Detection of Cardiac Arrhythmias Using Unsupervised Learning: A Preliminary Approach Based on PAM and CLARA Clustering Algorithms

This article presents a preliminary analysis of an ongoing study using unsupervised learning methods to detect cardiac arrhythmias. We carried out this study on ECG recordings from a data set with 10,646 records classified into 11 heart rhythms and grouped according to the suggestion of cardiology specialists into four groups based on the characteristics of each of the defined rhythms. By grouping arrhythmias, it is possible to identify common patterns, distinctive features, and similarities among different cases. We obtained two clustering models over four and seven groups, applying Partitioning Algorithms Around Medoids and Clustering Large Application. We used external evaluation metrics to compare these groups with those defined in the dataset and proposals made in the state-of-the-art review. Also, we evaluated the models with internal evaluation metrics.