According to the European society of cardiology (ESC), nearly 26 million people have been diagnosed with HF worldwide [14, 15, 16]. Furthermore, 3%–5% of hospital admissions have been attributed to HF incidents. Hence, early detection of the disease is of vital importance in shaping the medical and surgical interventions to reduce the high rates of mortality and morbidity. AI has been reported to correct any existing medication errors, and hence can be effectively utilized as an ‘assisting resource’ on which clinicians can rely, and use in daily clinical practice [17]. Furthermore, AI and human predictive skills have been compared in previous literature. American Heart Association (AHA) and American College of Cardiology (ACC) provide guidelines on primary prevention of cardiovascular disease using risk factors such as nutrition, obesity, physical factors, diabetes, and lipid profile. In a new study, ACC/AHA guidelines were compared with four machine-learning algorithms: random forest, logistic regression, gradient boosting, and neural networks. The results not only indicated NN to predict 7.6% more events than ACC/AHA criteria but also took into account twenty-two more data points including ethnicity, kidney disease, and arthritis which are not part of AHA/ACC guidelines thus demonstrating AI to be of great value in predicting risks [18]. Another study applied NN for the diagnosis of HF on 40 individuals with the dataset comprising age, gender, blood pressure, and smoking history to determine predictors for HF. The model proved to be highly accurate with 85% of the results correctly predicted [19]. The first-line investigation performed in suspected heart problems is usually ECG, after which the diagnosis gets narrowed down, and if HF is suspected, the physicians investigate for cardiac biomarkers such as natriuretic peptides. However, natriuretic peptides are not very specific markers for HF. Instead, they are subjective to factors such as obesity, age, kidney disease amongst others [15, 20, 21]. Heart failure with preserved ejection fraction (HFpEF) is a rapidly emerging global health issue with its prevalence increasing at a rate of 1% per year relative to heart failure with reduced ejection fraction (HFrEF) owing to risk factors including age, gender, weight, and hypertension.

HFpEF is a complex syndrome resulting from structural and functional cardiac disorders rather than one disease entity hence making it difficult to diagnose by clinicians treating HF patients. Kwon et al. [21] established an adequate tool to screen it reliably and economically. The study developed and validated a DL model (DLM) based on an ensemble NN using 12-, 6-, and single-lead ECG that demonstrated reasonable performance. The DLM was further visualized to determine the characteristics and regions of ECG that were used for HFpEF prediction and confirmed the important variable for the decision in other ML models based on logistic regression (LR), random forest (RF), and convolutional neural network (CNN). A sensitivity map was used to identify where DLM focused on primarily, and it was found to have focused on the R wave in the QRS complex as well as the T wave [21]. The study further showed that in the subgroup of 1412 patients without HFpEF at initial echocardiography, 246 patients developed HFpEF within 24 months. Patients whose DLM was defined as having a higher risk had a significantly higher chance of developing HFpEF than those in the low-risk group (33.6% vs. 8.4%, p 0.001). Hence, the application of DLM to echocardiographic and ECG finding for screening HFpEF proved valuable [21]. AI coupled with ECG findings can also be used to highlight patients with HFrEF. A randomized control trial demonstrated AI to diagnose EF 50% in greater number of patients (2.1%) compared to usual care (1.6%) [22].

HeartModel is another modern invention that makes use of echocardiography in allowing consultants to determine disease status and treatment options. It is a software package that provides an automated analysis of echocardiographic findings in terms of echocardiographic parameters such as chamber volumes and ejection fractions [23]. A study conducted on the Chinese population compared the results of automated vs. manual echocardiography. It was found that atrial and ventricular volumes were a bit overestimated by the automated echocardiography device in comparison to manual echocardiography, while left ventricular ejection fraction (LVEF) was the same for both employed methods [23]. Statistical results favored the incorporation of automated echocardiography devices in place of manual echocardiography to assess chamber volumes and ejection fraction to diagnose HF [23].

