Few topics are as consequential, even in medicine, as the potential ability of modern technology to develop capabilities similar to those of human beings. The ability of machines or computer systems to mimic human intelligence processes is called artificial intelligence (AI). AI is becoming highly advanced in other professional disciplines and our current challenge is to transfer all of this development to the medical field and specifically to cardiology.

This article aims to clarify for the reader various terms that may still seem foreign, not just AI, but also machine learning (ML), deep learning (DL), data science, and big data, and to describe in detail the concept of AI and its types, the learning techniques and methodology used by ML, cardiac imaging analysis with DL, the contribution of this technological revolution to classic statistics, and its current limitations, legal aspects, and, critically, initial applications in cardiology.

The terms data science, big data, AI, and ML have different meanings, despite belonging to the same discipline: the analysis and extraction of information from data. However, these concepts are sometimes used interchangeably and in ambiguous ways.

The term big data was first coined in 2005 by R. Magoulas, who described it as a massive volume of data that, due to size, overwhelms traditional storage and processing software. In turn, the world of big data revolves around 7 Vs: data volume, velocity, variety, veracity, validity, volatility, and value. That is, big data generate information at high speed, controlling the integrity of the data and exhibiting huge variety in terms of nature and type (eg, text, images, videos, different systems, and different providers). In cardiology, big data would be viable within a framework of national, European, or global collaboration, involving data homogenization and sharing by hospital entities to create large-volume repositories. These large databanks can support our daily practice by helping to establish protocols, promote standardized and early disease diagnosis, aid prognosis of disease progression, and support treatment planning for our patients.1

This type of information processing would not be possible without recent technological advances in AI. AI uses tools able to identify existing patterns in data. It was defined for the first time in the 1950s and encompasses multiple subdisciplines, from expert systems and robotics to ML. It is precisely this last term that we will focus on in this review due to the major advances in the field in recent years.

The term machine learning was first coined in 1959 by AL Samuel.2 Its objective is to develop algorithms enabling computer systems to make decisions and learn from the results; these systems would thus be able to learn to do something without having been explicitly provided with the programming required.

Among the many ML techniques, there is more and more interest in DL models.3 These base their predictive power on artificial neural networks (NNs) and are characterized by having multiple layers of information processing (transformations) that allow them to analyze datasets with more complex patterns. These algorithms are enabling pivotal advances in speech and image recognition; the latter is particularly important in cardiology.

The data science field encompasses all aspects related to the derivation of the existing information contained in data. It involves scientific methods, processes, and systems to extract knowledge or develop a better understanding of data, normally applying novel processing techniques such as AI. Although AI and ML are often used as synonyms, they are different. AI covers a broader scope of this technological revolution and includes both ML and DL (figure 1A). The right-hand panel of figure 2 illustrates the interest triggered by these concepts in the last 3 years, with increasing interest in ML, AI, data science, and DL. Big data is the most common term In Spain, southern Europe, and South America, maching learning in the United States, Canada, the United Kingdom, northern Europe, and Australia and deep learning in China and Japan.

Figure 1.

A: Venn diagram of the most commonly used terms in the data science disciplines. B: total searches (source: Google Trends) in the last 5 full years of terms related to artificial intelligence and data science; the vertical axis of the diagram represents the proportion of a topic with respect to the total number of searches on the topics. C: the most searched term in each country in the same period.

(0.24MB).

Figure 2.

Application methodology of a machine learning model. DB, database; ML, machine learning.

(0.43MB).

Nowadays, it is difficult to find a universal definition of what is known as AI. The term itself is often applied to the field of computer science that tries to mimic human cognitive processes, learning capacity, and knowledge storage. Other definitions are broader and encompass the understanding and construction of intelligent entities, generally understood as computer software. In terms of types of task, AI can refer to automated systems capable of, for example, translating a document, recognizing people by their facial features, or driving a car. However, AI is not limited to imitating human tasks: in some cases, it is able to beat the best expert in a field by making decisions with lower error rates than humanly possible or by identifying patterns imperceptible to the human eye.4 Therefore, AI allows information to be analyzed with a distinct approach from the traditional one. Thus, when responding to an event, we are no longer limited to describing the available information. AI allows other questions to be answered: what happened? (diagnosis), what will happen? (prediction), and what should I do? (prescription).

