Cluster analysis of patients’ states performed in order to develop treatment strategies for patients with atherosclerosis

Abstract

The article describes an approach to the implementation of the initial stage of solving the problem of finding and prescribing optimal treatment strategies using reinforcement learning. The approach involves the identification of the main groups of conditions of patients with diagnosed atherosclerosis by means of cluster analysis. The MIMIC-III database containing the clinical, laboratory, hemodynamic, and other data of patients was used as the initial data set. The main cluster analysis method used in the study was the k-medoids algorithm. The quality of clustering was assessed by means of silhouette analysis. At the preliminary stage of clustering, we reduced the dimensionality using principal component analysis (PCA). The results were visualized using the t-SNE method. An important part of the study was the calculation of the severity of the patients’ conditions for each of the identified clusters. The resulting estimates were then used to calculate the rewards in the model for assigning optimal treatment plans by means of reinforcement learning. The set of obtained clusters determines the set of the environment states. Thus, the clustering results allowed us to identify the main patterns in the initial dataset and to obtain the main components of the reinforcement learning model for prescribing optimal treatment plans for atherosclerosis.