Pointwise Estimation of Noise in a Multilead ECG Record

Abstract

Noise of certain characteristics can efficiently be extracted from a multilead physiological record. This paper presents a novel method for distilling random components of electrocardiogram. First, the algorithm calculates first principal component of the record as a noise-reduced reference. Next, a multiplication-addition network is trained and used for prediction of the value in each given lead based on remaining leads. Finally, the time series of prediction errors are decorrelated with the noise-reduced reference, what removes the geometry-related component of unexpectedness. In tests with poor electrode contact and muscle fibrillation noise patterns, the method shows static noise estimation accuracy of 80.3 % and 84 % respectively and in dynamic tests - the accuracy of 78.1% and 82.6%, what outperforms other known methods.