Medical Engineering & Physics 29 (2007) 26–37 A new approach to QRS segmentation based on wavelet bases and adaptive threshold technique Jo˜ ao P.V. Madeiro ∗ , Paulo C. Cortez, Francisco I. Oliveira, Robson S. Siqueira Department of Teleinformatics Engineering, Federal University of Cear´ a, Av. Mister Hull, S/N-CEP 60455-760, Fortaleza, Cear´ a, Brazil Received 30 August 2005; received in revised form 2 January 2006; accepted 17 January 2006 Abstract In this paper, we develop and evaluate a new approach to QRS segmentation based on the combination of two techniques: wavelet bases and adaptive threshold. Firstly, QRS complexes are identified without a preprocessing stage. Then, each QRS is segmented by identifying the complex onset and offset. We evaluated the algorithm on two manually annotated databases, the QT-database and the MIT-BIH Arrhythmia database. The QRS detector obtained a sensitivity of 99.02% and a positive predictivity of 99.35% over the first lead of the validation databases (more than 192,000 beats), while for the QT-database, values larger than 99.6% were attained. As for the delineation of the QRS complex, the mean and the standard deviation of the differences between the automatic and the manual annotations were computed. Using QT-database which contains recordings of annotated ECG with a sampling rate of 250 Hz, we obtain the average of the differences not exceeding two sampling intervals, while the standard deviations were within acceptable range of values. © 2006 IPEM. Published by Elsevier Ltd. All rights reserved. Keywords: Electrocardiogram (ECG); QRS complex; Wavelet transform (WT); Interval between beats (IBB); False positive (FP); False negative (FN) 1. Introduction Cardiac diseases constitute the main cause of mortality in many countries. The great amount of exams, which contains so much information of diversified nature, makes the visual evaluation difficult. Moreover, the manual analysis consti- tutes an error-prone task due to the visual fatigue. Hence, the convenience for developing automatic systems to process the electrocardiogram signal. Automatic feature extraction can improve the recognition of patterns of cardiac diseases in order to allow an early diagnose and, therefore, an efficient treatment. The QRS complex is the most characteristic waveform of the ECG signal [1]. Its higher amplitude makes QRS detec- tion easier than the other waves. Therefore, its detection is the first step for the complete segmentation of the ECG signal. However, in order to recognize patterns of cardiac diseases, ∗ Corresponding author. Tel.: +55 85 32496554. E-mail addresses: joaopdvm@yahoo.com.br, joaopaulo@deti.ufc.br (J.P.V. Madeiro), cortez@deti.ufc.br (P.C. Cortez), ivan@deti.ufc.br (F.I. Oliveira), siqueira@deti.ufc.br (R.S. Siqueira). the detection of occurrences of QRS complexes is insuffi- cient, and the complete information about its morphology is required. This also applies to the other waves from ECG sig- nal: P, T and U waves. The previous correct detection and a precise delineation of the QRS complex are fundamental conditions to perform an efficient detection and delineation of the other waves. Concerning QRS segmentation, we find in the literature very different approaches: based on mathematical models [2], signal envelop [3], matched filters [4], ECG slope crite- ria [5], second order derivatives [6], low pass-differentiation [7], the wavelet transform [8], non-linear time-scale decom- position [9], adaptive filtering [10], dynamic time warping [11], artificial neural networks [12] or hidden Markov mod- els [13]. These algorithms have the common characteristic of requiring a preprocessing stage over the entire signal before the QRS detection. For example, it is applied one or more filtering stages to the entire signal and, then, the decision rules are applied. Concerning on-line applications, where it is important the reduction of computing effort without effi- ciency loss, techniques without preprocessing are desirable. The technique of the wavelet transform, used in this work, 1350-4533/$ – see front matter © 2006 IPEM. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.medengphy.2006.01.008