1.1. Traditional rule-based algorithms apply rules to data, while machine learning algorithms learn patterns from data

The main difference between machine learning algorithms and traditional algorithms lies in their input and output. Traditional algorithms are based on rules and logical statements written by programmers that require precise definition of input and expected output. For example, a sorting algorithm requires an unordered list as input to generate a sorted list in ascending order as output.

In contrast, machine learning algorithms are based on statistics and data analysis and can automatically learn patterns and rules from data. Machine learning algorithms can handle raw or unclassified data and learn and adapt from it. The input for machine learning algorithms can be large datasets, and the output is typically a prediction or classification. For example, machine learning algorithms can learn natural language processing rules from large amounts of natural language data to perform natural language understanding or generation tasks more accurately (15).

Another difference is that traditional algorithms are usually deterministic, meaning they always produce the same output for a given input. In contrast, machine learning algorithms can be stochastic or probabilistic, meaning they may produce different output results for a given input because they are based on probabilistic models.

Finally, machine learning algorithms typically require a large amount of data to train the model to improve the accuracy of prediction or classification, whereas traditional algorithms typically use less data. There are several types of machine learning algorithms, from decision trees and support vector machines (SVMs), to highly complex, data-hungry algorithms called neural networks. Neural networks are used in deep machine learning (deep learning or DL), and their ability to analyze large amounts of highly complex data—electronic health record (EHR) data, for example, or the collection of pixels that make up medical images—are especially exciting for cardiology applications (16).

1.2. ML algorithms can learn patterns from labeled examples: supervised learning

Supervised learning algorithms need to be trained using existing datasets in order to learn regularity and patterns. These datasets are typically from hospitals or research institutions and include medical images, medical records, and experimental data. The accuracy of supervised learning algorithms depends on the quality and quantity of the datasets used, so a large amount of high-quality data is required to train the model (17).

The advantage of supervised learning algorithms is that they have a clear goal: to predict the labels of interest. But the downside of supervised ML algorithms is that their ability to find interesting patterns in the data is also limited by these labels. Training the right data and deciding on the right answer or label is critical to training and requires a lot of work. Similarly, a major challenge in supervised machine learning is the availability of datasets of sufficient size that have properly annotated labels of interest. However, this is also not necessarily accurate. Therefore, proper labeling of datasets requires active management by physicians and often requires consensus from more than one physician.

1.3. ML algorithms can learn patterns without labeled examples: unsupervised learning

Unsupervised learning is a machine learning method that uses unlabeled data for training to discover regularity and patterns in the data without first having labeled data. Unlike supervised learning, unsupervised learning does not require predefined inputs and outputs, but instead lets the algorithm learn the structure and features of the data on its own. The main goal of unsupervised learning is to discover hidden structures in data in order to better understand the data and extract useful information.

The advantages of unsupervised learning include the following:

• (i) No need to label data: Unlike supervised learning, unsupervised learning does not require labeled data for training, so it is easier to obtain a large amount of unlabeled data.
• (ii) Discover hidden structures: Unsupervised learning can help discover hidden structures and patterns in data to provide better data understanding and analysis.
• (iii) More comprehensive data analysis: Unsupervised learning can use a variety of algorithms to analyze data from different perspectives, so as to be able to understand the data more comprehensively.
• (iv) Applicable to a variety of fields: Unsupervised learning can be applied to a variety of fields, such as image processing, natural language processing, data mining, etc. (11).

3.2. Distribution by journal and author

Table 1 lists the journals of the top 100 cited papers in descending order, with the number of articles, average number of citations per publication, and impact factor (2021). In total, 49 journals were included. The following journals had >3 instances: Artificial Intelligence in Medicine (n=5), Journal of the American College of Cardiology (n=5), Computer Methods and Programs in Biomedicine (n=4), BMC Medical Informatics and Decision Making (n=3), European Heart Journal (n=3), European Journal of Heart Failure (n=3), JACC-Basic to Translational Science (n=3), and American Journal of Surgical Pathology (n=3). A total of 206 authors participated in the preparation of these most-cited articles, and 15 authors participated in the preparation of three or more articles. Daniel Rueckert wrote 5 of the 100 top-cited articles, topping this list. Table 2 includes a list of these 15 authors.

