] using a support vector machine classifier. The lowest specificity was 55.1%, again obtained by Moradkhani, Broumandnia, and Mirabedini [105].

Others
In this subsection, we group the proposals that focused on other diseases of lesser frequency (i.e., chronic heart failure, myocardial infarction, coronary artery disease, carotid artery, saturated oxygen, stroke disease, ECG abnormalities, and ECG noise). In Table 11, we list the proposals that correspond to different diseases using IoT/IoMT.
The best performance in accuracy (99.8%) was achieved by Yu et al. [118], using a deep neural network. The worst was 97.74%, achieved by Aranki et al. [111] in combination with an SVM classifier. The rest of the proposals reported precision, sensitivity, specificity, recall, F1-score, and root mean square in combination with classifiers such as SVM, random forest, long short-term memory, deep neural network, artificial neural network, an R-peak detection algorithm, and ResNet-9 semi-supervised learning. Sahani et al. [119] reported a reduction of the data dropout rate (21.09%) and an increment in the number of R-peak detections (15.33%).
The most commonly detected conditions were arrhythmia (38.82%) and cardiovascular disease (29.41%). Conversely, the fewest were cardiomyopathy (including hypertrophic car-diomyopathy), ischemic heart disease, valvular heart disease, left ventricular hypertrophy, chronic heart failure, and stroke, classified in the 8.23% of other diseases. Table 13 summarizes the abnormality and arrhythmia detection proposals using only machine learning approaches. There were 33 proposals, 26 of which used the MIT-BIH or PhysioNet data sets. Other data sets used were privately provided by hospitals or health centers. On the other hand, the machine learning approaches that were most frequently used were neural networks such as convolutional neural networks, long short-term memory, deep neural networks, bidirectional long short-term memory, lead convolutional neural networks, deep residual neural networks, densely connected convolutional networks, multiscale fusion convolutional neural networks, and one-dimensional neural networks. Other techniques included SVM, kNN, random forest, genetic algorithms, bacterial-foraging optimization, and particle swarm optimization.

Abnormality Detection and Arrhythmia
In terms of evaluation metrics, some of the higher values reported were achieved with neural networks. Haleem et al. [142], for example, achieved 100% accuracy for congestive heart failure events and sudden cardiac deaths. The rest of the proposals reported accuracy in the range of 98-99%. Other metrics reported included F1-score, precision, recall/sensitivity, specificity, AUC, positive productivity, and positive predictive value. Focusing on sensitivity and specificity values, the best performance for the former was 99.8%, obtained by Ma et al. [155], while the best specificity was 99.7%, obtained by Tung et al. [149]; both proposals used convolutional neural networks.
In general, the results ranged between 81% and 99%. However, Dias et al. [141] and Li, Qian, and Li [151] reported lower precision values: 36.8% and 65.88%, respectively. Li, Qian, and Li [151] also reported a lower sensitivity (35.22%). Both proposals achieved lower results in these metrics for supraventricular segment or ectopic beat.    Table 14 shows a summary of the proposals for blood pressure and hypertension detection using only machine learning approaches. Ten proposals were analyzed in this category for detecting blood pressure or hypertension. Nine proposals used the MIT-BIH data set, while only one proposal used the UCI data set. The more frequently used machine learning techniques included recurrent neural networks, convolutional neural networks, bi-directional long short-term memory, long short-term memory, multi-scale fusion neural networks, deep learning, artificial neural networks, and decision trees. In terms of results, only three proposals included accuracy as a metric: Alkhodari et al. [158], Zhang et al. [163], and Kim et al. [167]. Alkhodari et al. [158] and Kim et al. [167] reported an accuracy above 91%, while Zhang et al. [80] reported an accuracy for systolic blood pressure in normal, atrial fibrillation, and coronary arteriosclerosis subjects. Other proposals focused on MAE, RMSE, and mean absolute error or standard deviation metrics. The majority reported results for systolic and diastolic blood pressure, but Alkhodari et al. [158]; Landry, Peterson, and Arami [159]; Saleh et al. [160]; and Mahmud et al. [166] reported only general results. In terms of accuracy, precision, F1-score, sensitivity, and specificity metrics, Alkhodari et al. [158] obtained a better accuracy (97.08%) than Kim et al. [167] (91.44% and 94.66% for systolic and diastolic BP); however, Kim et al. [167] achieved a better performance in terms of sensitivity and specificity (above 94% for specificity and above 70.17%for sensitivity).

