New approaches for predicting outcomes in patients with atrial fibrillation

Aim: we aimed to assess the capabilities of “machine learning” methods in predicting remote outcomes in patients with non-valvular atrial fi brillation (AF).Methods. From 2015 to 2016 234 patients with non-valvular AF were included in the study (median age 72 (65; 79) years; 50.0% men). During the median follow-up of 2.9 (2.7; 3.2) years 42 patients died, 9 patients had non-fatal acute cerebral circulatory disorders and 3 patients had non-fatal myocardial infarction (MI). These events in 52 subjects (22.2% from all patients included) were combined into a combined endpoint (death and a nonfatal cardiovascular accident at the stage of remote observation). The first 184 patients comprised a “training” group. The next 50 patients formed the “test” group. The following methods of «machine learning» were used in the analysis: classifi cation trees, linear discriminant analysis, the k-nearest neighbor method, support vectors method, neural network.Results. Long-term outcomes were influenced by age, known traditional risk factors for cardiovascular diseases, the presence of these diseases, changes in intracardiac hemodynamics and heart chambers as evaluated by echocardiography, the presence of concomitant anemia, advanced stages of chronic kidney disease, and the administration of drugs associated with a more severe cardiovascular disease progression (amiodarone, digoxin). The best prognosis was created using the model of linear discriminant analysis, the complex neural network model, and the support vector machine.Conclusion. Modern methods aimed at prognosis estimation seem to be of importance in cardiology. These methods include big data analysis and machine learning technologies. The methods require further evaluation and confirmation, and in the future they may allow correcting cardiovascular risks, using data from real clinical practice and evidence-based medicine at the same time.

Atrial fibrillation (AF) is among the most common cardiovascular diseases [1]. The management of AF is described in Russian and foreign clinical guidelines [2][3]. Numerous methods are known for assessing the risk of cardiovascular accidents and treatment complications in AF patients (CHA 2 DS 2 -VASc, HASBLED, ATRIA scores) [4][5][6].
Most of the available risk scales are designed to assess the likelihood of complications. However, few scales affect clinical and anamnestic, demographic indicators, concomitant diseases and the therapy ta ken. In recent years, novel methods for predicting risks associated with the use of so-called. «machine learning» [7]. The combination of these methods allows us to quickly analyze large amounts of data to assess the risk of a specific event in a particular patient. These methods are successfully used in genetics, oncology, endocrinology [8][9][10].
This article will describe several examples of the use of such methods in a small study evaluating longterm outcomes in patients with a non-valve form of atrial fibrillation.

MATERIALS AND METHODS
On the basis of the University Clinical Hospital №1 of I.M. Sechenov First Moscow State Medical University (Sechenov University) for the period 2015-2016 included 234 patients with non-valvular AF, median age 72 (65; 79) years, 50.0% of men and women. The median follow-up was 2.9 (2.7; 3.2) years, during which time 42 patients died, 9 non-fatal acute cerebrovascular accidents and 3 non-fatal myocardial infarction (MI) occurred. These events in 52 patients (22.2% of all included) were summed into a combined endpoint (death and non-fatal cardiovascular catastrophes at the stage of remote observation).
The first 184 included patients formed a "training" group, the next 50 patients -the "test" group used for validation purposes. For the "training" group, regression analysis (Cox proportional risk model) was carried out to estimate the development of the combined.
Comparative analysis was carried out between the "training" and "test" groups (U -Mann-Whitney test for numerical data and Fisher's exact test for categorical data) to make sure that there are no significant intergroup differences.
Several "machine learning" methods were used for creating various predictive models. Factors, significant in Cox regression model, were used for model training. After-wards, the models predicted outcomes for each patient of the "test" group, thus validating the models. Because the outcomes in the "test" group were known to researchers; a comparison was made of the true results and the results predicted by the models.
The accuracy, positive and negative predictive values, Cohen kappa were calculated to assess model quality. was calculated as an independent assessment of the quality of the model. ROC analysis was carried out for a number of models.
Modeling was performed using the R language v. 3.5.1 [11].

