Novel Approaches to ECG-Based Modeling and Characterization of Atrial Fibrillation

Sammanfattning: This thesis deals with signal processing algorithms for analysis of the electrocardiogram (ECG) during atrial fibrillation (AF). Such analysis can be used for diagnosing patients, and for monitoring and predicting their response to various treatment. The thesis comprises an introduction and five papers describing methods for ECG-based modeling and characterization of AF. Paper I--IV deal with methods for characterization of the atrial activity, whereas Paper V deals with modeling of the ventricular response, both problems with the assumption that AF is present. In Paper I, a number of measures characterizing the atrial activity in the ECG, obtained using time-frequency analysis as well as nonlinear methods, are evaluated for their ability to predict spontaneous termination of AF. The AF frequency, i.e, the repetition rate of the atrial fibrillatory waves of the ECG, proved to be a significant factor for discrimination between terminating and non-terminating AF. Noise is a common problem in ECG signals, particularly in long-term ambulatory recordings. Hence, robust algorithms for analysis and characterization are required. In Paper II, a robust method for tracking the AF frequency in noisy signals is presented. The method is based on a hidden Markov model (HMM), which takes the harmonic pattern of the atrial activity into account. Using the HMM-based method, the average RMS error of the frequency estimates at high noise levels was significantly lower compared to existing methods. In Paper III, the HMM-based method is employed for analysis of 24-h ambulatory ECG signals in order to explore circadian variation in AF frequency. Circadian variations reflect autonomic modulation; attenuation or absence of such variations may help to diagnose patients. Methods based on curve fitting, autocorrelation, and joint variation, respectively, are employed to quantify circadian variations, showing that it is present in most patients with long-standing persistent AF, although the short-term variation is considerable. In Paper IV, 24-h ambulatory ECG recordings with paroxysmal and persistent AF are analyzed using an entropy-based method for characterization of the atrial activity. Short segments are classified based on these measures, showing that it is feasible to distinguish between patient with paroxysmal and persistent AF from 10-s ECGs; the average classification rate was above 95%. The ventricular response during AF is mainly determined by the AV nodal blocking of atrial impulses. In Paper V, a new model-based approach for analysis of the ventricular response during AF is proposed. The model integrates physiological properties of the AV node and the atrial fibrillatory rate; the model parameters can be estimated from ECG signals. Results show that ventricular response is sufficiently represented by the estimated model in a majority of the recordings; in 85.7% of the analyzed 30-min segments the model fit was considered accurate, and that changes of AV nodal properties caused by autonomic modulation could be tracked through the estimated model parameters. In summary, the work within this thesis contributes with new methods for non-invasive analysis of AF, which can be used to tailor and evaluate different strategies for AF treatment.