Determining Dominant Frequency with Data-Adaptive Windows Gagan Mirchandani and Shruti Sharma School of Engineering, College of Engineering & Mathematical Sciences, University of Vermont Abstract. Measurement of activation rates in cardiac electrograms is commonly done though estimating the frequency of the sinusoid with the greatest power. This frequency, commonly referred to as Dominant Frequency, is generally estimated using the short-time Fourier Transform with a window of fixed size. In this work a new short-time Fourier trans- form method with a data-adaptive window is introduced. Experiments are conducted with both synthetic and real data. Results for the for- mer case are compared with current state-of-the-art methods. Given the difficulty in identifying activation points in electrograms, experiments reported in the literature have so far used only synthetic data. The new method is tested by application to real data, with true activation rates de- termined manually. Substantial improvement is observed. An error anal- ysis is provided. Keywords: atrial fibrillation, data-adaptive windows, non-stationary signals, dominant frequency. 1 Introduction Measurement of temporal beat-to-beat variations in heart rate is one of the key factors in the determination of cardiac disorders. In atrial fibrillation (AF), the atria contract rapidly and irregularly. The corresponding heart rate, measured through intracardiac electrograms, shows rapidly changing frequencies between 3-15 Hz. Since AF can be associated with heart disease and heart failure, an ac- curate determination of heart rate is essential for proper management purposes. Activation rate or frequency, is defined as the inverse of the time period between two consecutive activation points. While manual determination of atrial activa- tion timings by experts is theoretically possible, the sheer volume of data that would need analysis make this an impractical option. Accordingly, automated methods for determining the beat-to-beat variations have received much atten- tion. This is a difficult task for many reasons: electrogram data is random and nonstationarity and there is no clearly identifiable spectral estimation technique that specifically matches the particular problem of atrial activation rate deter- mination. Methods using the short-time Fourier transform (STFT) suffer from the usual time duration, frequency resolution problem. Furthermore, complexity in identifying activation time, often due to morphology fragmentation, further complicates the spectral estimation task. A. Elmoataz et al. (Eds.): ICISP 2010, LNCS 6134, pp. 287–296, 2010. c  Springer-Verlag Berlin Heidelberg 2010