The LR model of AI has been applied in the diagnosis of congestive heart failure (CHF). Son et al. [24] included CHF patients presenting with complaints of dyspnea. LR- based decision-making model was used to generate certain decision rules based on predictors of CHF. A decision-relative reduct was selected to generate decision rules using the Rough Set (RS) based decision model. The RS-based model performed much better in the prediction of CHF patients than the LR-based model, with an accuracy of 97.5% and 88.7%, respectively. It was supported by the AUC for the two decision-based models, 97.5 $\pm$ 1.1% and 88.8 $\pm$ 3.1%, respectively [24]. RF model also provides 100% classification accuracy in detecting CHF. A study conducted by Masetic et al. [25] made use of long-term ECG findings extracted using the autoregressive Burg method. Various classifiers were tested including SVM, artificial neural network (ANN), and k-nearest neighbors (k-NN) amongst which RF algorithm was selected due to a higher level of accuracy, and specificity for the diagnosis of HF [25]. Detection of HF six months before the actual clinical diagnosis was achieved by Wu et al. 2010 [26]. The Geisinger Clinic’s electronic health records were obtained to be used in the model; 179 independent variables were expressed. The authors used SVM, boosting, and LR methods for HF identification. The study demonstrated that HF was detected more than six months before the actual date of clinical diagnosis using the LR model and boosting. SVM demonstrated the poorest performance, possibly owing to imbalanced data [26].

Decision-making models involving decision trees have been utilized for the diagnosis of CHF in patients presenting to the emergency department. Rough sets (RS) and LR methods have been used to construct a decision-making model with the RS-based model proved to be more accurate than the LR model [24]. Aljaaf et al. 2015 [15] put forward a 5-risk level assessment (1 = no risk and 5 = extremely high risk) for HF prediction by making use of the C4.5 decision tree classifier [15]. The dataset used for the project was further improved by considering lifestyle factors such as obesity, physical activity, and smoking. The cumulative accuracy for the 5 risk levels was observed to be 86.3% [15].

The heart responds to stimuli from inside and outside the body, which eventually leads to heart rate variability (HRV). The usage of ECG in HRV has limitations because HRV analysis in identifying cardiac problems is usually incomplete. By using linear and nonlinear HRV findings, the accuracy of results can be improved [27]. Linear Autoregressive (AR) makes use of the frequency and time domain. The time-domain makes use of R-R interval, while the frequency domain uses the oscillation analysis of 5-minute ECG recordings or the Holter ECG, which monitors heart rhythm throughout the day [28]. Nonlinear AR takes the humoral, electrophysiological, and hemodynamic variables, etc. into consideration. Support vector machines (SVM) can also be utilized for the detection of CHF as a study demonstrated it to be amongst the top three classification methods with an accuracy rate of nearly 98%. This study analyzed ECGs for normal rhythm, SV arrhythmia, and CHF amongst hundred patients. Vectors were employed for this purpose comprising SVM, artificial neural network (ANN), C4.5 decision tree, and RF which then underwent clustering and classification. Another study was conducted to explore the nonlinear AR in further detail, where the analysis of the spectrum was instead using ML models such as SVM, decision trees, k-NN, and ensemble classifiers. SVM was used with its kernel, enabling it to take data as input and transform it into the required form, i.e., linear, Gaussian, linear base function, and polynomial. This study also reported SVM to demonstrate the highest performance amongst the different methods employed [29]. Finally, Zheng et al. 2015 [30] introduced a computer-assisted system and made use of the least-squares SVM (LS-SVM), which utilized heart sounds and cardiac reverse features. The results showed ANN and Hidden Markov Models to be dominant over the LS-SVM model.

The role of AI in assisting HF specialists in diagnosis can be made easy by potential assistance from Artificial Intelligence-Clinical Decision Support System (AI-CDSS) [31]. It is a hybrid system that contains both expert-driven and ML-driven knowledge to grow the knowledge base with HF [31]. Retrospective cohorts and prospective pilot studies with HF and non-HF patients were carried out to observe the accuracy of AI-CDSS, which was converted into mind maps and then to a decision tree to be assessed by physicians [31]. Machine-derived learning involved the usage of 5 algorithms that selected LVEF left atrial volume index (LAVI) and left ventricular mass index (LVMI) as contributing factors [31]. All algorithms were ranked for accuracy, the number of rules extracted, and the number of attributes involved [31]. The classification and regression tree (CART) algorithm was selected owing to its highest accuracy. Combining the two methods (ED and ML), lead to hybrid knowledge, the clinical knowledge model (CKM), which focused on the physical findings in patients, while a prediction model from ML looked for LVEF. The overall diagnostic accuracy was found to be 90% (expert-driven), 88.5% (ML-driven) and 98.3% (hybrid CDSS) [31]. Direct learning can assess complicated patterns in dataset more than NN, which is why DL is being incorporated more often in the field of medicine [20]. These models have also been studied for diagnosis of HF using chest X-rays. Matsumoto et al. [32] used 952 chest x rays and established a DL model with an accuracy of 82% for HF detection.