Cardiovascular Disease/Heart Disease
In Table 15, we list the proposals that classified a common disease (cardiovascular disease or heart rate disease) using only machine learning approaches. There were 25 proposals, 14 of which used the PhysioNet and MIT-BIH data sets, 12 used the UCI data set, 5 used private data sets, and 2 used the Framingham data set. Many machine learning approaches were used to classify cardiovascular disease, including SVM, Naïve Bayes, random forest, multi-layer perceptron, hybrid random forest, XGBoost, convolutional neural network, long short-term memory, particle swarm optimization, neuro-fuzzy inference system, deep neural network, beetle swarm optimization, twin SVM, deep neural net-work, neural network, artificial neural network, decision tree, bidirectional long short-term memory, sine cosine kNN, gradient boost, deep learning, growing multi-layer network, and other approaches for big data processing, ensemble-based feature selection, and recursive feature elimination.
The most common metrics used for evaluation were accuracy, precision, recall, F1-score, sensitivity, specificity, AUC, MAE, and RMSE. The positive predictive value, the estimation error, and the equal rate of error was also used to evaluate the proposals. The best accuracy achieved was 99.45%, in combination with a modified salp swarm optimization-adaptative neuro fuzzy inference system proposed by Khan and Algarni [45]; they also achieved good precision (96.54%). Deepika and Balaji [182] used a multi-layer perceptron to obtain good performance in terms of accuracy, sensitivity, F1-score, specificity, and Kappa (all of them above 96.40%). The two proposals which reported lower performance were those of Theerthagiri [181] and Stankovic et al. [185], achieving a performance above 83% but lower than 89% in accuracy, precision, F1-score, and recall. Theerthagiri [181] also obtained lower performance in AUC, MSE, and RMSE, compared to other proposals.

Myocardial Infarction
A myocardial infarction (MI) is characterized by a typical rise and/or fall in cardiac troponin, with at least one value above the upper reference limit of the assay, and at least one other characteristic associated with ischemia. Also known as myocardial infarction, it occurs when the flow of oxygenated blood to a segment of the heart muscle is suddenly interrupted and the heart is deprived of oxygen. The segment of heart muscle begins to die if blood flow is not immediately restored [206]. Table 16 lists the six proposals that classified myocardial infarction using only machine learning approaches. Three of them used convolutional neural networks, one a long short-term memory, and two used traditional approaches: SVM and random forest. The most-used data sets were the Physikalisch-Technische Bundesanstal and PhysioNet data sets. Xiao et al. [192] used medical records retrieved from a hospital, but obtained lower performance in terms of accuracy and AUC (82% and 85%, respectively). The best performance was obtained by Fatimah et al. [193] in combination with kNN, where the accuracy, sensitivity, and specificity obtained were all above 99.94%. Meanwhile, Deng et al. [190] used a convolutional neural network and achieved good performance (above 98.1%), and Ibrahim; et al. [191] obtained a performance above 93.4% with an XGBoost classifier.

Coronary Artery Disease/Coronary Heart Disease
Coronary heart disease is a condition in which the heart muscle-or myocardiumdoes not receive enough oxygen as the coronary arteries do not have an adequate blood supply. Coronary artery obstruction occurs when the arteries become stiff and narrow due to the build-up of fatty deposits (plaques). These fatty deposits are mainly composed of cholesterol and fibrin and, when these deposits predominate, it is called atherosclerosis [207]. Table 17 summarized the proposals for coronary heart disease detection using only machine learning approaches. There were four proposals, each of which considered different approaches: logistic regression, elastic net, SVM, random forest, XGBoost, multi-layer perceptron, k-NN, bagging, binary logistic classification, Naïve Bayes, boosting, and random forest. The data sets also were from different sources, including Framingham, PhysioNet, and a cardiovascular disease data set. In terms of evaluation metrics, the most used were AUC, accuracy, sensitivity, and specificity. The best performance in terms of sensitivity was 100%, obtained by Dash et al. [198]. Meanwhile, Masih, Naz, and Ahuja [196] achieved better accuracy and specificity: 96.50% and 98.28%, respectively.

Others
In this section, diverse proposals from different diseases are grouped into the same table (Table 18), for a total of seven proposals. For example, Dai, Hwang, and Tseng [202] focused on classifying cardiomyopathy; Smole et al. [204] considered hypertrophic cardiomyopathy; Elias et al. [200] considered valvular heart disease; Duffy et al. [201] considered left ventricular hypertrophy; Loeffler and Starobin [203] considered ischemic heart disease; Li et al. [199] considered chronic heart disease; and Pathan et al. [205] considered stroke.
The most-used machine learning techniques were convolutional neural networks (1D, 3D, deep convolutional). Other proposals used more traditional techniques, such as random forest, SVM, and boosted trees. Elias et al. [200] used AUC-ROC as an evaluation metric, obtaining 88% for detecting aortic stenosis. Meanwhile, Duffy et al. [201] reported an AUC of 98% for hypertrophic cardiomyopathy. Dai, Hwang, and Tseng [202] achieved 99.84% accuracy-the highest of the seven proposals. Pathan et al. [205] used a support vector machine classifier and obtained the precision, sensitivity, F1 score, and accuracy of each attribute in the Mckinsey data set (i.e., gender, age, hypertension, heart disease, ever married, work type, residence, glucose level, BMI, and smoking status).