The most important application areas of AI in the health care field include the following: automatic speech recognition and natural language processing; prediction, recommendation, and diagnostic algorithms; computer vision and image analysis; robotics; and AI and expert systems.

ML is an area of data science that forms part of what is known today as AI. ML involves the creation of systems that learn automatically; they understand by autonomously “learning” the ability to recognize complex patterns, without the need for human intervention and in datasets of any kind, including numerical, visual, and textual. As the experience of these systems increases, that is, as they are provided with new data, their performance improves to a point that may even exceed the human capacity in that task.

Although there are several ML techniques, they are usually grouped into 2 types: supervised learning and unsupervised learning.

Supervised learning techniques are undoubtedly the most widely used methods in ML and those with the best results. These procedures rely on a dataset from which the response variable to be predicted (eg, diagnosis, parameter, segmentation) is determined through the correct labeling of examples. Depending on the type of prediction, classification algorithms or regressive algorithms are used. In the first case, the aim is to identify 2 or more classes using a series of variables. In contrast, regressive algorithms seek to approximate a continuous value as much as possible.

In unsupervised learning techniques, there is no information on the variable to be predicted. These techniques must learn from the relationships among the elements of a dataset and classify them without relying on labels or categories. To do this, they look for structures, patterns, or characteristics in the source data that can be reproduced in new datasets. For this task, clustering methods are the most commonly used approaches.

The construction of an ML model does not simply involve the application of a learning algorithm to a database, but is a whole process,9,10 which usually includes the steps shown in figure 2.

The first stages are common to conventional statistics; raw data are converted into information with structured data (preprocessed), and an initial database is constructed (step 1 of figure 2). From this database, a descriptive and exploratory analysis is performed to identify and select the most significant variables; these variables will be directly applied to the ML algorithms (step 2 of figure 2).

The next stage, now specific to ML techniques, is the division of the dataset into 3 subsets—training, validation, and testing—, typically 60%, 20%, and 20%, respectively (step 3 of figure 2).

The training set is the dataset used to adjust the various ML algorithms selected (step 4 of figure 2). A wide range of classification and regression algorithms is available11 (table 1), from more classic linear techniques such as logistic regression or linear discriminant analysis to more modern ones such as aggregation-type or bagging (random forest) algorithms and packaging or boosting (xgboost) algorithms. In addition, subsampling and oversampling techniques can be used when, as often occurs in health data, one category has many more cases than another. In theory, no algorithm is better than another; its ability to make a good adjustment will depend on the characteristics of the data (eg, number of variables, linearity, normality, missing values, and continuous or categorical variables).

Once the adjustment has been made, the validation data subset is used to evaluate the quality of the model (step 5 of figure 2). To do this, the aim is to maximize the metric of greatest interest in our particular case, such as area under the ROC curve, precision, sensitivity, and accuracy (figure 3). It is common for this training-validation process to be repeated a number of times while randomizing both subsets, which is known as k-fold cross-validation. The objective is to optimize the internal parameters of the algorithm used, evaluate the robustness of the model, and determine whether the model is subadjusting or overadjusting the data, trying to find a balance between the 2 scenarios.

Figure 3.

Metrics commonly used to evaluate supervised classification algorithms.

(0.2MB).

Once the final model has been constructed, the test data subset is used to verify that the final ML model behaves as expected with data that has not been used for its construction or validation (step 6 of figure 2). If this result differs from that obtained in the validation set, the dataset used for training is probably insufficient and should be expanded if a reliable estimator is needed before its generalized use (step 7 of figure 2).

For the implementation of ML, the most commonly used open source programming languages are currently Python and R. Both platforms can avail of libraries such as scikit-learn (Python) and caret (R) with implementations of the most widely used techniques and algorithms.

Image analysis is the field with the most rapid advances in AI and is therefore highly relevant in cardiology. The daily analysis of various cardiological images can be tedious and time consuming. However, our day-to-day work can already be improved by the various tools, based on NNs, available for automatic image processing.

