Thorough study of the cardiac arrhythmia by applying the methods of statistical analysis, machine learning and computational chemistry

Abstract

This thesis develops the phenomenon of cardiac arrhythmia, which is a condition with several subtypes; some of these types can be kept in control using pharmacological interventions with the help of a cardiologist, while others may lead to lethal outcomes. The prediction of arrhythmia was the main focus of this study, and by implementing machine learning methods combined with classification algorithms, it became possible to predict the events of arrhythmia. Therefore, through further statistical analysis, some indicators associated with the disease were derived, and as a result of the previously important recognition about the disease, their explanation and relation to the disease were examined. The dataset used, taken from Kaggle, consisted of 452 records and 280 features that tested whether the predictive model could detect one of the 15 types of arrhythmia. The results were good, with the best predictive model achieving an accuracy rate of 87.07%. Further studies on the disease thereafter began with the employment of the β2 adrenergic receptor, whose stimulation produces pathologies in both the respiratory and cardiovascular systems. The study aimed at finding those agents that act as agonists in deactivating the adrenergic receptor. Within this research, some preliminary computational experiments were performed using statistical analysis besides machine learning techniques to select the most relevant markers and evaluate their predictive capabilities. Then, in the molecular binding experiments, one certain branch of computational chemistry made an attempt to discover the binders from the library of natural ZINC compounds by the designated receptor known as 2RH1, which is known for its potent inhibiting activity upon binding to the receptor. The last attempt was made in the direction of combining the results obtained from the first two parts for the purpose of knowing whether candidate ligands combined with the major determinant points of the receptor act as agonists and how they can be identified in nature. The dataset consisted of 1470 records and 215 descriptors, with statistical analysis showing that 71 out of 215 descriptors had a statistically significant difference. In the machine learning approach, the prediction model created achieved an accuracy rate of 93.45%, with 7 of the 71 descriptors being the most suitable. Through the molecular docking experiments, out of the 2000 potential ligands screened from the ZINC database, only 10 were further investigated using the Maestro software. The 2 final ligands belong to the flavonoid and anthraquinone families respectively.