Premature Ventricular Conduction Detection and Localization From the ECG Using a Neural Network

Alexander Vieira Pereira1, Peter van Dam2, Roger Abächerli1
1HSLU, 2Arrhythmia Center, UCLA


Abstract

Objective: The localization of a premature ventricular conduction (PVC) is crucial in an electrophysiological exam. A pre-exam localization can be time saving. The PVC localization from the ECG using an automat-ic algorithm has not fully been addressed. We aim to detect the PVC and localize its anatomical origin from the ECG using a neural network (NN). Material and Method: We used 59 clinical 12-lead ECG recordings. We divided all recordings in consecutive 10s segments with annotation of a PCV presence forming the first NN-target. The ECGs were processed by an algorithm designed in the ALVALE research project. Its localization output was used as a second NN-target. The 466 10s 12-lead ECG seg-ments formed the NN input. A 65% learning, 35% verification and 5% test portion was defined. The NN-toolbox of Matlab (Mathworks, USA) was used for designing the NN. Results: The found NN consisted of two layers. The first layer needed 40 neurons while the second layer needed three neurons corresponding to the three outputs respectively targets called right ventricle (R), left ventri-cle (L) and no PVC (noPVC). 238 files contained a PVC. The correspond-ing detection sensitivity was 88.0% while the specificity was 87.6%. 165 files contained a PVC with an origin from the right ventricle. The corre-sponding locations sensitivity was 77.6% with a specificity of 95.7% while the localization sensitivity for the left ventricle was 55.6% with a specificity of 99.3%. Conclusion: It is possible to design a NN with only two layers being able to detect the PVCs from a 10s 12-lead ECG and localize its anatomi-cal origin. More NN layers might be added to improve the detection and localization performance. Further an adjustment of the NN to the clinical need (focus on high sensitivity or high specificity) will be performed in future.