Validation of a Novel Algorithm for Ventricular Repolarization Analysis: Use of Physionet Resources F Cantini, M Emdin, C Passino, M Varanini, F Conforti CNR Institute of Clinical Physiology, Pisa, Italy Abstract Ventricular repolarization analysis allows extraction from the ECG signal of quantitative indexes (namely the QT interval), of prognostic value in unselected populations and cardiac patients, being related with arrhythmic risk. Several attempts to improve automatic ECG waveform detection have been accomplished, using signal derivatives, digital filtering, wavelet analysis, neural network techniques, nonlinear approaches. In the present study, a single-lead low-pass differentiation detector of ECG significant points (PulseMeter) has been evaluated. The algorithm performance has been validated against the manual annotation of the "QT database" (http://www.physionet.org/), developed for validation purposes. QRS complex and other ECG waveform boundaries were independently evaluated in the present study. The mean values and standard deviations computed improve the result of automatic annotation in QT database, especially in T wave detection. The QRS detector has a sensitivity of 99.96% and a positive predictivity of 99.96% on the first lead and a sensitivity of 99.90% and a positive predictivity of 99.94% on the second lead, showing a better performance than the automatic annotation in the QT database. 1. Introduction The analysis of ventricular repolarization allows to compute quantitative indexes (such as JT, QT interval) with diagnostic and prognostic values in patients with systemic (i.e. diabetes) or cardiovascular diseases, because of their linkage with arrhythmias or sudden death. The development of automatic algorithms for ventricular repolarization analysis from the electrocardiographic signal is relevant to the diagnostic process in subsets of patients with potential arrhythmic or ischemic risk, either during provocative tests, ambulatory or Intensive Care Unit ECG monitoring. In the present study, a single-lead, low-pass differentiation detector of ECG significant points (PulseMeter) (figure 1) has been evaluated. 2. Methods The “Physionet QT database” [1] has been used for PulseMeter algorithm performance evaluation. The database is a collection of 15-minute long, 2-lead, 105 selected ECG recordings from other databases such as MIT-BIH Arrhythmia Database [7], European ST-T Database [8] and Boston Beth Israel Deaconess Medical Center database. Signals have been selected to cover a wide range of QRS, ST segment and T wave morphologies. For each record a set of annotation for the QRS complex and a set of manual annotations for T, U and P wave boundaries for selected beats are provided. For 11 records a second set of manual annotations is provided. Also, automatic annotations obtained using the “ecgpuwave” application are available [5][6]. 2.1. Preprocessing While processing biological signals [2], filter cut-off characteristics are not critical in order to extract information from the signals. Thus, for our purpose, moving-average digital filters were used. This kind of Figure 1: Application of PulseMeter to the ECG signal: automatic annotation of significant points 0276-6547/03 $17.00 © 2003 IEEE 509 Computers in Cardiology 2003;30:509-512.