RESULTS
The study included 234 patients with a non-valvular form of atrial fibrillation, 50% of them were men. Median age was 72 (65; 79) years. A high prevalence of cardiovascular risk factors was noted: 11.5% were smokers, 92.7% diagnosed with arterial hypertension, 24.4% had type 2

analysis between groups, assessment of factors significantly influencing on the "training" group prognosis (Cox regression analysis)
Description: p # -significant differences between groups, CI -confidence interval, p $ -significant of prognosis, CADcoronary artery disease, CHF -chronic heart failure, CKD -chronic kidney disease, CVD -cardiovascular disease, GFR -glomerular filtration rate, MDRD -Modification of Diet in Renal Disease Study, LA -left atrium, RA -right atrium, SPPA -pulmonary artery systolic pressure, * -at discharge JOURNAL OF ARRHYTHMOLOGY, 27(E), 2020 diabetes mellitus. Dyslipidemia was detected in 61.5% of patients. Among the whole group, cardiovascular risks and vascular catastrophes were widely recorded. A history of coronary heart disease (CHD) was diagnosed in 26.9% of patients, while 20.9% had a history of prior MI. In 41.0% of patients, symptoms of heart failure (HF) were noted, 13.7% previously had stroke, 5.1% had transient ischemic attack. Median score on CHA₂DS₂-VASc was 4 (3; 5) points, on the HASBLED -2 (1; 2) points.
The first group consistently included 184 patients (78.6%) in the study group, the next 50 patients (21.4%) were included in the test group. Many works show the initial ratio of the data of «training» to the «test» data of 70.0% / 30.0% to 80.0% / 20.0% [7]. Also, the separation in accordance with the sequential switching time brings the experiment somewhat closer to real practice in the form of patients who are sequentially hospitalized or who came to see a doctor.
For the "training" group, a one-dimensional Cox regression analysis was performed to identify factors that significantly affect the development of the combined endpoint during remote observation. These factors, as well as the relative risk value, are presented in Table 1: Long-term outcomes were influenced by age, changes in intracardiac hemodynamics by echocardiography, the presence of concomitant anemia and more severe stages of CKD, as well as prescriptions associated with more severe disease (prior amiodarone, digoxin on the long-term use at discharge).
According to these factors, a comparative analysis was carried out between the groups of «training» and «test», the results of which are given in table. 2: Generally groups were comparable, with the exception of age (where differences were noted on the verge of significance). It is important to note that there were no significant differences either in the frequency of the combined endpoint or in the cardiovascular anamnesis.
Brief description of the methods used for "machine learning" Classification trees Trees combine regression and classification methods splitting of values and finding optimal threshold factor by factor. The result of numerous splitting is ultimately the classification of the object of interest Schemes of classification trees are quite clear and understandable both by researchers and doctors There are several algorithms for constructing classification trees (Random forest, building trees using boosting, and many others) [12].
Linear discriminant analysis Methods underlying the linear discriminant analysis are somewhat similar to analysis of variance (ANOVA) and are associated with the search for best linear relationships between predictors. The analysis is sensitive to group size. The resulting linear interactions are somewhat reminiscent of linear regression analysis.  Table 2.

Evaluation methods predictive value classification
Description: PPV -positive predictive value, NPV -negative predictive value, 95% CI -95% confidence interval a b c

Figure 2. ROC curves for the most accurate forecasting methods in the study: a -linear discriminant analysis, bsupport vector machines, c -multilayer perceptron.
JOURNAL OF ARRHYTHMOLOGY, 27(E), 2020 The method of k-nearest neighbors This method is close to cluster analysis when setting the number of outcome classes (in our case, there are two of them: the patient who has not reached the combined point and the patient who has reached the combined point). Each factor included in the analysis (for example, the age of the patients) is split according to the given class in terms of its average values and the distance between the average values in each class. Factors of interest are splitted thresholds are computed. Thresholds correlate with one of the predicted out come classes.

