Donets V. Methods and models of elements stratification of computer medical monitoring systems based on a multi-agent approach

Donets V. V. Methods and models of elements stratification of computer medical monitoring systems based on a multi-agent approach. – Qualification scholarly paper: a manuscript. The dissertation submitted for obtaining the Doctor of Philosophy degree in Informational Technology: Speciality 122 – Computer science. V. N Karazin Kharkiv National University, Ministry of Education and Science of Ukraine, Kharkiv, 2024. The dissertation is devoted to developing methods and models of data element stratification in medical monitoring computer systems using a multi-agent approach. Stratification is defined as the determining process of the possible patient states, their classification, and identifying the influence of state variables. This process includes data clustering, classification of patient conditions, and variable selection. The multi-agent approach refers to the elitist selection approach, which is implemented in the clustering method and consists of selecting the best clusters, which are agents in the space of generated data, according to a certain metric among the defined states. The first chapter contains an overview of existing research in medical monitoring computer systems, including systems based on fuzzy logic, machine learning, and deep learning methods. It was determined that automatic data analysis could improve the treatment quality, considering the limited number of specialists. This analysis allowed us to determine the study's purpose of improving the patient's condition diagnosing accuracy by developing stratification methods and models. The tasks of clustering, classification, and determining the state variables informativeness are necessary to solve for this purpose. Further, taking into account the specified study's goal and tasks, a stratification procedure was proposed, according to which a medical monitoring computer system model with a separate stratification subsystem was developed. The role of each module in the model of the medical monitoring computer system is explained, and the modes of operation of the stratification subsystem are defined, depending on the availability of information about possible states or their number. The second chapter describes the stratification subsystem components, namely methods of clustering, classification, and determining informativeness. A multi-agent method of fuzzy clustering is proposed to solve the data clustering problem. It is proposed to check the accuracy of its work using the classification method. Next, methods of training and hyperparameters configuration of an artificial neural network were shown to speed up gradient ascent and hyperparameters selection, which in general should improve classification accuracy. Methods for determining the general and current informativeness of state variables were also proposed. These methods solve the problems of determining the set of the most influential variables and reasons for decision-making in a computer medical monitoring system. The third chapter analyzes software tools for implementing stratification methods and models, including Python, associated libraries for data processing and machine learning, and the development environment. The datasets for validation are presented, including standard test datasets, medical monitoring data, and data for expanding functionality with economic monitoring data. The chapter concludes with a comprehensive methodology for verifying the developed software, allowing quality checks of both individual methods and their combination. The fourth chapter presents the practical application results of methods with medical monitoring data. General testing shows that the developed multi-agent clustering method forms target clusters with satisfactory accuracy. The developed method of learning and setting ANN model hyperparameters leads to high classification accuracy on both clustered and original data. The developed method for determining general informativeness compares favorably with existing methods but has a more linear nature for selecting informativeness weights. The modified method of integrated gradients accurately assesses the influence of certain input variables on ANN model classification results. This proves the method's applicability for justifying decisions in the medical monitoring system. The possibility of expanding the application of methods based on economic monitoring data has been identified. Also, at the end of the chapter, practical recommendations are given for using the developed methods and the stratification subsystem as a whole in the computer system of medical monitoring.