• QRS classification and clustering (research in progress)

Rhythm analysis is one of the most important features of Holter monitoring software. It allows to determine health-threatening conditions and helps to correctly diagnose pathological cases. Rhythm analysis is based on ECG structural components classification and clustering.

This research is aimed at QRS classification and in-class clustering by means of classical machine learning algorithms. The R&D pipeline is shown in Figure 1. The process requires MIT-BIH arrhythmia database. All QRS complexes are spread into four classes and for each of them set of 20 features (time and frequency domain) are extracted. The obtained dataset is subjected to exploratory data analysis in order to find out the optimal feature set. This reduced dataset is used in comparative study of classical machine learning algorithms. According to preliminary results, support vector machine classifier is able to deal with classification task better than other models.

Figure 1. R&D pipeline of QRS classification and clustering

The next step of the research is to determine an appropriate model for in-class clustering of QRS complexes. To achieve this, specifically prepared datasets are used in comparison of several clustering algorithms. According to the carried out comparison, the following aspects are determined:

> DBSCAN does not need prior knowledge about the number of clusters, however it assigns large amount of data points to noise group;

> k-means clustering needs prior knowledge about the number of clusters, which is not applicable in this specific case;

> agglomerative (hierarchy) clustering can be used either with the number of clusters or with distance between objects. The trained agglomerative clustering model can be used to select the required number of clusters.

Once the research is done, it is expected to deploy the selected ML models into Holter monitoring software.