NNs3 are a type of AI algorithm that is analogous to the learning process occurring in the neurons of the brain. Since its development as a computational model, training algorithms and network architectures have emerged that have considerably improved learning accuracy and efficiency, functioning with ever smaller amounts of training data. When these architectures consist of numerous layers of neurons, the term DL is used. Currently, the most commonly applied NNs include convolutional NNs, recursive NNs, generative adversarial networks, and U-nets, each with different uses and architectural subtypes. NNs are also very flexible and can be used in supervised, unsupervised, and reinforcement learning contexts. For their implementation, various open source platforms are available, such as TensorFlow, Pytorch, Keras, and Caffe. The disadvantages of DL techniques include their high computational cost. In addition, a high degree of expertise is required for their correct adjustment and, in the case of supervised learning, a set of manually annotated images is required, which can sometimes be extremely expensive.

Within the field of cardiac imaging, AI techniques have several objectives. These include segmentation and identification of the different structures of the heart (figure 4),12,13 classification of images cataloged with different conditions, lesion detection and segmentation, image registration, and classification of tissues from histological images. Other tasks are related to the generation of artificial images that are as realistic as possible.

Figure 4.

Example of automatic segmentation and identification of the left and right ventricle through deep learning performed in our department from images obtained with a 1.5-T Philips Achieva resonance system.12 From original images (A), the free neural network13 was able to identify and segment the left and right ventricles (B, white color for the left ventricle, light gray color for the myocardium of the left ventricle, and dark gray color for the right ventricle).

(0.14MB).

Research applying DL techniques to cardiological imaging has burgeoned in the last 3 years. However, the use of DL techniques in medical imaging has made more of a mark in other disciplines, such as neurology and pneumology.14 Among the challenges likely to be explored in the coming years are combined analyses of different imaging sources, the incorporation of clinical data and medical reports, and the study of temporal changes in cardiac images.

The main characteristic of AI in terms of ML or DL models is that learning is based on the identification of patterns in datasets. This philosophy is simultaneously a strong point, because computers are extremely efficient and precise at finding such patterns when they exist, and a limitation, for several reasons. First, because the amount of data required to obtain an accurate model can be substantial. This can be a problem in medicine, where the implementation of automatic data collection systems is just beginning, such systems must meet established legal and ethical criteria, and some rare diseases inevitably have few studies. In addition, even with algorithms with good results for a dataset, ML and DL models suffer from an inability to correctly detect and classify cases that they have not previously seen. Along these lines, the reliability and quality of the data source are essential for an algorithm to be realistic and correct. If the dataset used to adjust the model is itself biased, the model may not be well generalized to other populations.15 This problem is nothing new in our daily work because we accept risk scales based on data from other populations, such as the Framingham score for cardiovascular risk, resulting in erroneous results that overestimate or underestimate the risk upon application to other populations.16 Initiatives are underway to standardize the requirements that must be met by an AI algorithm for its possible practical application.17 For all of these reasons, it seems clear that particular effort will be needed for the integration of datasets from different populations.

Another important limitation is the opacity and interpretability of the ML models, particularly the DL models. These techniques are used as “black boxes”, which are fed inputs to obtain an output, namely a prediction. Thus, ML offers us answers in the form of predictions, but not a biological explanation. Although there are various methodologies to interpret the results of a model and verify its correct functioning through analysis of the specific weights of parameters or variables and to highlight the most discriminating parts or the individual explanation of each prediction,18 it is not currently possible to know exactly why the more complex ML and DL models make a certain decision. This limitation makes it difficult to generate knowledge and to apply AI to an area as critical as medicine.

Although these 2 specialties have different origins, they also have common features, despite their distinct purposes.

The cornerstone of classic statistics is inference. In inference, the observations in a sample are extended to the entire population, generally creating a mathematical model that defines the relationship among variables. The strength of ML lies in a prediction based on the available information, without the need to know the mechanisms relating the variables to each other. These characteristics are not exclusive because, to a greater or lesser extent, both disciplines use inferential techniques to improve their results (ML) or predictive algorithms to corroborate inferential conclusions (statistics).

Compared with statistical techniques, those of ML do not require prior suppositions about the variables, have the ability to handle cases with missing data, and improve their reliability when large volumes of data are available by detecting complex relationships among variables.

When AI is applied to the world of medicine, issues can arise that are not strictly related to the quality of the ML and DL algorithms. These are connected to legal problems associated with the automatic processing of personal data or how to apply ML in day-to-day practice.

Practically all AI techniques studied to date require a certain amount of data for the training and validation of predictive models. In our case, the information used is particularly sensitive because it generally involves patients’ personal and clinical data. Therefore, the proper use of this information from a legal point of view is a critical aspect.