Results and Discussion
As a result of this systematic review, 162 articles were analyzed and classified into two categories: those using IoT/IoMT technologies and those applying machine learning techniques. We obtained 78 proposals for the first category and 84 for the second. In each of these categories, we evaluated the articles according to six parameters: Study area, disease, data set used, wearable device/smart device/medical device, approach, and outcomes. We utilized these categories to answer the research questions presented in Section 3.1. For RQ1, we collected information about wearable devices/smart devices/medical devices; for RQ2, we considered the data set used, approach, and results; and, finally, for RQ3, we focused on the predicted disease.
Our reading of the 162 papers revealed that several devices have been used to monitor or detect CVD, including smartphones, microcontroller boards with sensors, and even experimental devices built specifically for CVD. We also found that neural networks were one of the most commonly used machine learning approaches, including convolutional neural networks, long short-term memory, bidirectional long short-term memory, multilayer perceptrons, and deep neural networks.
Unlike other works, our comprehensive review details the machine learning methods used in conjunction with wearable/smart devices/medical devices and data sets for the detection, prediction, and/or monitoring of CVD, including the most-researched disease types (i.e., abnormality detection/arrhythmia, aortic stenosis, arterial stiffness, atrial fibrillation, blood pressure/hypertension, heart disease in general, chronic heart failure, myocardial infarction, coronary heart disease and stroke), in a single work. We place focus on the technology used, presented in the form of wearable devices/smart devices/medical devices and machine learning approaches for the classification of CVD, rather than a specific disease. Our results indicate that CVD/heart disease in general was the disease most commonly captured by wearable devices/smart devices/medical devices. In contrast, abnormality detection/arrhythmia was the disease most focused on when applying machine learning methods, thanks to the public MIT-BIH data set. We found that there is a need for more public CVD data sets, as the MIT-BIH and Cleveland-UCI public data sets were the most widely used. On the other hand, we also found that neural networks were the most-used approach, while other approaches included traditional machine learning methods, statistical approaches, and optimization techniques.
We now summarize the main results presented in Sections 4 and 5, with the aim of answering the research questions introduced in Section 3. The discussion is divided into three parts, related to the detection, prediction, or monitoring of CVD using IoT/IoMT technology or machine learning techniques.

