Artificial Intelligence Enabled Classification and Prediction of Atrial Arrhythmias using ECG Signal Morphological Features

Abstract

The Electrocardiogram (ECG) is a diagnostic tool that non-invasively captures and records the heart’s electrical impulses. Examining the electrical characteristics of the heart can unveil the presence of cardiovascular ailments and furnish insights into the performance of the intracardiac conducting tissue. Cardiovascular diseases (CVDs) are currently the primary cause of death globally. CVDs were responsible for the demise of 17.9 million individuals globally in 2019, accounting for 32% of all mortalities and further the prevalence will double in 2030. The causes of mortality for 85% of these people were heart attacks and strokes. According to recent data, in 2019, CVDs accounted for 38% of premature mortalities (under 70) attributed to non-communicable diseases. The most commonly occurring CVD globally is Atrial Fibrillation (AF), a type of atrial arrhythmia. Despite several attempts to diagnose CVDs precisely, many go undetected. The limited based knowledge acts as a drawback in diagnosing fast-evolving CVDs. Mostly CVDs are reported in the last stage, thus limiting the choices of methodological diagnosis to the healthcare fraternity. Prior research has primarily centered on ventricular arrhythmias, with insufficient attention given to arrhythmias associated with the atria. Furthermore, there is a dearth of morphologically based classifications for atrial arrhythmias. The purpose of this study is to analyze, classify, and predict various heart rhythms generated by the sinoatrial (SA) Node, atrial disorders, and atrial arrhythmias using Artificial Intelligence (AI) techniques such as Machine Learning (ML) and Deep Learning (DL) based on morphological features (e.g., peaks, waves, intervals) obtained from ECG signal. Initially, this work analyzed Heart Rate Variability (HRV) in Sinus Rhythm (SR) and Exercise-induced Sinus Tachycardia (e-ST) conditions in two phases. The first phase consisted of statistical analysis of the time and frequency domain parameters and morphological features such as P-wave and PP Interval (PPI). The analysis revealed a significant alteration in the durational aspects of the HRV features when compared to SR and ST. The subsequent phase of the study involves the creation of a Long Short Term Memory (LSTM) based Recurrent Neural Network (RNN) to forecast the heart rate of the e-ST condition. The model utilized HRV features as input for multivariate time series forecasting to predict the heart rate of e-ST volunteers. The work focused on classifying e-ST with other atrial arrhythmias, such as Atrial Tachycardia (AT), because of the heart rate is the same in both cases and it is a herculean task to differentiate manually based on the ECG signal. Primarily to classify SR, e-ST, and AT, the three ML algorithms used were Extra Trees (ET), Ridge Classifier (RC), and CatBoost (CB) classifier. The classification methodology relies on morphological characteristics extracted from the ECG signal. Subsequently, these features were prioritized based on their significance in distinguishing SR, e-ST, and AT. In the feature ranking plots, the atrial features were given utmost importance in classifying as these conditions directly influenced them. Later the focus of the work shifts to analyzing and classifying precursors of AF, which are AT and LAE. The morphological-based analysis and classification are needed to understand the occurrence of AF. The morphological features consisted of temporal and amplitude aspects of the ECG signal and features obtained from P-Wave Indices (PWI). The PWI gives better insights into the atria functions. The initial phase of the work was classifying SR, e-ST, AT, and LAE by developing stacked ML models, and the later half ranked the morphological features by a pie formula-based technique. The PWI-based features obtained the highest importance compared to the temporal and amplitude aspects of the ECG signal. Finally, this work classified types of AF such as Paroxysmal Atrial Fibrillation (PAF) and Persistent Atrial Fibrillation (PsAF) from Non-AF cardiac rhythm. The classification of AF subtypes and Non-AF is necessary to better the clinical decision management and delineate the A ’s clinical condition. The work proposes a 2D custom Convolutional Neural Network (CNN) model to classify AF subtypes automatically based on the time-frequency spectrum. After further converting ECG signals to time-frequency spectrum using Constant-Q Transform (CQT), the obtained plots acted as the input to the CNN model. The advantage of the time-frequency spectrum is that the features extracted by the CNN model are Intermediate Frequency (IF) and Spectral Delay (SD), thus giving advanced information about the ECG signals, such as instantaneous bandwidth, duration, and amplitude. Medical professionals may find it helpful to use morphological feature-based ML models to get insight into crucial clinical ECG features for early atrial arrhythmia prediction. Additionally, the generated models may help medical professionals choose individualized AF treatment and decrease misdiagnosis.