Modern and compact implanted sensors are highly efficient in measuring vitals that are highly specific to HF’s course and prognosis. These sensors are wireless and battery-less. Raised Left Atrial Pressure (LAP) is the earliest and most specific sign of HF before any symptoms appear. Implantable sensors can now easily record LAP and print a waveform. V-LAP is amongst the first battery-less cardiac monitoring devices which track LAP and provides remote HF care in such patients [33]. With the help of V-LAP, physicians can now continuously monitor a patient’s LAP and identify HF even before the onset of HF symptoms [33]. In recent years, pulmonary artery pressure (PAP) measurements at home have been possible through the implantation of pressure sensors in the pulmonary artery. CardioMEMS system is a monitoring device most commonly utilized to measure PAP [33]. It was used in the CHAMPION trial and a 33% decline in cardiac hospitalizations was noted over 18 months [34]. Apart from measuring PAP, devices such as defibrillators and pacemakers have shown inefficacy in preventing hospitalization/rehospitalization in contrast to the CardioMEMS [35]. The level of physical activity determines the course of HF to a great extent. Informative accelerometers are non-invasive devices that can be used to track the physical activity of HF patients which has been found to predict the risk of hospitalization [33]. Remote Dialectic Sensing (ReDS) uses electromagnetic waves to detect the extent of pulmonary congestion which infers lung field concentration, helping in the interpretation of CT scans of lung field concentration [33].

Various objects around us have been incorporated with an electronic software broadly termed the internet of things (IoT). It consists of 3 layers, a sensing layer that is present in the particular sensor used by the patient, a transport layer which is composed of connectors transporting data from sensor to remote device, and an application layer which is the server [36]. Physicians with the help of IoT can now monitor various vitals in their patients and forecast medical emergencies. This way, patients can now have medical care available at their doorstep, reducing the frequency of hospitalizations. When combined with specific algorithms, IoT can also warn us of heart attacks beforehand [36]. Cardiac resynchronization therapy (CRT) devices can additionally detect intrathoracic impedance, which further helps in predicting hospitalization in HF patients [37]. The MultiSENSE study presented an algorithm based on data from CRT defibrillators which showed 70% sensitivity in hospitalization prediction [38]. To avoid false-positive predictions in HF, the research highlighted this issue and provided that a telephone triage questionnaire can easily eliminate false-positive results [39]. Multisensor Non-invasive Remote Monitoring for Prediction of Heart Failure Exacerbation (LINK-HF) study made use of a non-invasive sensor placed on the patient’s chest using adhesive tape. The sensor could record ECG, 3-axis accelerometry, skin impedance, body temperature, and posture [40]. The Bluetooth-enabled connection between the sensor and cellular phone allowed for data transfer. Data-enabled cell phone syncs the transferred data to an encrypted cloud for viewing and storage. Cloud-based data gets converted to similarity-based modeling (SBM), and a baseline model is constructed within 72 hours of hospital discharge. Following this, the sensor on the patient’s chest switches to the scrutiny phase and monitors the vitals [40]. Variations in vitals due to normal activities are noted to remove false alarms. This is done by creating a multivariate change index range, from –1 to 1. The change in the patient’s physiology would indicate a greater index change, while no change in physiology would be closer to 0. The change in index range (–1 to 1) allows for continuous monitoring of patient’s vitals and provides indications for rehospitalizations [40].

