DenseNet-ABiLSTM: Revolutionizing Multiclass Arrhythmia Detection and Classification Using Hybrid Deep Learning Approach Leveraging PPG Signals

Abstract

Arrhythmias (AM) are heart conditions that can lead to fatal cardiac arrest. Automated identification of arrhythmias is crucial for detecting cardiac diseases. Previous studies have used photoplethysmography (PPG) signals to identify arrhythmias, but there is limited research on their application for multiclass arrhythmia classification. This study introduces a Hybrid Deep Learning (HDL) model called DenseNet-ABiLSTM, which uses densely connected convolutional networks and Attention-based Bidirectional Long Short-Term Memory (ABiLSTM) to categorize various types of arrhythmias. The model uses 1D convolutional kernels to acquire multiscale conceptual features, followed by BiLSTM to understand temporal relationships among features. The Attention Mechanism layer is presented to improve detection performance. The model categorizes arrhythmia rhythms into six types: Sinus Rhythm (SR), Early Ventricular Contraction (EVC), Early Atrial Contraction (EAC), Ventricular Tachycardia (VT), Supraventricular Tachycardia (ST), and AF. Various metrics were assessed and compared with Electrocardiogram (ECG) results to determine AM rhythms. The mean performance measures showed strong overall performance, with a mean F1 score and accuracy of 87.74% and 89.14%, respectively.