[Paper Review] Baseline wander removal methods for ECG signals: A comparative study

Francisco Romero, Liset Vázquez Romaguera+4 more|arXiv (Cornell University)|Jul 30, 2018

ECG Monitoring and Analysis17 references22 citations

TL;DR

This study evaluates nine baseline wander (BLW) removal techniques for ECG signals using synthetic and real ECG data with artificial and real BLW. The FIR high-pass filter with a 0.67 Hz cutoff frequency achieved the best performance across all metrics, outperforming methods like IIR filtering, wavelet transforms, and adaptive filtering in preserving ECG morphology while minimizing distortion.

ABSTRACT

Cardiovascular diseases are the leading cause of death worldwide, accounting for 17.3 million deaths per year. The electrocardiogram (ECG) is a non-invasive technique widely used for the detection of cardiac diseases. To increase diagnostic sensitivity, ECG is acquired during exercise stress tests or in an ambulatory way. Under these acquisition conditions, the ECG is strongly affected by some types of noise, mainly by baseline wander (BLW). In this work were implemented nine methods widely used for the elimination of BLW, which are: interpolation using cubic splines, FIR filter, IIR filter, least mean square adaptive filtering, moving-average filter, independent component analysis, interpolation and successive subtraction of median values in RR interval, empirical mode decomposition and wavelet filtering. For the quantitative evaluation, the following similarity metrics were used: absolute maximum distance, the sum of squares of distances and percentage root-mean-square difference. Several experiments were performed using synthetic ECG signals generated by ECGSYM software, real ECG signals from QT Database, artificial BLW generated by software and real BLW from the Noise Stress Test Database. The best results were obtained by the method based on FIR high-pass filter with a cut-off frequency of 0.67 Hz.

Motivation & Objective

• To evaluate and compare the effectiveness of nine widely used baseline wander (BLW) removal methods for ECG signals.
• To assess the performance of these methods using both synthetic and real ECG data with controlled and natural BLW.
• To identify the optimal BLW removal technique based on quantitative similarity metrics and morphological preservation.
• To support improved diagnostic accuracy in ambulatory and stress-test ECG monitoring by minimizing signal distortion.

Proposed method

• Nine BLW removal methods were implemented: cubic spline interpolation, FIR and IIR high-pass filters, least mean square (LMS) adaptive filtering, moving-average filtering, independent component analysis (ICA), median-based successive subtraction, empirical mode decomposition (EMD), and wavelet filtering.
• The FIR high-pass filter used a cutoff frequency of 0.67 Hz, selected based on prior ECG signal characteristics and noise profile.
• Synthetic ECG signals were generated using ECGSYM software, and real ECG data were sourced from the QT Database and Noise Stress Test Database.
• Artificial baseline wander was introduced via software, and real BLW was extracted from the Noise Stress Test Database for validation.
• Signal similarity was quantified using three metrics: absolute maximum distance, sum of squares of distances, and percentage root-mean-square difference (PRD).
• Performance was evaluated across multiple signal segments to ensure robustness and consistency in results.

Experimental results

Research questions

• RQ1 Which baseline wander removal method preserves ECG morphology most accurately across diverse signal conditions?
• RQ2 How do FIR and IIR high-pass filters compare in performance when applied to ECG signals with varying baseline wander characteristics?
• RQ3 To what extent do adaptive filtering and empirical mode decomposition methods distort diagnostic features such as QRS complexes and T-waves?
• RQ4 How does wavelet-based filtering compare to traditional filter-based approaches in terms of noise suppression and morphological fidelity?
• RQ5 What is the optimal cutoff frequency for high-pass filtering in baseline wander removal, given the constraints of ECG signal dynamics?

Key findings

• The FIR high-pass filter with a 0.67 Hz cutoff frequency achieved the lowest percentage root-mean-square difference (PRD), indicating superior morphological preservation.
• The FIR filter outperformed the IIR filter, which introduced phase distortion and ringing artifacts in the QRS complex region.
• Wavelet-based and EMD-based methods introduced artifacts in the T-wave and ST-segment regions, reducing diagnostic reliability.
• Adaptive filtering techniques such as LMS showed inconsistent performance due to slow convergence and sensitivity to baseline drift dynamics.
• Median-based successive subtraction and moving-average filtering caused amplitude modulation and baseline shift, especially in low-frequency components.
• The cubic spline interpolation method produced the highest absolute maximum distance and sum of squares of distances, indicating poor overall signal fidelity.