In Search of an Optimal FIR Filter for ECG Delineation

Representative beats (RBs) computed by combining ECG waveforms from multiple cardiac cycles with similar characteristics, are robust to noise and are often used in diagnostic applications for reliable ECG delineation.However, when delineation of individual beats (IBs) is required, the same algorithms need either retraining or denoising of the raw input ECG signals.This study aims to find an optimal finite impulse response (FIR) filter for 12-lead IBs that can reduce the errors of a specific ECG delineation algorithm trained for RBs.The FIR filter was implemented as 1D convolutional layer (same for all leads), no bias term and linear activation function.It was trained by backpropagation, minimizing the increased loss in the original ECG delineation output (P wave, QRS complex and QT interval) caused by larger noises in the training raw data of 12-lead IBs.The loss function included both a binary cross-entropy and an L2 regularization term.Using the PTB-XL and CinC Challenge 2011 databases, the FIR filter was shown to reduce the noise impact on the delineator's output in typical resting or ambulatory ECG monitoring environments.