The goal of this research was to classify cardiac excitation patterns during atrial fibrillation (AFib). For this purpose, virtual models of intracardiac mapping catheters were moved across in-silico cardiac tissue to extract local activation times (LATs) of each catheter electrode from simulated cardiac action potential (AP) signals. The resulting LAT patterns consisting of the LATs of all electrodes resemble patterns measured in clinical cases. The LATs represent the input information for features that were used to separate four different excitation patterns during AFib. Those four excitation patterns were plane wave, ectopic focus (spherical wave), rotor (spiral wave) and block. A feature selection algorithm was used to investigate the features concerning their power to classify the different simulated excitation patterns. The scores of the selected features were used to train and optimize a support vector machine (SVM). The optimized and cross-validated SVM was then used to classify the simulated cardiac excitation patterns. The achieved overall classification accuracy of this SVM model was 98.4 %.
Aiming for patient specific treatment of atrial fibrillation, cardiologists in the EP-lab (ElectroPhysiology-lab) intend to identify the pattern of depolarization waves in the atria by measuring endocardial electrograms with multichannel catheters. Hereby the pattern of plane waves, ectopic foci, lines of block, or rotors are of special interest. Data acquisition is performed with various multichannel catheters, and all four patterns leave different fingerprints in the electrograms. In this work we extract features from the activation sequence in the electrograms that can support the cardiologist to identify the underlying depolarization pattern. To this end computer simulations of fundamental depolarization scenarios were carried out and the corresponding activation patterns were analyzed.
Student Theses (2)
C. Reich. Classification of Cardiac Excitation Patterns during Atrial Fibrillation using Multichannel Mapping Data. Institute of Biomedical Engineering, Karlsruhe Institute of Technology (KIT). Masterarbeit. 2015
Atrial fibrillation (AFib) is the most common type of cardiac arrhythmia worldwide. However, mechanisms for the initiation and maintenance of AFib are still not completely understood. Different theories favoring multiple wavelets, ectopic foci and rotational sources as potential drivers for the maintenance of AFib are present. To find those drivers clinicians use different kinds of intracardiac mapping catheters including Singleloop, Doubleloop and PentaRay catheters.Virtual models of those catheters were used to extract information in form of the local activation times (LATs) of each catheter electrode from simulated electrograms (EGMs). The resulting LAT patterns consisting of the LATs of all electrodes were used as input information for different features that were introduced to separate four different excitation patterns during AFib. Those four excitation patterns were plane wave, ectopic focus (spherical wave), rotor (spiral wave) and block. A feature selection algorithm was introduced to investigate the features concerning their usefulness to classify the different simulated excitation patterns. The scores of the selected features were used to train and optimize a Support Vector Machine (SVM). This trained SVM model was used to classify the artificial data as proof of concept and different clinical databases afterwards.For the artificial data an overall classification accuracy of 97.4 % was achieved for the optimized SVM. When the LAT patterns for the artificial data were artificially corrupted by additive Gaussian noise a reduction of the classification accuracy dependent on the noise level was observable. For three clinical databases different results were obtained. For the atrial flutter (AFL) database with LAT patterns similar to the noise-free artificial data a high predictive accuracy was achieved. For another database containing pacing episodes misclassifications occurred. LAT patterns with outliers were often misclassified as block or rotor excitation patterns. For the AFib database LAT patterns very different from those in the noise-free artificial data were observed. Most of the AFib data records were classified as block or rotor excitation patterns.