Analyses of the data were descriptive in nature (22). According to the total link strength (TLS), which measures the overall strength of connections between a certain researcher and the co-authors of other articles, projects were divided into clusters. The citation map of journals and the co-authorship map of authors for the 2,392 articles are shown in Figures 3A, ​,4A.4A. A total of 43 journals have contributed to machine learning in heart failure. Among them, Frontiers in Cardiovascular Medicine was the leading journal with 78 documents, followed by Circulation with 58 documents, and PLoS One with 52 documents. The top three co-authorship triads of authors were Pandey Ambarish (TLS=38), Segar Matthew W. (TLS=36), and Acharya U. Rajendra (TLS=35). In comparing highly cited articles to all articles, we found a big difference. The impact rankings for each country and each journal provided are based on citation rates (23).

Open in a separate window

Figure 3

Visualization knowledge maps of citation. (A) Citation map of journal; (B) citation map of documents. Different color indicates different clusters. The size of the nodes represents the count of citations. The distance between the two nodes indicates their correlation.

Open in a separate window

Figure 4

Visualization knowledge maps of the co-authorship. (A) Co-authorship map of authors that indicates the authors that cooperate in the field of heart failure. (B) Co-authorship map of countries. (C) Co-authorship map of organizations. Different colors indicate different clusters and the size of nodes indicates the number of publications. Thickness of the lines represents link strength of the countries.

3.3. Analysis of high-cited references

The most-cited reference was published in the New England Journal of Medicine and authored by Alan S. Maisel in 2002. The second most-cited reference was published in Circulation and authored by Rahul C. Deo in 2015. The third most-cited reference was published in the Cochrane Database of Systematic Reviews and authored by Jasvinder A. Singh in 2011.

These top three articles were analyzed. The most-cited reference by A. S. Maisel in 2002 (24) reported that measurements of B-type natriuretic peptide added significant independent predictive power to other clinical variables in models predicting which patients had congestive heart failure, using multiple logistic regression analysis. Used in conjunction with other clinical information, rapid measurement of B-type natriuretic peptide is useful in establishing or excluding the diagnosis of congestive heart failure in patients with acute dyspnea.

The second most-cited reference by R. C. Deo in 2015 reported that part of its effort was to identify what obstacles there may be in the change of the practice of medicine through statistical learning approaches, and discuss how these might be overcome (19).

The third most-cited reference by J. A. Singh in 2011 used mixed-effects logistic regression using arm-based random-effects models within the empirical Bayesian framework, and planned to combine the results of biologics used in many conditions to obtain much-needed risk estimates (25). The citation map documents are shown in Figure 3B.

3.4. Countries and organizations of top 100 cited articles

Countries and institutes of origin of the top 100 cited articles are listed in Tables 3, ​,4.4. A total of 24 different countries contributed these 100 articles. The United States (n=50) with half of the articles was the most prolific country, followed by China (n=7), England (n=5), Singapore (n=4), Canada (n=4), Spain (n=4), Australia (n=3), Turkey (n=3), Italy (n=3), and others. The institutes with the largest number of publications were Harvard University in the United States (n=17) and University of California in the United States (n=10). Icahn School of Medicine at Mount Sinai in Egypt produced eight articles. Two institutes produced seven articles: Mayo Clinic in the United States and Stanford University in the United States. As for all the articles, the co-authorship map of countries and institutes is summarized in Figures 4B,C. The top three countries in co-authorship were the United States, China, and England. The top three institutes in co-authorship were Harvard Medical School in the United States, Duke University in the United States, and University of Toronto in Canada. The field of AI healthcare attracts research from all over the world, but high-income countries are a major force in healthcare-related AI research. The United States alone contributes about half of the research in healthcare-related AI research (26), staying far ahead of other countries in terms of both quantity and quality of its contribution in this area. For researchers, understanding the ranking results can help them better choose suitable research partners, publish research results in appropriate journals, and participate in academic conferences.