To manage and prevent in-hospital cardiac arrest, a track and trigger system (TTS) and rapid response system (RRS) were introduced to identify cardiac arrest. Although TTS proved to be of great help, it was associated with high false alarms and low sensitivity. Single parameter TTS is defined as when anyone’s vital sign concerning the heart, i.e., blood pressure, heart rate, respiratory rate, body temperature, and mental status is found to be abnormal [41]. To avoid a high rate of false alarms, a deep learning-based early warning system (DEWS) was introduced. This system has better sensitivity and a low rate of false alarms. The DEWS make use of ML by identifying trends in the dataset, including specific vitals, as mentioned above [42]. The model was used in two hospitals, a cardiovascular teaching hospital, and a community general hospital. Patients who had cardiac arrest had experienced death within 30 minutes of admission, or who were outside the study period were excluded. DEWS is a 3 layered neural network system which makes use of the time series data [42] by going over records to conclude a patient’s current state. DEWS was compared with modified early warning score (MEWS) at 3 sensitivities and SPTTS at 1 sensitivity. It was also compared with logistic regression and random forest at 75% sensitivity [43]. The area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), and confidence intervals (CIs) were greater for DEWS compared to other models [42]. The specificity, positive and negative predictive value, F measure, MACHP, and net reclassification value were assessed using 75% sensitivity. DEWS was found to be the most accurate amongst all systems, hence DEWs can be used in the RRS to identify cardiac arrest [42, 43].

Algorithms can be used to classify HF into many types in the same way they have been used to diagnose patients with HF, well beforehand, and prevent hospitalization. Unsupervised ML performs grouping, i.e., observes similarities in the total dataset and compiles similar data in one place whereas supervised ML identifies trends and predictions [5, 44]. A study used model-based clustering (MBC) to classify HF in different phenotypes. HFpEF patients were taken in the study along with their variables such as echocardiography and laboratory results [45]. MBC made use of R in the mclust package. After the application of MBC, classifiers were made to assign patients in their respective groups using Elastic Net, Neural Networks, and Naive Bayes. Resampling was done using cross a validation procedure in which four-fifth of the sample was used in the optimization of the model parameters, while the remaining of the sample was used in the prediction performance assessment, i.e., the ability of the ML model to put samples in the correct phenogroups [45]. Six phenogroups were identified via MBC; phenogroup 1 consisted of young patients with cardiovascular risk factors, including left-sided heart changes along with patients that progressed to chronic kidney disease. Phenogroup 2 patients included severe HF with the highest degree of diastolic dysfunction, and deteriorating right ventricular function. Phenogroup 3 included young patients with milder HF variants. Phenogroup 4 included male patients with hypertension and enlarged left atrium contributing to the risk of atrial fibrillation. Phenogroup 5 and 6 included females with hypertension, atrial fibrillation, and low body mass index (BMI) [45]. After establishing the protocol, a test run was carried out by clustering patients and using Elastic Net to put them into phenogroups accordingly [45].

In another study, unsupervised clustering was used with the addition of CRT devices. The variables including echocardiography values were recorded, and the ML algorithm was used, after which the samples were clustered using the K-algorithm to identify 4 phenogroups undergoing CRT [44]. Phenogroups 1 and 3 involved females with cardiomyopathy and left bundle branch block (LBBB), while phenogroup 1 was found to have prolonged QRS interval in comparison. Phenogroups 2 and 4 involved males with HF, predominantly caused by ischemia but had a lesser percentage of patients with LBBB [44]. Cluster analyses on patients with HFpEF using exercise tolerance have also been performed in published literature. A study included HFpEF patients who underwent echocardiography during rest and following exercise to monitor heart function at both states. Clustering was done of noted variables which lead to the separation of the total dataset into two distinct phenogroups based on data similarities and differences. The first phenotype had a decreased chronotropic/diastolic reserve, ventricular arterial coupling (indicating global heart efficiency), and abnormal longitudinal deformity, while the other phenotype showed a preserved heart rate reserve combined with low left ventricular systolic reserve [46]. Few studies have classified HF using conventional tress and ML classifiers. Classification trees divide the total dataset into many subsets but are found to have questionable accuracy, while ML uses an aggregate of classification trees [24]. Classification trees are simple to use and can utilize binary methods to dividing the dataset into two subsets. The classification of HF has been performed in two parts in a study with all classification techniques used in a manner to keep two classifications as their result, i.e., HFpEF and HFrEF.

