POSSIBILITIES OF USING ARTIFICIAL INTELLIGENCE TO PERFORM ECG ANALYSIS DURING POPULATION SCREENING

Abstract

The aim of this study is to discuss the evolution and possibilities of modern methods of electrocardiogram analysis, that is, methods based on the mathematical transformation of the ECG signal, first of all a modern method of machine learning, which is called the anomaly detection method. Materials and methods: Five different groups were examined (in total 1211 persons): patients with proven coronary heart disease, military personnel with landmine and explosive injuries, those who suffered from COVID-19, as well as two subgroups that involved participants of a large-scale screening in one of the rural areas of Ukraine. Subgroup 1 consisted of persons, who died during five-years follow-up (all-cause mortality), subgroup 2-persons, who didn’t die during this period. Control group consisted of 181 people (males, aged from 18 to 28) is used in this study. Each ECG can be presented as a vector in the 204-dimensional feature space. In the case of having the group of ECG with similar characteristics, the corresponding feature vectors will form in the cluster in the space. If the particular ECG is located far from the cluster, this might indicate that their features are distinct from those of the cluster members. The vector of ECG which is similar to the group of ECGs forming the cluster will be located within the cluster. The concept of outlier/inlier is proposed to be used for detecting the deviations of the ECG from the group of other ECGs. To define whether the particular ECG is an outlier or not, the Isolation Forest anomaly detector is used. The negative values of the anomaly score indicate that the ECG is an anomaly; this is interpreted as the substantial deviation of the ECG from the norm. Results. When estimating distance between the studied groups and normal controls it was found that the largest distance takes place between healthy volunteers’ group and CAD patients group and group of subjects who died within 5 years of follow-ups (all-cause mortality). COVID group is in an intermediate position. The minimal distance from NC was detected in the Combatants group. Conclusions. The high sensitivity of the proposed machine learning algorithm based on Isolation Forest anomaly detection to detect a small pathologic changes in the electrocardiogram was demonstrated. The further large-scaled study is planned.