ABC | Volume 114, Nº4, April 2020

Marques et al. Artificial Intelligence in Cardiology Arq Bras Cardiol. 2020; 114(4):718-725 Review Article Table 1 – Brief description and classification of the main ML tools Tool Description Learning SVM It is useful for two-group classification problems. The idea is to find a function called hyperplane from the resolution of a linear system built from the various lessons of the training subset. 40 This hyperplane is used to cluster the lessons of the test subset into two disjoint groups. Supervised NB It was inspired in the studies of the reverend Bayes on conditional probability. 41 These probabilities are used to identify the category (out of a total n possible) that a particular lesson belongs to. 42 Supervised KNN It is said that a vector norm is a mathematical function, which satisfies specific properties, and associates a vector with a value greater than or equal to zero. 43 The norm of the difference between two vectors is the distance between them. The KNN uses a norm to calculate the distance between all the vectors (lessons) that make up the database. Then, for each vector of the database, the k vectors closest to it are determined. The inclusion in a given group is obtained from a majority voting system among the neighbors. 44,45 Supervised AG Algorithms inspired by the biological evolution of species, in which each possible candidate to solve the problem is modeled as a chromosome consisting of a set of genes, which during the execution of the algorithm undergoes operations of crossing-over and mutation in order to obtain better solutions than the current ones. 46 This way, they allow a database to be separated, for example, into two distinct groups – which have or do not have a particular characteristic. Supervised RF This method is based on the construction of several decision trees. The first step is to get several random samples (with reposition) of lessons to build other databases, a process that is called bootstrapping. Each of these new databases will give rise to a decision tree, which is obtained iteratively, from a subset of variables (features). After the construction of all trees, a new lesson in the database should be allocated to the group that has the largest number of decision trees, showing that it belongs to this group (majority of votes). 47,48 Supervised K-means It allows partitioning a database into k groups with similar characteristics. To do so, it is necessary to update, in an iterative way, a set of vectors, called reference centroids of each group and to calculate the distance of each lesson to each one. A lesson is always allocated to the centroid for which it has the shortest distance. The elbow chart is generally used to determine the ideal number of groups to separate from the database. 49 Unsupervised ANN Inspired in biological nerve systems, a structure called a graph - a set of nodes and edges - is used in which nodes are layered and connected by valued edges, which represent a weight assigned to a given connection. The idea is that from a set of inputs, these weights are used properly to produce an output. Several architectures have been proposed for neural networks, from simpler ones such as the perceptron, to more sophisticated ones, such as the radial basis function, convolutional networks and deep learning. In deep learning, in addition to the input and output layers, there are hidden layers that increase significantly the number of weights to be updated and often require huge computational efforts. Convolutional network is a type of deep leaning inspired in visual cortex of animals that have an important role in image analysis. Autoencoders and Kohonen neural networks are examples of unsupervised learning. 1,7,50-52 Unsupervised or Supervised GB It is a tree-based method that uses gradient, vectors related to the direction of maximum increase in a math function, to produce sequential decision trees to be combined to brush up on the prediction. Variants of this approach include Stochastic Gradient Descent that incorporates a random subsampling to GB. 53,54 Supervised Some artificial intelligence tools and applications Currently, there is a multiplicity of models of ML each of them with diverse particularities, varied uses and limitations. The applications of some of these models in Cardiology are explained in the following paragraphs, while a brief description of each of them and their type is shown in Table 1. a) Support Vector Machine (SVM): used by Samad et al., 22 to predict with success the deterioration of ventricular function in patients with repaired tetralogy of Fallot from a database of 153 patients with clinical, electrocardiographic and cardiac magnetic resonance imaging data. In relation to predicting any deterioration (minor or major) vs. no deterioration, the mean area under the curve (AUC) was 0.82 ± 0.06. 22 Berikol et al. 23 used clinical, laboratory (troponin I and CK-MB levels), ECG, and echocardiographic data from 228 patients who presented at the emergency department with chest pain for classification regarding the presence or absence of Acute Coronary Syndrome. Accuracy, sensitivity and specificity were, respectively, 99.19, 98.22 and 100%. 23 Betancur et al. 24 also used SVM to more precisely define mitral valve plane (VP) positioning during left ventricular segmentation in Single-Photon Emission Computed Tomography (SPECT) exams. Images of 392 patients were analyzed and the good results obtained were compatible with the opinion of experts in the area – AUC: 0.82 [0.74-0.9] for regional detection of obstructive stenosis and ischemic total perfusion deficit areas. 24 b) Naive Bayes (NB): Paredes et al., 25 used an NB fusion and genetic algorithm to predict the risk of occurrence of cardiovascular events (e.g., hospitalization or death) based on data from 559 Acute Coronary Syndrome-Non-ST Segment Myocardial Infarction (ACS-NSTEMI) patients. Sensitivity and specificity were, respectively, 79.8, 83.8. 25 c) K-nearest neighbors (KNN): Al-Mallah et al. 26 compared the prediction of all-cause mortality in 10 years between the classical logistic regression model and the KNN, considering a database of 34,212 patients with clinical information and information obtained after the treadmill test using the standard protocol of Bruce. 26 The results obtained by this ML tool showed a sensitivity of 87.4% and specificity of 97.2%, better than the predictive performance of the traditional Atherosclerosis Cardiovascular Disease Risk Score (ASCVD). d) Genetic algorithms (GA): Smisek et al. 27 developed a wearable device to detect arrhythmias from the information record of a single-lead electrocardiogram. The data were analyzed from a combination of the (SVM), decision tree and 720