Interpretable Deep Learning Models for Arrhythmia Classification Based on ECG Signals Using PTB-X Dataset.

Area of Science:

• Cardiology
• Medical Diagnostics
• Artificial Intelligence

Background:

• Cardiovascular diseases necessitate early detection, with electrocardiogram (ECG) arrhythmias being critical indicators.
• Manual ECG interpretation is time-consuming and prone to human error, driving the need for automated diagnostic solutions.
• Deep learning (DL) excels at identifying complex patterns in raw ECG signals, offering a scalable approach.

Purpose of the Study:

• To develop and evaluate deep learning models for automated ECG arrhythmia classification.
• To assess the performance of custom CNN and modified VGG16 models using demographic data and advanced preprocessing.
• To enhance clinical transparency through interpretable ECG analysis.

Main Methods:

• Developed a custom dual-branch CNN and adapted a VGG16 model for multi-branch input, processing ECG signals and demographic data.
• Utilized the large-scale PTB-XL ECG dataset for binary, multiclass, and subclass classification tasks.
• Applied advanced preprocessing techniques and integrated demographic features to improve model performance.

Main Results:

• The custom CNN model achieved 97.78% accuracy in binary and 79.7% in multiclass classification tasks.
• Both models outperformed existing benchmarks, including CNN-LSTM and CNN entropy features.
• Interpretability analysis provided lead-specific insights into ECG contributions.

Conclusions:

• The developed deep learning models demonstrate high accuracy and potential for reliable, explainable arrhythmia detection.
• These models show promise for real-world application in cardiovascular diagnostics.
• Automated ECG analysis using DL can significantly aid in early detection and management of heart conditions.