Research Question 1
RQ1: What types of devices with IoT and IoMT technologies have been used to detect and predict cardiovascular disease using machine learning? was focused on discovering the technologies and wearable devices/smart devices/medical devices used to detect or monitor CVD in real-time as preventive care.
The technologies discovered in the systematic review included medical devices, smartphones, microcontroller boards, sensors, smartwatches, radar, and wearable devices. These technologies were combined with frameworks, applications, and systems to monitor or detect CVD. Some technologies used the cloud to store and process the data obtained from the devices. In these cases, data privacy and security are also important, considering the importance of patient confidentiality. The devices and the system need to be secure and reliable, as they manage medical and physiological data. Preferably, the devices must be FDA-approved to avoid significant measurement errors; however, in monitoring cases, some patients may not be able to afford medical devices and a framework including commercial wearable devices needs to be developed. This is especially important for rural communities which, in some cases, do not have easy access to health services. In this case, the systems must be responsive to end-user requests for data extraction, downloading, and consultation. They must also be non-intrusive and easy to wear (body placement).
After analyzing the 162 papers selected as relevant, we observed an increase in relevant research output starting around 2019. These proposals started to use IoT/IoMT devices more frequently for the monitoring of chronic diseases in real-time, with promising results obtained after the COVID pandemic. Some of the proposals included in the review used IoT/IoMT technologies in combination with clinical decision support systems. The use of these systems can help clinicians to assess the risk of heart disease and provide treatments to further manage risk. Implementing models into decision support systems can improve preventive care, allow for the collection and analysis of data in real-time, and reduce the likelihood of misdiagnosis. The proposed models uncovered in this systematic review were developed to provide high-performance predictions of the presence or absence of heart disease, given the current condition of the participants.
Other proposals used frameworks in combination with IoT/IoMT technologies. Khan [43] proposed modules in a prediction system that integrates hardware devices, microcontrollers, and LoRa communication hardware to transmit data to the cloud system. Forkan et al. [78] presented a generalized framework for personalized healthcare connected to a personal cloud server. The local server only collects low-level data (e.g., ECG data, blood pressure monitor data, accelerometer data) from the ambient assisted living system, then forwards them directly to the context aggregator or the personal cloud server.
Microcontrollers were one of the most common devices used for the detection or prediction of CVD. The two most common microcontrollers were Arduino and Raspberry Pi, in combination with various sensors (e.g., photopleytmography, ECG, heart rate, temperature, electroencephalagram, SpO 2 , respiration, EMG). Some authors have also developed healthcare monitoring systems presenting good results in the evaluation metrics. In such instances, the medical devices performed well in the evaluation process, such as Holter devices, biomedical sensors, and Omron, achieving an accuracy of over 94%. To the contrary, sensors and microcontroller boards only achieved an accuracy above 72%. Other devices, such as the Lenovo Smart ECG Vest, achieved an accuracy above 86%. In terms of specificity, the Kardiamobile and SmartCardia devices achieved 74.9% and 78.82%, respectively. The Zigbee device achieved an accuracy of 92.11%, while the SmartCardia NYU device achieved accuracy, precision, specificity, F1-score, and sensitivity values above 90%.
The most convenient device we detected for the monitoring of CVD were smartphones, being a non-intrusive device that most of people have access to. Unfortunately, these devices do not provide good measurements, when compared to medical devices, and wrong measurements can lead to mortally serious outcomes. Ideally, only medical-grade devices should be used to monitor chronic diseases such as CVD, precisely due to the importance of acting in a timely manner in the event of an emergency. People who monitor their health on a daily basis can survive if the system alerts them (or their caregivers) to an impending stroke or heart attack. Devices that use sensors or microcontrollers provide a good solution for an unobtrusive device, but a huge amount of testing is necessary to approve the reliability of their measurements. The goal of any device that monitors a chronic disease is to prevent an incoming emergency and send alerts to a caregiver, relatives, and even hospitals, in order to prevent a loss of life.

Research Question 2
RQ2: What machine learning techniques have been applied to detect and predict cardiovascular disease? was focused on determining the machine learning approaches used to classify, predict, and/or analyze CVD data, in order to compare the approaches as well as the evaluation metrics and their results.
Machine learning techniques allow for the exploration of an immense amount of data, feeding a computer algorithm to analyze the input data. As such, they allow software applications to predict diseases. As described in Section 5, many machine learning techniques have been used to detect CVD. Some of the most popular approaches were neural networks and classifiers such as random forest, XGBoost, k nearest neighbors, or support vector machine. Other proposals have used a combination of techniques, such as a modified salp swarm optimization-adaptive neuro-fuzzy inference system (MSSO-ANFIS) and a beetle swarm optimization-adaptive neuro-fuzzy inference system (BSO-ANFIS). It was found that the proposals using a combination of machine learning approaches generally obtained good performance in the accuracy metric, over 96%. Meanwhile, those utilizing neural networks (deep neural networks, multi-layer perceptrons, convolutional networks, long short-term memory) obtained accuracy values above 90%. Some recent algorithms have been applied to CVD detection; for example, particle swarm optimization has been used in two proposals-those of Dang et al. [133]-and obtained an accuracy of above 95%.
Other approaches, such as linear-kernel SVM and hybrid random forest with a linear model, have been used for CVD detection and achieved good performance in terms of evaluation metrics. Linear kernel SVM achieved specificity and sensitivity of 99.6% and 79.3%, respectively, while the hybrid random forest achieved an accuracy of 88.7%. The proposals that used an SVM as a machine learning classifier obtained an accuracy above 88%.
We noticed that the best performance in evaluation metrics was obtained with a combination of techniques and neural networks. The most popular classifiers (Random Forest, SVM, XGBoost) achieved good performance. Some of the neural networks used included deep neural networks, convolutional neural networks, long short-term memory, multi-scale fusion convolutional neural networks, end-to-end deep multi-scale fusion convolutional neural networks, lead convolutional neural networks, recurrent neural networks, bidirectional long short-term memory, deep learning-modified neural networks, and convolutional neural network-long short-term memory. The most widely used neural networks were the convolutional neural network and long short-term memory.
In general, deep learning using neural networks, as some of the most-used techniques in the state-of-the-art, obtained good results. However, some of the proposals solely focused on the accuracy metric, describing how the model performs across all classes without focusing on other metrics such as sensitivity, specificity, precision, or F1-score. Some algorithms based on the behavior of animals were used to detect or predict CVD, such as the dragonfly algorithm used by Deepika and Bajaji [182], which obtained 97.47% accuracy, 98.92% sensitivity, 96.45% F1-score, 96.47% specificity, and 96.75% kappa. Another example is the study of Singh et al. [172] who obtained 99.91% accuracy, 99.37% precision, 99.4% specificity, and 99.21% sensitivity for heart disease classification using a beetle swarm optimizationadaptive neuro-fuzzy inference system. Raj [171] also used an animal-inspired algorithm in the modified salp swarm optimization-adaptative neuro-fuzzy inference system to obtain 99.45% accuracy and 96.54% precision. Khan and Algarni [45] used the biologically inspired particle swarm optimization algorithm to achieve 99% accuracy. Similarly, Kora, Abraham, and Meenakshi [132] used the particle swarm optimization in addition to a bacterial-foraging optimization to obtain 99.1% accuracy for atrial fibrillation classification applying wavelet transform, 98.9% accuracy for myocardial infarction using SVM classifier, and 99.3% accuracy for bundle branch block using the same classifier.

