Quick ECG Segmentation, Artifact Detection and Risk Estimation Methods for On-Line Holter Monitoring Systems

Computer-aided bedside patient monitoring requires real-time vital function analysis. On-line Holter monitors need reliable and quick algorithms to perform all the necessary signal processing tasks. This paper presents all the methods that were conceptualized and implemented at the development of such a monitoring system at Medical Clinic No. 4 of Targu Mures. The system performs the following ECG signal processing steps: (1) Decomposition of the ECG signals using multi-resolution wavelet transformation, which also eliminates most of the high and low frequency noises. These components will serve as input for wave classification algorithms. (2) Identification of QRS complexes, P and T waves using two different algorithms: a sequential classification and a neural-network-based clustering algorithm. This latter also distinguishes normal R waves from abnormal cases. (3) Localization of several kinds of arrhythmia using a spectral method. An autoregressive (AR) model is applied to estimate the series of R-R intervals. The coefficients of the AR model are predicted using the Kalman filter, and these coefficients will determine a local spectrum for each QRS complex. By analyzing this spectrum, different arrhythmia cases (bigeminy, trigeminy, ventricular flutter, etc.) are identified. The algorithms were tested using the MIT-BIH signal database and own multi-channel ECG registrations. The QRS complex detection ratio is over 99.6%. Using the output of the above mentioned methods, the on-line monitor system performs heart rate variability (HRV) and heart rate turbulence (HRT) analysis, which can help the diagnosis and can predict dangerous states of the patient.