Convolutional Autoencoder for Filtering of Power-Line Interference with Variable Amplitude and Frequency: Study of 12-Lead PTB-XL ECG Database

Abstract

This study aims to explore a new deep learning strategy for electrocardiogram (ECG) denoising under adverse conditions of non-stationary power-line interference (PLI) with amplitude changes or nominal frequency deviations. The study presents an exhaustive training strategy of deep convolutional autoencoder (CAE), while input with one of the largest PhysioNet 12-lead ECG databases contaminated by simulated sinusoidal PLI noise with augmented settings. Twelve ECG leads (I, II, III, aVR, aVL, aVF, V1–V6) from 14890 PTB-XL records, divided patient-wise to training (50%, 7441 records), validation (20%, 2979 records) and test (30%, 4470 records) are superimposed by PLI with five signal-to-noise ratios (SNR) (–2.5, 0, 2.5, 5, 7.5 dB), nine frequencies (48, 48.5, 49, 49.5, 50, 50.5, 51, 51.5, 52 Hz), 12 amplitude slew rates ±(50, 100, 250, 500, 750, 1000 μV/s), and 14 frequency slew rates ±(0.01, 0.025, 0.05, 0.075, 0.1, 0.15, 0.2 Hz/s). CAE receptive field inputs one ECG lead with 1024 samples (2.048 s for 500 Hz sampling rate). CAE architecture is designed with seven 1D-convolutional layers, including three encoder layers (filters x kernel size = 16 × 8, 8 × 8, 8 × 8) and four decoder layers (8 × 8, 8 × 8, 16 × 8, 1 × 8) with linear activation function and same padding. CAE non-linear operations for max-pooling and up-sampling (pool size of 2) follow each encoder and decoder convolutional layers, respectively. Adam optimizer and mean squared error loss function are applied for CAE training over 250 epochs. The quality of clean ECG reconstruction in CAE output is evaluated by root-mean-square error (RMSE), percentage-root-mean-square difference (PRD) and improvement in signal-to-noise ratio (SNR_imp). Statistical test results for denoising of all 12 ECG leads present median RMSE = 5.3 μV, PRD = 3.5%, SNR_imp = 22–32 dB for SNR = –2.5 to 7.5 dB. The results do not substantially change for PLI frequencies 48–52 Hz, amplitude slew rates up to ±1000 μV/s and frequency slew rates up to ±0.2 Hz/s with median value divergence of $\Delta$RMSE < 2 $\upmu{\text{V}}$, $\Delta$PRD < 1.5%, $\Delta$SNR_imp ≤ 3 dB. The observed performance stability justifies the deep learning strategy for training a CAE with generalizable application for denoising of ECG signals with non-stationary PLI.