Sinusoidal rhythms identification from intracardiac electrograms using machine learning techniques

Abstract

Our heart muscles continuously experience small electrical changes, including a sequence of depolarization followed by repolarization. This process allows the charged ions to pass through cells, triggering contractions during each cardiac cycle, producing what we called heartbeat. The accepted limits of the heart beat have been conventionally set at 60 to 100 pulses per minute,[1]In heart disease known as atrial fibrillation (AF), this regular sequence is disrupted. In many patients, the atria with the pulmonary veins (PV) can fire rapid, spontaneous bursts. The activation rate could be as high as 400-600 pulses per minute, and these abnormal impulses can overrun the normal sinus rhythm, jolting the atria into chaotic, unsynchronized activation. Each impulses fire from the atria, conducted to the ventricles, and then the rapid ventricular rate might lead to ineffective cardiac contraction and instant death.[2] Theories[3] have been supporting an origin from a rapid tachycardia from the PV, with the remodeling process promoting the transition to re-entry AF.[4] Besides the PV, the ectopic foci from other sources such as the superior vena cava has been identified as potential trigger of AF.[5] Ablation with high radiofrequency to isolate the PV region have been considered as the most conventional methods for AF treatment. [6]The process is guided by the electro-anatomical mapping system to provide information on the catheter navigation as well as the voltage maps throughout each stage.[7] The electrocardiogram (ECG) which is mapped on the body surface could also help with diagnosis for cardiac arrhythmia[8], however, the ECG only records depolarization of the large atrial masses, the atria and ventricles as a whole, without showing the origin of the cardiac impulse. In contrast, the intracardiac electrogram (iEGM) amplifies electrical activity unrecordable in ECG recordings,[9] therefore makes it a robust reference in post ablation analysis. The iEGM are recorded separately for each catheter location, and classified by their hallmarks of left atrium (LA), right atrium(RA), right ventricular, left ventricular. Both left and right atrium have veins drained from their respective PV region, and are of great importance of guidance to the interventional decision in an ablation procedure. [10] Although in many studies for iEGM analysis, RA has been treated as a replacement for LA, since it’s more easily acquired by avoiding exposing the patients to the risk of trans-septal puncture. [11] In studies for arrhythmia detection, the arrhythmia iEGM have been compared against the sinus rhythms with no differentiation between RA and LA signals. However, research has shown that the left lateral ridge might be a potential site for ectopic foci, leading to AF, as well there are structural differences on the LA appendage in shape and length.[11] In my project, I aimed to identify the difference between RA and LA baseline sinus rhythms, as well as their correlation with the respective signals during AF active stages. First selected features have been extracted on the iEGM signals, capturing their frequency, entropy and shape information in a scalable time window. Then I use this validated feature map as the basis for comparing across different stages of signals from the LA and RA . In this work, we demonstrate how machine learning methods such as Support Vector Machine(SVM) and random forest could be used to evaluate the iEGM in the high-dimensional parameter space and effectively discriminate between the sinus rhythms recorded from LA and RA, and the classification would be essential to the futuristic analysis for arrhythmia identification and post-ablation diagnosis, and the electrophysiology features shown in the RA baseline signals before ablation are similar to the features for the LA signals post ablation. We are developing a clustering system to use machine learning techniques to classify the signals of LA and RA baseline signals in real time.