The bootstrap technique, also known as bagging is a modern technique that has been previously implemented in literature. Multiple bootstraps were applied to the sample size, and a classification/regression tree was utilized for each bootstrap. The bootstrap samples were compiled with help of majority votes in each sample, and the classification was obtained [24]. Boosting classification technique makes use of multiple weak classifiers and “boosts” their result. A weak classifier has a margin of error slightly better than of guessing [24]. When a classifier puts a subject in the wrong class, the class gets weighted heavier hence at the end when it’s all summed up, the samples get correctly classified and placed in subsets [24]. SVM makes use of the hyperplane concept, where the dimensional space gets divided into two distinct planes. Hence two distinct subgroups are made in a single sample. The sample on one side of the hyperplane and the one on the other side are considered as two classifications in a sample [47].

AI can be of advantage in identifying potential cardiac events faster than a clinician could, using predictive analytics. As described above, ML; the most common application of AI, can recognize patterns in data to improve cardiovascular diagnosis and detection [11]. Mobile health and the internet of things have allowed patients to keep a track of their health parameters, which have been particularly useful for home-based long-term monitoring, detection, and subsequent alertness for abnormal cardiovascular findings, and timely contact with doctors when necessary.

According to the World Health Organization (WHO)’s Global Observatory for eHealth, Mobile Health (mHealth) is defined as “medical and public health practice supported by mobile devices, such as [cell] phones, patient monitoring devices, personal digital assistants (PDAs), and other wireless devices” [48]. mHealth enables patients to track their health, e.g., in the form of keeping a check on their heart rate, blood glucose levels, medication dosages, and sleep cycles. It also allows for remote consultations and maintaining electronic health records [49]. It has also been used in managing chronic diseases such as diabetes, hypertension, and chronic obstructive pulmonary disease [50]. The mHealth-based interventions have several factors, potentially making them effective for cardiovascular disease; according to Dicianno et al. [51], these factors include: (i) ‘interactivity that allows the bidirectional communication and the care delivery, as a form of a personal coach’, (ii) ‘personalization’ that allows for the mHealth interventions to be customized according to the individual’s needs, (iii) ‘timeliness’ which allows for delivery of information within the right time frame, (iv) ‘context sensitivity’: the ability of interventions to be adapted and modified according to the circumstances and/or the individual’s needs, and (v) ‘ubiquity and accessibility of the technology to all segments of population’ [51]. The advantages of mHealth in the management of HF has been demonstrated in Fig. 2. In addition, wearables (e.g., smartwatches, wristbands, patches and sensors) are being used to measure and record a range of parameters, such as heart rate, blood pressure and temperature, and to track diseases such as HF and diabetes [52]. Wearables not only track and transfer biometric data into a shareable and comprehensible user interface but also allow for the continuous monitoring of more than one dozen biometric parameters as reported by several developers of wrist-worn sensors [53].

Fig. 2.

The advantages of mHealth in the management of HF. mHealth based interventions has been effective in managing cardiovascular diseases owning to the various benefits including interactivity, personalization, context sensitivity, timeliness, ubiquity, and accessibility, adapted from [51].

A study conducted in 2019, under the collaboration of Stanford University and Apple, assessed the ability of a smartwatch to detect atrial fibrillation or atrial flutter using algorithms that use pulse wave data detection of these episodes [48]. According to this study, only 0.52% (2161) of the total participants received an irregular pulse notification. Among those who received an initial notification and returned an ECG patch, it was reported that 84% of their subsequent notifications were confirmed as atrial fibrillation. Moreover, 76% of the participants who received notification were reported to have contacted either a telemedicine provider or a non-study healthcare provider.