Research Question 3
RQ3: What types of diseases have been detected and predicted? was focused on gathering the cardiovascular diseases that are more reliable to monitor or detect through the use of devices/wearable devices/smart devices/medical devices. Likewise, we also aimed to determine the diseases that are hardly monitored using IoT/IoMT technologies, as an opportunity to develop real-time systems or applications.
The most frequently detected diseases using IoT/IoMT technologies were cardiovascular diseases in general, arrhythmia, and blood pressure/hypertension. On the other hand, the most common conditions detected by machine learning alone were arrhythmia, cardiovascular disease in general, and blood pressure. A total of 59 proposals focused on CVD in terms of arrhythmia, 55 on CVD, 18 on blood pressure/hypertension, 8 on myocardial infarction, 6 on coronary artery disease, 3 on chronic heart disease, 3 on aortic stenosis, 3 on arterial disease, 2 on stroke, 2 on cardiomyopathy, 1 on valvular heart disease, 1 on left ventricular hypertrophy, 1 on ischemic heart disease, 1 on saturated oxygen, and 1 on carotid disease.
We noticed that some proposals did not specify the detected disease and considered cardiovascular disease as a general disease, which was one of the most-detected diseases using IoT/IoMT technologies and using only machine learning. However, these proposals utilized public data sets, such as UCI Cleveland and PhysioNet MIT-BIH. The less common diseases detected included chronic heart disease, myocardial infarction, coronary artery disease, stroke, carotid disease, ischemic heart disease, left ventricular hypertrophy, cardiomyopathy, and valvular heart disease.
A significant number of proposals found, detected, and/or predicted CVD, arrhythmia, and hypertension, but there is an opportunity to detect specific diseases, such as stroke or myocardial infarction, that may be useful to predict in the context of healthcare systems.
Other observations from this systematic review are as follows. Most of the data sets used were from the UCI and PhysioNet sites, while data sets from hospitals and private companies were included in some cases. The devices used to monitor or diagnose CVD ranged from medical devices to smart watches and microcontroller cards. Finally, some devices were developed for specific diseases and evaluated in specific groups (including control groups).

