Left ventricular hypertrophy detection system based on electrocardiogram signal

Abstract

Abstract Background Left ventricular hypertrophy (LVH) is associated with increased risks of cardiovascular diseases. Electrocardiography (ECG) is generally used to screen LVH in general population and electrocardiographic LVH is further confirmed by transthoracic echocardiography (Echo). Purpose We aimed to establish an ECG LVH detection system that was validated by echo LVH. Methods We collected the data of ECGs and Echo from the previous database. The voltage of R- and S-amplitude in each ECG lead were measured twice by a study assistance blinded to the study design, (artificially measured). Another knowledge engineer analyzed row signals of ECG (the algorithm). We firstly check the correlation of R- and S-amplitude between the artificially measured and the algorythm. ECG LVH is defined by the voltage criteria and Echo LVH is defined by LV mass index >115 g/m2 in men and >95 g/m2 in women. Then we use decision tree, k-means, and back propagation neural network (BPNN) with or without heart beat segmentation to establish a rapid and accurate LVH detection system. The ratio of training set to test set was 7:3. Results The study consisted of a sample size of 953 individuals (90% male) with 173 Echo LVH. The R- and S-amplitude were highly correlated between artificially measured and the algorithm R- and S-amplitude regarding that the Pearson correlation coefficient were >0.9 in each lead (the highest r of 0.997 in RV5 and the lowest r of 0.904 in aVR). Without heart beat segmentation, the accuracy of decision tree, k-means, and BPNN to predict echo LVH were 0.74, 0.73 and 0.51, respectively. With heart beat segmentation, the signal of Echo LVH expanded to 1466, and the accuracy to predict ECG LVH were obviously improved (0.92 for decision tree, 0.96 for k-means, and 0.59 for BPNN). Conclusions Our study showed that machine-learning model by BPNN had the highest accuracy than decision trees and k-means based on ECG R- and S-amplitude signal analyses. Figure 1. Three layers of the decision tree Funding Acknowledgement Type of funding source: None