According to a meta-analysis conducted in 2020 among HF patients (n = 1683), mobile phone-based interventions were associated with a significantly lower rate of all-cause hospital admissions at six months (31% vs. 36%, OR 0.77, 95% CI 0.62–0.97, p = 0.03, I${}^2$= 0). A significant difference was also demonstrated for HF admissions (14.0% vs. 18.5%, OR 0.69, 95% CI 0.48 to 0.98, p = 0.04, I${}^2$ = 26%) [54]. However, another meta-analysis found mobile monitoring interventions, mostly consisting of a mobile communication device, a blood pressure measuring device, a weighing scale, and an ECG recorder, to be inconclusive for monitoring HF outcomes. It is also difficult to achieve a 100 percent adherence to mHeath based interventions, as seen in many study groups reported by this meta-analysis, with only one study reporting 100% adherence [50]. However, the popularity of mhealth and its potential utilization cannot be undermined since, in the US alone, 91% of adults have been reported to own some kind of cellphone, with 56% reported to be smartphone users [48]. Moreover, a study conducted in 2017 also stated that older individuals had a positive outlook towards the benefits of mHealth and were eager to utilize mhealth interventions, despite facing usability issues with mHealth such as in navigation and visualization, etc. [55].

The Internet of Things or IoT is simply a network of different devices in various physical objects or locations, that allows for the exchange and storage of data, using network connectivity [56]. A study conducted in 2017 proposed the structure of an IoT-based system comprising of 3 different layers: the sensing layer, the transport layer, and the application layer, and 4 different modes that could be utilized for heart-disease monitoring. This system could allow for more cost-effective and efficient transmission of data, enabling communication with healthcare practitioners in real-time, in a more structured and systematic way [36]. For cardiovascular management, IoT enables healthcare practitioners to monitor their patients remotely and in real-time, and to collect data, such as blood pressure, pulse rate, oxygen saturation, and ECG simultaneously. This could drastically change the way patients receive medical assistance, making it easier for them to get healthcare at their homes in a timelier fashion, reducing hospital visits concurrently. If IoT is utilized in cohesion with certain real-time algorithms, this could be helpful to alert about potential attacks beforehand [5].

IoT is particularly useful in the detection and diagnosis of certain conditions that may require monitoring throughout the day. To make an accurate diagnosis for arrhythmias, for example, the patient’s ECG needs to be monitored for an entire day, at minimum. Hence, making them especially challenging to detect. Moreover, they can be intermittent, so detection and diagnosis can be markedly enhanced by long-term continuous monitoring, using IoT techniques. According to an article published in 2019, mobile cardiac telemetry (MCT) devices that providing real-time monitoring of a patient’s heart rhythm over a longer period are essential for AFib detection. MCT devices are the only heart monitoring devices that provide complete arrhythmia detection, with the highest diagnostic yield at 61%, compared to Event monitors at 23% and Holter monitors at 24% [57]. In a study conducted by J. Stehlik et al. [58], it was suggested that non-invasive devices such as wearables may be more useful and cost-effective in predicting and recognizing the risk of rehospitalizations for heart failure.

ML techniques including supervised and unsupervised networks as well as DL has been incorporated in different cardiac imaging procedures for accurate quantitative and qualitative evaluation of cardiac diseases, particularly, HF.

ML models have been trained to recognize specific images of a wide variety of cardiac diseases [59]. This helps in the interpretation of unused data in three-dimensional imaging, thus leading to faster analysis and better outcomes. Furthermore, DL has been mostly utilized in imaging for segmentation of the ventricles. Chen et al. [60] demonstrated the use of convolutional neural networks (CNN) to segment the ventricle into 5 different 2D views. ANN model was previously used for segmentation which was further extended in studies by Carneiro and Nascimento [61]. Dong et al. [62] and Ghesu et al. [63] also combined DL with traditional methods to segment the left ventricle and the aortic valve. Dezaki et al. [64] utilized electrocardiograms in combination with RNN and CNN to predict end-systolic and diastolic volume. Table 2, Ref. [46, 65, 66, 67, 68, 69, 70, 71] identifies some models for prediction, differentiation of cardiovascular diseases, accurate measurement using echocardiography features providing increasing automaticity, and the ability for segmentation of cardiac structures.