Personal data protection is a fundamental right enshrined in the Spanish Constitution. However, such data may be used under certain conditions if a service is being provided to society. If both the organization collecting the information and that processing the data are fiscally resident in Spain, they must, since 2018, comply with Organic Law 3/2018 of December 5th, 2018, on the Protection of Personal Data and the Guarantee of Digital Rights.19 According to this legislation, the use of personal data is allowed for statistical purposes, as long as the results are aggregated. It is also necessary to adopt measures to anonymize and protect these data, and a number of rights of the individuals concerned is recognized, such as to be informed, have access, and be allowed rectification. Therefore, given that the use of cloud computing services is very common in AI due to their flexibility and greater computing power, the largest providers have adapted their data-processing policies to provide a completely secure service. In addition, these companies adhere to the CISPE (Cloud Infrastructure Services Providers in Europe) Code of Conduct to guarantee that their data protection standards comply with the legislation. Likewise, in situations where an AI model is used as a diagnostic or decision-making tool, the patient must be informed and provide consent.

The subtle discrimination inherent in the provision of health services can also be a problem when AI models are being developed and applied.15 For example, age is considered when scarce resources are being rationed, such as heart transplants. Such a consideration can lead to self-fulfilling prophecies: if physicians withdraw care for patients due to their advanced age, ML systems may conclude that the care of older patients is always fatal. On the other hand, it is also possible for ML models to help to resolve disparities in the provision of medical care if algorithms can be constructed to compensate for known biases or identify pressing areas of research.

There is broad discussion about how AI will affect the workflow of medical staff. AI might not eliminate jobs, but rather displace medical practice tasks. Routine and tiring jobs could be performed by machines to free up time and allow medical professionals to carry out more complex and sensitive tasks. However, the use of automatic diagnostic tools can lead to problems of bias in decision-making and in the assignment of responsibility in the case of error. Physicians may be inclined to review a diagnostic test in a more relaxed way if an algorithm has previously returned a negative result. The design of decision-making systems that involve both machines and humans is a crucial aspect for AI, and there are intermediate models between conventional clinical practice and fully automated systems, each with different characteristics of cost and time efficiency, risk, and interpretability.20 The implementation of these systems will be an area with a fundamental role for clinicians as validators.

Examples of AI using ML or DL are becoming more and more common in cardiology.21 This section presents the evolution of the contributions of AI to the different application areas of cardiology (figure 5).

Figure 5.

Temporal changes in the publications indexed in PubMed on artificial intelligence, machine learning, and deep learning according to the area of interest in cardiology. The details of the publications by area of interest and the search methodology are described in the supplementary material.

(0.07MB).

The supplementary material for this article includes a list of the contributions of the techniques described in several areas of cardiology in the last 5 full years and from January to March 2019, as well as a detailed description of the publications that we consider of most interest and that will help to explain the ML and DL models (table 2).22–40 Many of the publications reviewed are the result of common initiatives in AI, such as challenges promoted in conferences24,41 and datathons. In datathons, which are public and open competitions, people interested in the subject, from beginners to experts, work together with common databases to achieve an established goal. Participants must find a solution to a defined objective in the datasets provided. The most accurate and creative solutions are awarded a prize and made publicly accessible.

Although AI is often viewed as a futuristic and distant concept, the truth is that this technology is already in use in all types of areas, including cardiology. Due to the digitization of large amounts of data, the development of ML algorithms, and improvements in computer power in recent decades, AI can provide excellent opportunities for task automation, the application of precision medicine, or research progress through the detection of complex patterns in medical databases.67 A particular case is that of medical image analysis, with DL techniques having undergone a true revolution and their application to the area of cardiology already yielding excellent results. However, there is still a long way to go before these techniques can be widely applied to clinical practice. Essential prerequisites are large databases with high-quality information and the evaluation and integration of AI into realistic clinical contexts; therefore, the understanding of AI and its applications in our field is essential for the present and future development of cardiology.

This work was funded by the Carlos III Health Institute (Ministry of Science, Innovation, and Universities) within the CIBERCV, through which P.I. Dorado-Díaz and J. Sampedro-Gómez were hired.

V. Vicente-Palacios works for Philips Healthcare Ibérica. The remaining researchers have no conflicts of interest to declare.