4.2. The United States is leading the way in the area of heart failure and machine learning

The United States accounts for 50 of the 100 top-cited articles globally, which confirms the tremendous impact on medical science research in the United States with its large scientific population and the sufficient financial resources which are available to the scientific communities, demonstrating its dominance in the fields of research, technology, and medicine. The United States is a pioneer in many other fields as well (29). The United States hosts many prestigious journals, which may be the reason why publications there account for the vast majority. Nevertheless, researchers prefer to cite papers from their own countries. American researchers may similarly favor papers from their own country (30, 31). For papers published in WOS, this work is with significant per-capita contributions from all industrialized nations as well as work in the majority of developing nations. Global and similarly multi-institutional collaborations in cardiac AI/ML exist within the United States.

It is undeniable that citation analysis within a specific subject may always provide a wealth of information about journals, organizations, and authors that is useful for identifying important publications and high-impact journals. It shows the trajectory of heart failure research as well as a historical perspective on scientific advancement in the subject. Through the analysis of the statistics of the journals, authors, and countries of the first 100 citations, we can learn advanced medical technologies and concepts, learn the advanced nature of the country that plays a leading position in this field, understand the future development trend of the field, and timely improve the shortcomings and shortcomings of the data analysis methods used by medical scientists. Most importantly, our statistical results are intended to provide these healers with the direction and orientation to solve problems (32).

4.3. Range of machine learning methods in heart diseases in United States, Europe, and Asia

We divided the 100 articles based on their original countries for analysis: the United States, Europe, and a few Asian countries (mainly from China). Figure 6 displays the articles from the three areas that examine the main cardiac conditions and machine learning techniques. The tables on the left of the three regions show heart failure and its common causes (33). It demonstrates that the worldwide research results about origins of heart diseases are much the same. SVM, a binary classification method that uses machine learning techniques, occurs most frequently. Its basic model is defined in the feature space of the largest interval linear classifier. Its advantage is that only a small number of support vectors determine the outcome, which not only helps identify important samples but also eliminates redundant samples (34). These techniques have been widely adopted by medical research to predict disease and survival.

Open in a separate window

Figure 6

Major heart diseases and machine learning methods in the United States (A), Europe (B), and Asia (C).

However, the interesting thing is that in the article “Comparing different supervised machine learning algorithms for disease prediction,” we found the following description. They discovered that the SVM algorithm is applied most frequently (in 29 studies) followed by the Naïve Bayes algorithm (in 23 studies). However, the Random Forest (RF) algorithm showed superior accuracy comparatively. Of the 17 studies where it was applied, RF showed the highest accuracy in 9 of them, i.e., 53%. This was followed by SVM which topped in 41% of the studies it was considered (35). We also found that different diseases use different machine learning models to achieve different accuracy.

4.4. Analysis of the application of machine learning in heart failure

Millions of individuals die from cardiovascular disease yearly worldwide. Heart attack (induced by obstruction of blood vessels), stroke (caused by occlusion or rupture of cerebral blood vessels), and heart failure are the three main heart and blood vessel diseases (caused by the inability of the heart to pump enough blood to the body). Prediction of patients’ status and prognosis based on clinical and laboratory data is crucial since severe heart failure can result in mortality (36).

Using machine learning, heart failure may be predicted, detected, and treated with high reliability and accuracy. Among these, data analysis and model building for heart failure frequently use methods like Neural Networks, Support Vector Machines, Decision Tree, and Random Forests. Physiological data, electrocardiograms, medical records, and images are just a few of the numerous data sources that offer rich data resources and research foundations for application of machine learning in heart failure (37).

