An explainable deep learning framework for trustworthy arrhythmia detection from ECG signals.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Cardiovascular diseases (CVDs) are a major global health concern, with cardiac arrhythmias increasing mortality and morbidity.
• Accurate detection of arrhythmias from Electrocardiogram (ECG) signals is crucial but challenging due to data complexity.
• Current Deep Learning (DL) models for ECG analysis lack interpretability and face adoption barriers.

Purpose of the Study:

• To develop an explainable Deep Learning (DL) framework for accurate and reliable cardiac arrhythmia detection.
• To enhance the interpretability of DL models in ECG analysis for clinical adoption.
• To improve the generalization and performance of DL models using advanced data balancing techniques.

Main Methods:

• Integration of Convolutional Neural Network (CNN) and Dense Neural Network (DNN) architectures.
• Implementation of a multi-stage pipeline including data preparation, signal preprocessing, and multi-strategy data balancing (ADASYN, SMOTE, SMOTETomek, Random Over-Sampling).
• Incorporation of Explainable Artificial Intelligence (XAI) methods (SHAP, LIME, Feature Importance Analysis) for model transparency.

Main Results:

• The Random Over-Sampling combined with CNN (ROS+CNN) model achieved high classification accuracies: 99.74% (MITDB), 99.43% (PTBDB), and 99.98% (NSTDB).
• The framework demonstrated superior performance on benchmark ECG datasets.
• XAI components provided actionable insights into the model's decision-making process.

Conclusions:

• The developed explainable DL framework offers accurate and reliable arrhythmia detection.
• The integration of XAI fosters clinical trust and facilitates the adoption of AI in cardiovascular diagnostics.
• This approach paves the way for more impactful AI-driven solutions in cardiology.