Future Trends
Over the past few years, several trends in Internet of Things (IoT) and Internet of Medical Things (IoMT) technologies for CVD classification and prediction have emerged. These trends include clinical decision support systems, wearable devices, data analytics, data security, remote patient monitoring, and risk assessment combined with predictive analytics.
A clinical decision support system is a computer-based tool designed to assist healthcare providers in making clinical decisions by providing relevant information, knowledge, and recommendations at the point of care. One of their benefits is providing healthcare providers with real-time insights and recommendations for the diagnosis and management of CVD. These systems require the development of intelligent algorithms and decision support tools to help healthcare professionals make accurate and timely decisions. They also integrate patient-specific data from multiple sources, such as electronic health records (EHRs), medical imaging systems, laboratory results, and IoT/IoMT devices. Such integration enables the system to easily provide a comprehensive view of a patient's health information at any time and provide treatment recommendations based on established guidelines and individual patient characteristics. It can suggest potential diagnoses, differential diagnoses, help to rule out certain conditions, and provide appropriate medications, dosages, treatment plans, and lifestyle changes for the management of CVD by analyzing patient data and symptoms. The key benefit of such a system is that they can generate alerts and reminders for healthcare providers, notifying them of critical information such as abnormal test results, drug interactions, or upcoming preventive screenings. These reminders help to ensure timely intervention and adherence to clinical guidelines. It can also help to monitor patient progress and provide follow-up recommendations. Remote patient monitoring can track vital signs, lab results, and IoT/IoMT data to assess treatment effectiveness and suggest adjustments as needed. However, the wearable devices, smart devices, or sensors integrated into microcontroller boards must be reliable and highly accurate-and preferably FDA-approved-as the vital signs of patients can change rapidly in the context of chronic disease, and a major change may even lead to death. Future studies may focus on improving the accuracy and reliability of these devices, as well as integrating them with IoT/IoMT platforms for real-time data analysis and the development of early-warning systems.
One of the major benefits of integrating wearable devices into clinical decision support systems is the ability to monitor patients at home without the need for intrusive medical devices and hospitalization; however, the main drawback is the amount of data generated by IoT/IoMT technologies. Researchers need to thoroughly explore the use of advanced data analytics techniques such as machine learning, deep learning, and artificial intelligence to analyze the large amounts of data collected from IoT/IoMT devices. In terms of data analytics, researchers can explore the development of predictive models that use historical patient data, genetic information, lifestyle factors, and IoT/IoMT data to assess an individual's risk of developing CVD. These models could help to identify high-risk individuals for targeted interventions and preventive measures. As these techniques are expected to help in identifying patterns, correlations, and predictive models for the early detection and accurate classification of CVD, they also involve the collection and transmission of sensitive patient data. Therefore, future studies should also focus on developing robust security measures and privacy frameworks to protect data from unauthorized access and ensure compliance with privacy regulations.
Granular computing based on IoT/IoMT technologies has recently emerged, along with the advent of machine learning gaining significant attention in medical studies, with the potential to revolutionize various aspects of healthcare, including the diagnosis, treatment, and management of CVD. It allows data to be aggregated and grouped into granules that can be analyzed more efficiently and effectively. Granular computing can be viewed as a unified framework of theories and methods that utilize granularity in the problem-solving process. Granularity leads to information compression. Therefore, computing with granules instead of objects leads to faster computation times, making granular computing an important aspect of knowledge discovery and data mining [208].
Granular computing enables the representation of complex and heterogeneous CVDrelated data collected from various IoT/IoMT devices, such as wearable sensors, remote monitoring systems, and health records. It also facilitates the fusion of multiple data sources in the IoT/IoMT ecosystem, such as physiological parameters, medical history, lifestyle factors, and environmental data. Granular computing can help to identify hidden patterns, correlations, and dependencies relevant to CVD through data clustering for medical data classification. It can provide great support in feature selection and extraction from the large amounts of data generated, improving the accuracy and efficiency of CVD classification and prediction models by reducing the dimensionality of the data and identifying the most informative features. These techniques support the development of algorithms that generate rules and decision models based on the extracted granularity, in order to provide insight into the relationships between CVD risk factors, symptoms, and outcomes. Granular computing enables continuous monitoring to assess the risk of developing CVD by analyzing granular data and identifying high-risk individuals. Continuous monitoring takes into account various factors such as age, gender, genetic markers, lifestyle habits, and physiological parameters. It provides personalized risk assessments that lead to personalized treatment and intervention strategies for individuals with CVD, taking into account individual patient characteristics. This helps to optimize treatment plans, recommend appropriate interventions, and support shared decision-making between patients and healthcare providers. It also allows for the analysis of large volumes of electronic health records and the classification of patients into different disease categories. CVD risk prediction enables targeted interventions and preventive measures to reduce the burden of CVD based on factors such as demographics, medical history, lifestyle data, genetic markers, and biomarkers.
In contrast, granular computing has been used in several machine learning approaches, such as disease diagnosis, risk prediction, treatment adaptation, prognosis detection, image analysis, decision support, precision medicine, biomarker discovery, and clinical trial design. Another approach is outlier detection, an important process for dealing with the outliers in a data set. More importantly, in disease prediction, the outliers are sometimes the most valuable data, as they indicate the values of those individuals with the disease.
With the continuous growth of technology, it is possible to classify, monitor, and predict chronic diseases, including CVD, in real-time and with high accuracy through the use of sensors and devices connected to the cloud. Recent proposals have focused on obtaining high accuracy with ensembles of algorithms or even new approaches to deal with the large amount of data generated; however, some of them have not placed an emphasis on privacy. They can classify and monitor vital signs, predict major risks, and send reminders or emergency alerts to healthcare providers, but they cannot handle the privacy of the sensitive and confidential data generated by IoT/IoMT devices in real-time. The data generated by IoT/IoMT devices contain details about an individual's medical conditions, treatments, test results, and other personal identifiers. Privacy ensures that this information remains confidential and accessible only to authorized individuals. If the clinical decision support system can guarantee the privacy and confidentiality of the data generated, it will be a great tool for improving the remote monitoring of chronic diseases and helping to reduce the risk of death in even remote areas, where health care is expensive or not easily accessible, as will highly accurate devices or sensors for the tracking of vital signs.
Limitations of this study: First, we conducted a literature search in only four databases (PubMed, IEEE, Springer Link, and Science Direct). Second, we excluded studies that were not written in English, written before January 2016 and after May 2023, and articles other than JCR.

