ABC | Volume 114, Nº4, April 2020

Marques et al. Artificial Intelligence in Cardiology Arq Bras Cardiol. 2020; 114(4):718-725 Review Article threshold-based rules. Genetic algorithms were used to select the most appropriate characteristics to be used in the work. In relation to the detection of atrial fibrillation, an F1 score (harmonic mean of positive predictive value and sensitivity) of 0.81 was obtained. 27 Stuckey et al., 28 used the Cardiac Phase Space Tomography Analysis - a pioneering method that dispenses with the use of radiation and contrast, as well as performing exercises or pharmacological stress - combined with ML models (e.g., genetic algorithms) to analyze the thoracic phase signals. In this study, the authors used this tool to evaluate patients with coronary disease and chest pain who were referred by the physician for angiography. 606 patients were studied, and the results showed sensitivity of 92%, specificity of 62% and predictive value of 96% for coronary disease. 28 e) RandomForests (RF): Samad et al., 29 analyzed a database consisting of clinical and electrocardiographic variables to evaluate survival in 10 different periods of time (ranging from 6 to 60 months), considering a total of 171,510 patients. RF was used, with excellent results, better than those obtained through traditional scores such as the Framingham risk score and ACC/AHA guideline score. The area under the curve (AUC) was superior to 0.82. 29 Ambale-Venkatesh et al. 30 used information from noninvasive tests, questionnaires, biomarkers and imaging tests from 6,814 patients to construct 739 variables (features) in order to apply a variant of RF – called survivor random forests 31 – for predicting cardiovascular events (all‑cause death, stroke, all cardiovascular disease, coronary heart disease, atrial fibrillation and heart failure), having performed better than established risk scores, e.g., MESA-CHD, AHA/ASCVD and Framingham, with increased prediction accuracy (decreased Brier score by 10%-25%). 30,31 f) K-means: Cikes et al. 32 used a database consisting of clinical variables and echocardiographic parameters for which twomodels of ML, Kmeans and Multiple Kernel Learning were applied, in order to categorize the patients into mutually exclusive groups to evaluate the response to resynchronization therapy cardiac. A total of 1,106 patients were analyzed and four disjoint groups were identified, two of them with the best response to therapy. 32 g) Artificial Neural Networks (ANN): Kwon et al., 33 in a multicenter study of 52,131 patients, constructed a deep learning-based early warning system capable of predicting the occurrence of cardiac arrest in a hospital. Themodel showed high performance when compared to traditional track-and-trigger systems. The area under the curve was 0.82. 33 Rubin et al., 34 had promising preliminary results with the use of neural networks with convolutional architecture to evaluate electrocardiographic signs and to classify them in atrial fibrillation, sinus rhythm (normal) or noise – the F1 score achieved was 0.82. 34 Zhang et al. 35 also used convolutional neural networks to analyze a database with 14,035 echocardiographic exams to detect the presence of diseases such as hypertrophic cardiomyopathy, cardiac amyloidosis and pulmonary arterial hypertension with a high performance: C statistics were respectively, 0.93, 0.87, and 0.85. 35 Nakajima et al. 36 used an ANN to evaluate the presence of coronary disease after performing myocardial scintigraphy. Results were obtained with high accuracy and superior performance to the traditional scores used. For example, the AUC for patients with oldmyocardial infarction based on defects in rest stage was 0.97. 36 h) Gradient Boosting (GB): Mortazavi et al. 37 used GB for prediction of risk of bleeding after percutaneous coronary intervention and demonstrated that these tools can help to identify patients who would benefit from strategies aiming to reduce the bleeding risk. A total of 3,316,465 procedures were analyzed and a C statistic of 0.82 was obtained. 37 Hernesniemi et al., 38 also proposed a GB to predict mortality in acute coronary syndrome, analyzing 9,066 consecutive patients. The AUC was 0.89 and the model performed better than GRACE traditional score. 38 It is important to note that when using any ML model, one should keep in mind a major problem that may arise, called overfitting. It occurs when a model describes the examples very well (training subset) and performs poorly when applied to other instances of the same phenomenon. 39 In addition, it is worth saying that there is no theoretical result that ensures that any of the AI algorithms is better than the others in any application. Thus, this choice depends on several variables, such as the nature of the problem under analysis, the time and resources available to solve the problem. The combination of techniques generating hybrid models can also be of great value. On the other hand, the use of tools for parallel processing, such as the Graphic Processing Unit (GPU), has been of great value in improving the performance of ML models, especially in relation to computational time needed to run them. Challenges and future prospects As previously highlighted, AI applications in cardiology have increased greatly in recent years and their growth potential is enormous. However, this scenario brings with it the need to overcome some challenges, such as: ethical limits of use (misuse), improvement of mathematical knowledge, acquisition of healthy data, development of security, need for collaboration, attention to errors and data-based care. All of this is discussed below and is summarized in Figure 2. a) Challenge 1 – ethical limits of use (misuse): like all disruptive technology, the limits of ethics need to be rethought and widely discussed. ML algorithms can be misused and misleading. As an example, a work of great repercussion was published by Wang and Kosinski (2018). The authors used deep learning and obtained expressive results in the prediction of whether an individual is gay or not from a database of images of the study participants’ faces. 55 Similarly, the same AI algorithms can be used to detect, for example, whether or not a patient will develop atrial fibrillation or any future cardiomyopathy. Could this information be used by companies to increase the amounts of their health plans or even deny membership to the plan due to a high cost? What if it is detected that a baby will be born with congenital heart disease due to the analysis of the genetic, clinical-laboratory and image (or other) data of its parents? This could open space for a kind of neoeugenia. This debate has gained an additional emphasis with the emergence of the CRISP-Cas9 technique, which allows DNA editing. 56 In this context, by stimulating a debate with society on the subject, transparency and regulation are fundamental pillars to be preserved. b) Challenge 2 – improve math knowledge: the advent of this new kind of unbelievable human being ( Homo incredibile ), which supports its decisions in data, carries with 721