Real-time filtering adaptive algorithms for non-stationary noise in electrocardiograms

A non-stationary noise with a previously unknown intensity level often contaminates electrocardiogram (ECG) signals. Therefore, provision of high quality suppression of the non-stationary noise in ECG is a vital task to be performed. A new lightweight adaptive method has been proposed for real-time filtering of non-stationary (from the point of view of its variance) noise in ECG with noise-and signal-dependent switching filters, appropriate for processing a local vicinity of the current input signal sample. This method does not require time for filter parameters adaptation and a priori information about the noise variance. A one-and a two-pass algorithm on the simple optimal Savitzky & Golay filters and on the linear averaging filter have been developed on the basis of the proposed method. There is also an algorithm suggested applying a re-filtering only when the identifiers used in the method define a not low noise level. The integral and local statistical estimates of filters' efficiency have been obtained from numerical simulations over mean-square error (MSE), maximum absolute error (MAE), and signalto-noise ratio (SNR) for a model ECG signal under different levels of Gaussian noise. Filtering efficiency was estimated with the real signals taken from physionet.org database. The filter parameters were chosen by numerical simulations for a typical P-QRS-T cycle with corresponding signal sampling rate and scale considered. For a wide range from low to high noise levels (input SNR belongs to the interval from 25 to 0 dB), the statistical estimates of efficiency have been obtained as follows: for an ECG sampled at 360 Hz (taken from NSTDB), inside QRS-complex, the SNR increases by 2.5-6.7 dB, the MSE decreases in 1.7-4.3 times and the MAE decreases in 1.3-2.2 times; inside the segments prior to and following QRS-complex, the SNR, on an average, increase by 8.6-13.2 dB and the MSE decreases in 7.1-19.2 times, and the MAE decreases in 2.4-5.1 times. For an ECG sampled at 1 kHz (taken from PTB), inside QRS-complex, the SNR increases by 5.2-8 dB, the MSE and the MAE decrease in 3.2-6 times and 1.9-2.7 times, respectively; outside QRS, the SNR increases by 10.2-15.8 dB, the MSE and the MAE decrease in 10.5-36 times and in 2.6-5.2 times, respectively. For an ECG sampled at 250 Hz (from CUDB), the local indicators of efficiency are: inside QRS-complex, the SNR increases by 2.9-6.6 dB, the MSE and the MAE decrease in 1.8-4 times and in 1.4-2.3 times, respectively; outside QRS-complex, the SNR, on an average, increase by 4.7-11.6 dB, the MSE and the MAE decrease in 3.3-10.9 times and in1.7-3.9 times. Additionally, the filters' efficiency has been estimated as to suppression of real electromyographic (EMG) noise with significantly different variance and the proposed algorithms have been compared with other filters. A noise free ECGs during 5 min sampled at 360 Hz were contaminated with highly non-stationary EMG noise from a muscle artifact (MA) record of different intensity (input SNR varies from 20 to-5 dB), the SNR improvement at the proposed algorithm output is 10-14 dB. The calculated quantitative estimates of efficiency confirm the high quality of non-stationary EMG noise suppression obtained with the adaptive algorithms suggested. Minute signal distortions and a high degree of noise suppression have been demonstrated. Good performance and high filter quality for various real signals with non-stationary EMG noise have been shown. ECG amplitude-time parameters and waveforms, including pathological changes, are shown to be well-preserved.