Support vector machines
The method is associated, on the contrary, with the maximum contrasts of 2 or more classes using linear and nonlinear methods. The method aims to find the largest possible separation between classes.
Neural networks Neural networks are based on non-linear programming algorithms. The construction of neural networks occurs in analogy with the work of neurons of the nervous system, with the presence of signals, synapses, etc. The simplest neural networks solve classification problems (simplest option is a binary classification) of an object in accordance with its characteristics. Due to the logistic functions used, the "strong" and "weak" characteristics of an object can be equalized in terms of their influence on the classification decision. The size of the neural network determine the resource consumption for its creation and operation. Usually there are "hidden» nodes structures with complex nonlinear transformations. The construction of such a network is relatively slow.
Using several layers of «hidden» nodes, allow to use nonlinear methods for more flexible and complex classification problems.
A schematic representation of the neural network used in the trial is shown in Fig. 1. Fig. 2 on the left shows the input nodes with the main patient factors, on the rightthe possible predicted outcomes -"1", when the patient is likely to develop a combined endpoint, and "0" when the patient's state does not change. In the middle we see 3 "layers" of 6, 4 and 3 neurons. These neurons are connected to each other by the so-called coefficients -"weights" (blue and black numbers), in some way reminiscent of coefficients in regression equations. The neuron itself is a complex non-linear function that calculates the probability of a particular outcome.

Assessment of the prognostic value of classification methods
Assessment is made on the predictive accuracy of the model, positive and negative predictive value, Cohen kappa coefficient. The results are presented in table. 3: According to the value of 95% CI, the models of linear discriminant analysis, complex neural network, and support vector machines showed the best quality. Based on the Cohen's Kappa, the most accurate was the linear discriminant analysis model, followed by models using the support vector machines and a complex neural network.
To assess the sensitivity and specificity of the methods, ROC analysis was performed with the calculation of the area under the curve (AUC). In fig. Figure 2 shows some ROC curves of the most accurate methods:

Scientific and practical novelty
The aim of this trial was to demonstrate the possibilities of using machine learning methods in predicting long-term outcomes using typical clinical, medical, demographic characteristics of patients.
This approach will allow us to use routine examination data that the cardiologist enters in the electronic medical history, without the use of complex special scales and risk calculation techniques that require additional time from the doctor. An extremely important aspect is that the cardiologist receives information not about some abstract cardiovascular risks, but about the risk of a specific outcome or combination of outcomes in a given patient.
The scientific novelty lies in the fact that the forecasting methods are rarely used in the intersection of cardiology, epidemiology and practice. Classification prognostic models are usually built for a narrow specific problem [13]. At the same time, there can be wide perspectives for making prognosis in patients in "grey zones" of clinical guidelines. Of course, at the moment, comprehensive verification of machine learning in cardiology is needed.

DISCUSSION
Due to the fact that cardiovascular diseases are still among the three main causes of disability and mortality worldwide, the analysis of predicted outcomes is extremely important for the patient's life, physical condition and social activity. In articles where similar forecasting methods were used, it is carefully emphasized that the prognosis should be achieved by several independent modeling methods to reduce the likelihood of false results. Factors that influenced the risk of adverse long term outcomes in this study are not unexpected (for example, more than half of those included in the history of cardiovascular accidents), and have been described in modern clinical guidelines for many years.
The analysis above showed that machine learning methods show good results with respect to prognosis. The limiting factor was a small amount of the "training" group (184 patients), the choice of endpoints (mainly the combined point without analysis of the causes of mortality). Larger sample and shorter periods can theoretically increase the accuracy of the forecast.
Currently, the abilities of large medical centers and hospitals allow aggregating a large amount of patient data, introducing electronic document management, which can serve as a "playground" for testing and implementing such methods. An important aspect is the ability of the model to provide assessment of individual risks, taking into account many factors that significantly affect the forecast.

CONCLUSION
Evaluation of long-term outcomes in patients with AF appears to be an extremely important task due to the high prevalence of the disease and the severity of complications. Modern methods aimed at predictive assessment, using big data and machine learning technologies, represent great potential for cardiology. These methods require further critical confirmation, as in the long term, they can allow correcting cardiovascular risks, using both the data of real clinical practice and the concept of evidence-based medicine.