iampedroalz / classification-atrial-beats

Atrial tachyarrhythmias such as atrial fibrillation (AFib) predispose to ventricular arrhythmias, sudden cardiac death and stroke. The complex and rapid atrial electrical activity makes it difficult to obtain detailed information on atrial activation during fibrillatory conditions. However, ectopic foci are often involved in initiating and sustaining AFib and therefore identifying the origin of atrial ectopic activity can help in diagnosis and treatment of AFib. Currently, invasive catheter mapping and ablation remains the cornerstone for the treatment of atrial arrhythmias. Non-invasive tools to guide electrophysiologists could significantly shorten catheter mapping procedures and may ultimately decrease the recurrence rate of ablations. Existing approaches are based on the analysis of the main characteristics of body body surface potential maps (BSPMs), such as the P-wave polarity, or rely on inverse reconstructions of the electric activity of the heart from BSPMs (ECG imaging). However, these methods have not yet shown to be accurate and reliable enough to be implemented in standard clinical practice. As the 12-lead electrocardiogram (ECG) is already routinely recorded in clinical settings, its use for classification of atrial ectopic foci into spatially differentiated atrial segments has been explored in this thesis. Atrial segments were defined with an automatic geodesic algorithm and a neural network (NN) was used for classification. Our simulation results with 8 atria-torso models show that ectopic foci with similar 12-lead ECG naturally cluster into differentiated atrial regions and that new patterns could correctly be classified into 29 segments with an accuracy of approximately 85%. Results also suggest that it is possible to predict whether the ECG signal belongs to the left atrium (LA) or the right atrium (RA) with an accuracy of 95%. If the classifier is applied to a new geometry that has not been used for training, however, the classification accuracy decreases drastically

Classication of Atrial Ectopic Beats into Spatial Segments based on the 12-lead ECG

P. Álvarez KIT - Master's thesis (09/2018)

Atrial tachyarrhythmias such as atrial fibrillation (AFib) predispose to ventricular arrhythmias, sudden cardiac death and stroke. The complex and rapid atrial electrical activity makes it difficult to obtain detailed information on atrial activation during fibrillatory conditions. However, ectopic foci are often involved in initiating and sustaining AFib and therefore identifying the origin of atrial ectopic activity can help in diagnosis and treatment of AFib. Currently, invasive catheter mapping and ablation remains the cornerstone for the treatment of atrial arrhythmias. Non-invasive tools to guide electrophysiologists could significantly shorten catheter mapping procedures and may ultimately decrease the recurrence rate of ablations. Existing approaches are based on the analysis of the main characteristics of body body surface potential maps (BSPMs), such as the P-wave polarity, or rely on inverse reconstructions of the electric activity of the heart from BSPMs (ECG imaging). However, these methods have not yet shown to be accurate and reliable enough to be implemented in standard clinical practice. As the 12-lead electrocardiogram (ECG) is already routinely recorded in clinical settings, its use for classification of atrial ectopic foci into spatially differentiated atrial segments has been explored in this thesis. Atrial segments were defined with an automatic geodesic algorithm and a neural network (NN) was used for classification. Our simulation results with 8 atria-torso models show that ectopic foci with similar 12-lead ECG naturally cluster into differentiated atrial regions and that new patterns could correctly be classified into 29 segments with an accuracy of approximately 85%. Results also suggest that it is possible to predict whether the ECG signal belongs to the left atrium (LA) or the right atrium (RA) with an accuracy of 95%. If the classifier is applied to a new geometry that has not been used for training, however, the classification accuracy decreases drastically.