Conclusions
In this study, we examined the IoT/IoMT technologies and machine learning techniques used to detect, predict, or monitor CVD. We also identified several diseases that have been commonly detected or monitored. We classified 162 proposals obtained from the IEEE, PubMed, Science Direct, and Springer Link databases into the following categories: CVD detection using IoT/IoMT and CVD detection applying machine learning techniques. We found 78 articles in the former category and 84 in the latter. For these two categories, we extracted the following information: Study area, disease, data set, wearable device/smart device/medical device, machine learning approach, and results.
We noticed that the technologies and the machine learning approaches used in the proposals are continuously evolving, especially in terms of mobile systems, web applications, and frameworks. Some of the articles proposed a combination of techniques to achieve reliable performance in evaluation metrics (i.e., accuracy, precision, sensitivity, specificity, F1-score). Others used medical devices or commercial devices for the real-time monitoring of CVD. In these cases, we found that medical devices were a better recommendation for monitoring or detecting CVD due to their reliability.
Our review also revealed the type of diseases detected by wearable devices/smart devices/medical devices or machine learning approaches. The most-used wearable devices were medical devices, sensors, and microcontrollers. In some proposals, CVD or heart failure was detected as a common disease. In others, arrhythmia was the most commonly detected disease. The most-used machine learning or deep learning techniques were neural networks, including long short-term memory, convolutional neural networks, recurrent neural networks, artificial neural networks, bi-directional long short-term memory, and deep neural networks. The most used traditional classifiers included kNN, SVM, and random forest. However, after analyzing the results reported by the authors regarding the evaluation metrics, it is possible that some of the models were subject to over-fitting, due to the lack of data used for training and testing. We also identified the data sets most commonly used to train the models: UCI Cleveland, PhysioNet MIT-BIH, PhysioNet MIMIC-II, and Framingham. However, other proposals also used private data sets provided by hospitals.
The most important finding is the lack of public data sets focused on CVD. As more public CVD data sets become available, relevant machine learning and deep learning models could gain improved performance metrics, increasing the possibility of timely prediction of risk situations or even myocardial infarction. For example, after the COVID-19 pandemic, the demand for IoT/IoMT sensor devices in healthcare increased for the treatment and monitoring of critical patients, leading to better and more complete systems and frameworks for remotely monitoring chronic patients. Some commercial devices have recently begun to add features such as temperature sensing, measuring blood oxygen, and taking an ECG at any moment, allowing for caregivers to be alerted to any emergency detected. This leads to more smart options for the tracking of diseases.
From the results collected in this article, it seems clear that the detection and monitoring of CVD is possible. With the help of IoT/IoMT technology, it has recently become possible to monitor CVD in real-time and even alert caregivers in the case of an emergency. Thus, it is possible to triage the condition of the patient, recommend the nearest hospitals in the case of emergency, and send notifications to doctors and caregivers.
Finally, as the leading cause of death worldwide, it is of great importance to propose new machine learning approaches or better methods focused on CVD, in order to achieve good performance in terms of evaluation metrics and facilitate its real-time prediction.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.

Conflicts of Interest:
The authors declare no conflicts of interest.

Abbreviations
The following abbreviations are used in this manuscript:

