Objectives. Holter monitoring is mainly used for medical follow-up and diagnosis of patients with suspected cardiac arrhythmia such as heart rhythm irregularities that can be missed during classical electrocardiogram recording (ECG). However, these long-term continuous recordings represent a large amount of data that cannot be processed by hand. The development of an automatic tool able to delineate and detect ECG events has become essential. Our objective is to present a new method based on Non-Negative Matrix Factorization (NMF) to detect R-peaks in Holter signals.
Methods. The approach consists in two stages namely source separation and detection of R peaks. First, NMF uses the spectrogram of the Holter signal and the different time-frequency patterns between the QRS complexes and the other waves of the signal (P and T waves) to separate the two sources (QRS complexes and non-QRS parts). Using the QRS source, Automatic Objective Thresholding (AOT) is applied to detect R-peaks. This computational method was applied to clean and noisy Holter signals to assess its reliability in a realistic context.
Results. The proposed approach is validated on the MIT-BIH Arrhythmia database and achieves comparable results in terms of sensitivity (Se = 99.59%), precision (Pr = 99.69%) and accuracy (Acc = 99.22%) with other existing methods. Using the MIT-BIH Noise Stress Test database, we also show the ability of our approach to discriminate R-peaks in noisy signals and in presence of complex baseline wanders.
Conclusion. This paper shows the practical relevance and effectiveness of NMF to detect R-peaks in Holter signals. A main advantage of this technique is that no pre-processing is needed since NMF-AOT is able to remove measurement noise as well as artifacts and baseline wanders. Future work will focus on the use of the second source in the detection of atrial fibrillation.