The following is a summary review of “QRS detection using adaptive filters: A comparative study” from Volume 66 of ISA Transactions (January 2017). If you would like to learn more about ISA Transactions, please visit this link.
See Samuel Ko Tak Shun's other summary on “Design of minimum multiplier fractional order differentiator based on lattice wave digital filter” here.
Electrocardiogram (ECG) is the electrical activity of the heart, measured by placing electrodes at several positions on the human body. The ECG signal is a quasi-periodic wave that comprises of a few individual characteristic waves, given as P, Q, R, S, and T. These waves initiate with P wave that originates from the atrial depolarization of heart muscles; the Q, E, and S waves originate from the ventricular depolarization, and are together interpreted as the QRS complex. The T wave is due to the re-polarization of the heart, and this cycle repeats itself with the next P wave.
ECG signal monitoring and analysis is widely used to explore and identify various heart diseases such as atrial and ventricular premature contractions, atrial fibrillations, bradycardia, tachycardia, etc. Variations from a normal sinus rhythm (standard heart cycle) are the symptoms of arrhythmia that represents some form of cardiovascular disease (CVD). CVD has been identified as one of the dominant causes of death by the World Health Organization (WHO). CVD is a primary health concern in most countries, with large expenditures on research and development of CVD monitoring and control equipment.
Such equipment and devices rely on digital ECG signal processing methods. Digital ECG also plays an important role in compressing the ECG signal for teleporting to tertiary healthcare centers under low bandwidth requirements. To enhance ECG monitoring for easy identification of abnormalities, automated and adaptive computerized methods have been developed and adopted.
The process of QRS detection starts with the mean subtraction of ECG signals, after which adaptive filtering, also known as adaptive linear prediction, is applied to get the instantaneous prediction error (PE). This then undergoes Savitzky-Golay (SG) filtering for peak enhancement and noise suppression.
The mathematical equations of variants consist of the following least mean square (LMS) variants:
- Basic LMS
- Variable step size LMS
- Leaky LMS
- Variable leaky LMS
- Fractional LMS
- Normalized least mean square (NLMS)
- Sign error LMS
- Sign-sign LMS
- Recursive least square (RLS)
The mathematical equations of the above variants can be found on page 364 of this research article.
The performance of the proposed QRS detection algorithm is evaluated on the standard MIT-BIH Arrhythmia Database for a data length of 60 seconds (21,600 samples). Its results are as follows:
- Optimized initialization of predictor-coefficients
- Optimized initialization of parameters of predictor variants
- Detection performance
Its discussions are divided into four sections:
- LMS variants versus fidelity parameters
- LMS variants versus time complexity
- LMS variants versus averaged weights
- Proposed improved algorithm versus state-of-the-art algorithms
The above discussions can be found in more detail on pages 368-374 in the research article.
In conclusion, the improvement of the QRS detection system is done by utilizing new variants of LMS algorithms; a novel combination of a few variants have been tested and analyzed. The performance of all variants are compared to find out the best variant to be used in adaptive filtering-based QRS detection systems. The performance of LMS variants gets degraded when either the ECG signal is highly corrupted with noises due to EMG and motion artifacts resulting in false positive (FP) beat detection, or it incorporates extraneously wide fusion (F) beats that generally result in false negative (FN) beat detection.
Jain, S., Ahirwal, M. K., Kumar, A., Bajaj, V., & Singh, G. K. (2017). QRS detection using adaptive filters: A comparative study. ISA transactions, 66, 362-375.