AI and ML algorithms have proved to be of crucial importance in other imaging techniques like Cardiac MRI particularly for ventricular segmentation. Machine learning (ML) can segment the heart chambers from Cardiac MRIs and these segmentations yield imaging biomarkers to predict CHF. Laser et al. [72] used knowledge-based reconstruction of the right ventricular volumes using images from echocardiography and cardiac MRI and compared them with the gold standard direct cardiac MRI and found that knowledge-based reconstruction has excellent accuracy for right ventricular 3D volumetry. Right ventricle has a complex shape which many times could not be visualized with 2D imaging echocardiography techniques. 3D visualization and cardiac image reconstruction with the help of AI can help in the diagnosis of a variety of similar diseases. Similarly, calculation of left ventricular mass, papillary muscle identification, common carotid artery, and descending aorta measurements with fully automated AI programs have produced far more accurate results. Another study extracted CMRIs from 350 subjects and used CNNs to segment the heart in short and long axis Cardiac MRI. He concluded that cardiac MRI provides a reliable method of extracting indicators for cardiac measurements and function [73, 74]. Table 3, Ref. [13, 72, 75, 76, 77, 78, 79, 80, 81] summarizes the findings of studies that utilized DL and SVM models using images obtained from Cardiac MRI.

ML based image analysis has been increasingly used for cardiac CT specially for evaluation of coronary artery disease and atherosclerosis. ML models were used for coronary artery calcification scoring and risk stratification. Commandeur et al. [82] used 2 CNNs to segment epicardial and thoracic adipose tissue using cardiac CT images. They evaluated their method in a large cohort of 1638 patients. Multi scale patch-based CNN was also used by Zerik et al. [83] to segment the left ventricle myocardium. Wu et al. [84] used short term memory RNN to label segments of coronary artery tree. Itu et al. [85] and Coenen et al. [86] also used ML models for quantitative analysis of fractional flow reserve in coronary artery disease. The implication of ML models using datasets from Cardiac CT has played a major role in diagnosis of coronary artery stenosis, and risk stratification of future events.

ECG is the most widely used diagnostic modality for detection of heart diseases. The application of ML models particularly DL has helped reduce the time taken in diagnosis, and speed up urgent care. Supervised learning models have been used to classify heart rhythm. Algorithms involving unsupervised learning analysis have also been used. In this method, unlabeled data has been grouped according to its ECG phenotype. This was demonstrated by Lyon et al. [87] who classified ECGs with arrhythmia risk factors in patients of cardiomyopathy. Hannum et al. [88] used a 34-layer deep NN to classify 12 different types of arrhythmias. Attia et al. [89] used a 6-layer deep NN to diagnose left ventricular systolic dysfunction. In this study, investigators attempted to diagnose asymptomatic left ventricular dysfunction by EKG alone using an AI-based CNN deep learning method. This model proved to be superior to the conventional method of using BNP levels for estimation.

Early detection of future mortality and episodes of destabilization can help provide quality care earlier in the management of the patient and also allow doctors to make crucial clinical decisions when needed. Using statistical analysis, multiple risk calculation scoring systems have been formed for estimation of mortality.

Candelieri et al. 2008 [90] used knowledge discovery (KD) models to predict if a patient with HF in a stable phase would decompensate or not. A group of 49 CHF patients was used for the evaluation of the KD approaches. Decision trees, Decision Lists, SVM and Radial Basis Function Networks were used, and the leave-patient-out approach was followed to evaluate the performance. Of the models, decision trees proved to be superior to other models and provided an accuracy of 92.03%, sensitivity 63.64%, and False Positive Rate 6.90%. In 2009 Candelieri et al. [90] used decision trees and SVM on an independent testing. Their results showed that SVM is more reliable in predicting decompensation events with an accuracy of 97.37%, 100.00% sensitivity. This was further extended through the “SVM hypersolution framework” which proved to be more accurate on minority class than Tabu search. Guiti et al. [91] also predicted the frequency of HF decompensation during the year after the first visit using five machine learning techniques (NN, SVM, Fuzzy -Genetic Expert System, Random Forests and CART). CART algorithm proved to be the superior one with 87.6% accuracy.

Re-hospitalizations add to the burden on the healthcare system and lead to poor quality of life for the patient. Predictive models have been developed to predict the risk of future hospitalization so that adequate monitoring and management can be done to prevent such adverse outcomes. At the same time, mortality can be improved using risk prediction models and discrimination of patients as demonstrated in the studies shown Table 4, Ref. [92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104]. Table 5, Ref. [5, 13, 24, 25, 30, 31, 34, 37, 40, 42, 44, 45, 46, 47, 51, 66, 69, 71, 75, 88, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128] provides the summarized evaluation of the advantages and limitations of some of AI models and devices for HF.