Since SVM appears the most, we analyzed the articles in which this keyword appeared. We found that most of the SVM used by these authors are not traditional SVM models. Here are some examples: (i) SVM and boosting algorithms (38); (ii) an expert system that stacks two (SVM) models. The first SVM model is linear and L-1 regularized and the second SVM model is L-2 regularized (39); (iii) RBF kernel-based SVM (40); (iv) a boosted C5.0 tree, as the base classifier, was ensembled with a SVM, as a secondary classifier (41); (v) the proposed prediction SVM model with particle swarm parameter (42). These are some of the representatives of the 100 top-cited articles. It is all known to us that just building a working AI model does not make it usable. We have to combine and compare different models to improve our accuracy, and we have to pay attention to the development of technology and update our technology, because each model has its advantages and disadvantages. For example, random forests are suitable for handling large datasets, and this ensemble-based classifier may outperform a single classifier, while SVM is less prone to overfitting and performs well in classifying semistructured or unstructured data such as text, images, etc. (35). We can learn more about the differences from this paper.

4.5. Specific applications of machine learning

As most machine learning algorithms are “data agnostic” (43), the case studies are organized not by clinical data types or disorders, but rather by machine learning use cases.

Machine learning has some of the following key capabilities

Denoising and image enhancement: Machine learning has been used to address the clinical issue of image post-processing and denoising across a variety of modalities. For ultrasound, CT, MRI, and clinical imaging, neural network-based denoising has proven beneficial due to the complicated patterns of noise found in clinical imaging (44). In terms of clinical utility, machine learning has the ability to reduce the time and effort required for image post-processing as well as the discrepancies in data processing between operators, vendors, and institutes (45). Most machine learning denoising techniques now employ supervised learning techniques, where the model is trained to approximately represent proprietary denoising software as a real label (46). A crucial next step to demonstrate its usefulness in clinical practice is the use of machine learning–based image denoising to monitor changes in diagnostic performance. To test image denoising on a broader scale, several clinical trials are now underway (47).

Feature extraction, feature selection, and feature representation: Feature selection is the process of choosing the most pertinent characteristics from a vast number of features, aiming to increase the precision and interpretability of the prediction model. For instance, feature selection techniques based on Decision Trees or Logistic Regression algorithms might be used. Representing important features in a simpler and more unified form is called feature extraction. For example, myocardial wall thickness measurement is a crucial diagnostic, monitoring, and therapeutic tool for the identification of coronary heart disease and cardiomyopathy. The outlines and thickness of the myocardial wall can be automatically extracted and determined using feature extraction methods in medical image analysis (48).

Measuring cardiac flow is vital for evaluating cardiac function and lesions as it provides important diagnostic information about the volume of blood the heart pumps each minute (49). Heart disorders including coronary artery disease and heart failure may be diagnosed and treated using feature extraction algorithms. These algorithms can automatically extract the contours of the ventricles and blood arteries and determine the flow rate and velocity (50).

Deep learning algorithms achieve higher classification performance through automatic extraction of these characteristics through intricate nonlinear combinations of input data (51). Data may be compared between institutes using these extracted features, and by integrating features from various data types, it is possible to create a multimodal image of a specific patient or disease by integrating features from various data types. This study will continue to benefit from concurrent experiments using small datasets as well as synthesizing larger datasets from smaller datasets, because training a model to execute these fundamental tasks frequently requires a lot of data labeling.

With a small amount of data available, synthetic knowledge synthesis (SKS) can be a useful tool for heart failure feature extraction. Existing data can be combined with domain expertise to create more data samples, enlarging the dataset and enhancing the model's performance. Synthetic knowledge synthesis can also aid in avoiding issues such as overfitting. However, SKS cannot deal with all issues. There is just one viable answer and more experimental testing is needed to determine whether it applies to certain datasets. For effectiveness, it is necessary to combine SKS with other machine learning approaches, such as feature selection, classifier selection, and so on (52).