Appendix A
In the Table A1, we present the 68 potential studies that answered the research questions described in Section 3, sorted by date.     [136] Improving the safety of atrial fibrillation monitoring systems through human verification 2017 11 [137] Validating the robustness of an internet of things based atrial fibrillation detection system 2020 20 [138] Extracting deep features from short ECG signals for early atrial fibrillation detection 2020 22 [139] Detection of Atrial Fibrillation in Compressively Sensed Electrocardiogram Measurements 2020 15 [140] A Multi-tier Deep Learning Model for Arrhythmia Detection 2020 115 [141] Arrhythmia classification from single-lead ECG signals using the inter-patient paradigm 2021 50 [142] Time adaptive ECG driven cardiovascular disease detector 2021 21 [143] ECG arrhythmia classification by using a recurrence plot and convolutional neural network 2021 91 [144] Stacking segment-based CNN with SVM for recognition of atrial fibrillation from single-lead ECG recordings 2021 33 [145] Exploring Deep Features and ECG Attributes to Detect Cardiac Rhythm Classes 2021 91 [146] Automated ECG multi-class classification system based on combining deep learning features with HRV and ECG measures 2022 17 [147] Deep learning-based multidimensional feature fusion for classification of ECG arrhythmia 2021 15 [148] ECG signal classification to detect heart arrhythmia using ELM and CNN 2022 - [149] Multi-Lead ECG Classification via an Information-Based Attention Convolutional Neural Network 2020 5 [150] Fuzz-ClustNet: Coupled fuzzy clustering and deep neural networks for Arrhythmia detection from ECG signals 2023 - [151] Inter-patient arrhythmia classification with improved deep residual convolutional neural network 2022 21 [152] DeepArr: An investigative tool for arrhythmia detection using a contextual deep neural network from electrocardiograms (ECG) signals 2023 - [153] Two-stage detection method of supraventricular and ventricular ectopic beats based on sequential artificial features and heartbeats 2023 - [154] Automatic varied-length ECG classification using a lightweight DenseNet model 2023 - [155] Arrhythmia detection based on multi-scale fusion of hybrid deep models from single lead ECG recordings: A multicenter dataset study 2022 2 [151] A deep learning approach to cardiovascular disease classification using empirical mode decomposition for ECG feature extraction 2023 1 [157] Prediction of paroxysmal atrial fibrillation using new heart rate variability features 2021 23 [158] Predicting  [161] Cuffless blood pressure estimation based on composite neural network and graphics information 2021 10 [162] PPG-based blood pressure estimation can benefit from scalable multi-scale fusion neural networks and multi-task learning 2022 5  [168] An IoT based efficient hybrid recommender system for cardiovascular disease 2019 75 [37] Effective Heart Disease Prediction Using Hybrid Machine Learning Techniques 2019 888 [169] An efficient IoT based patient monitoring and heart disease prediction system using Deep learning modified neural network 2020 92 [170] HDPM: An Effective Heart Disease Prediction Model for a Clinical Decision Support System 2020 130 [45] A Healthcare Monitoring System for the Diagnosis of Heart Disease in the IoMT Cloud Environment Using MSSO-ANFIS 2020 153 [46] Comprenhensive electrocardiographic diagnosis based on deep learning 2020 146 [171] An Efficient IoT-Based Platform for Remote Real-Time Cardiac Activity Monitoring 2020 35 [172] Multi-disease big data analysis using beetle swarm optimization and an adaptive neuro-fuzzy inference system 2021 17 [173] Cardiac disease detection using cuckoo search enabled deep belief network 2022 2 [174] Automatic diagnosis of cardiovascular disorders by sub images of the ECG signal using multi-feature extraction methods and randomized neural network 2021 20 [175] An intelligent heart disease prediction system based on swarm artificial neural network 2021 23 [176] A multi-label learning prediction model for heart failure in patients with atrial fibrillation based on expert knowledge of disease duration 2023 - [177] WoM-based deep BiLSTM: smart disease prediction model using WoM-based deep BiLSTM classifier 2023 - [178] A scalable and real-time system for disease prediction using big data processing 2023 3 [179] A novel blockchain-enabled heart disease prediction mechanism using machine learning 2022 15 [180] Heart rate estimation network from facial videos using spatiotemporal feature image 2022 2 [181] Predictive analysis of cardiovascular disease using gradient boosting based learning and recursive feature elimination technique 2022 3 [182] Effective heart disease prediction using novel MLP-EBMDA approach 2022 22 [183] LightX3ECG  [190] ST-Net: Synthetic ECG tracing for diagnosing various cardiovascular diseases 2020 6 [191] Explainable Prediction of Acute Myocardial Infarction Using Machine Learning and Shapley Values 2020 47 [192] Prognostic Value of Machine Learning in Patients with Acute Myocardial Infarction 2022 8 [193] Efficient detection of myocardial infarction from single lead ECG signal 2021 21 [194] Near real-time single-beat myocardial infarction detection from single-lead electrocardiogram using Long Short-Term Memory Neural Network 2021 10 [195] Predicting cardiac disease from interactions of simultaneously-acquired hemodynamic and cardiac signals 2021 8 Table A1. Cont.

Ref. Title
Year Cites [196] Multilayer perceptron based deep neural network for early detection of coronary heart disease 2021 17 [197] Early Detection of Coronary Heart Disease using Ensemble Techniques 2021 37 [198] Non-invasive detection of coronary artery disease from photoplethysmograph using lumped parameter modelling 2022 2 [199] [203] Reaction-diffusion informed approach to determine myocardial ischemia using stochastic in-silico ECGs and CNNs 2021 3 [204] A machine learning-based risk stratification model for ventricular tachycardia and heart failure in hypertrophic cardiomyopathy 2021 19 [205] Identifying Stroke Indicators Using Rough Sets 2020 18

Appendix B
In the Table A2 we depict the 68 Published articles sorted by journal.