# Deep learning-based early prediction of life-threatening ventricular arrhythmias using long-term Holter ECG signals

## TL;DR

This paper introduces a deep learning model combining graph neural networks and transformers to accurately detect life-threatening heart rhythm disorders from long-term ECG recordings.

## Contribution

A novel hybrid GNN + transformer framework for real-time arrhythmia detection with improved accuracy over traditional methods.

## Key findings

- The hybrid model achieved 98.27% accuracy in arrhythmia classification using Holter ECG data.
- Outperformed traditional machine learning models like CNN-Bi-LSTM and Bi-LSTM in precision and recall metrics.
- Demonstrated potential for real-time wearable healthcare applications with high classification accuracy.

## Abstract

Ventricular arrhythmias (VAs) are among the primary reasons for sudden cardiac death, and their early detection requires a key factor to reduce patient mortality. Conventional tools used to identify arrhythmias (manual and rule-based) are not only time-consuming but also have become dependent on an expert to interpret; thus, their utility is constrained in terms of their scalability and applicability in viewing arrhythmias in real-time. The increasing rate of cardiovascular diseases (CVD) and the desire to have an efficient and real-time environment are evidenced in the weaknesses of current systems.

The current research introduces a new deep learning framework on the basis of graph neural networks (GNNs) and a transformer to improve the detection of arrhythmia with the Holter ECG signal. The data are collected via Holter ECGs and the Sudden Cardiac Death Holter Database (SDDB) as a basis to start the workflow. It involves preprocessing of data, such as elimination of noise, normalization of the signals, and segmentation of the data to pertinent ECG segments. Features are then extracted by a combination of time domain, frequency domain, and non-linear techniques, followed by classification using the hybrid GNN + transformer model to incorporate both the spatial and the temporal dependencies. In comparison to classical methods of the rule -based approaches, machine learning algorithms, such as the hidden Markov models (CNN-Bi-LSTM) and recurrent neural networks (Bi-LSTM), the hybrid model of GNN + transformer automatically determines arrhythmias, including spatial and temporal dependencies, to enhance the classification accuracy by a significant margin.

Model training and testing were performed on MIT-BIH and SDDB, and the accuracy, precision, recall, and F1-score were 98.27%, 98.08%, 98.27%, and 97.76%, respectively. This evidence proves the framework to be powerful in practical arrhythmia identification, providing a trustworthy way of monitoring the heart.

The hybrid model is efficient compared with the traditional models and offers an extensible solution to wearable healthcare systems that would have the quality of detecting arrhythmia in real-time with a high degree of accuracy.