Interpretable Automated Arrhythmia Detection: An Assistive Framework for Clinicians.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Automated electrocardiogram (ECG) analysis for cardiac arrhythmia detection faces challenges with subtle pathological changes and limited interpretability of deep learning models.
• Traditional static ECG features often fail to capture complex patterns indicative of arrhythmias.
• The need for interpretable and accurate automated diagnostic tools in cardiology is paramount.

Purpose of the Study:

• To develop and validate a unique feature selection algorithm for identifying potent ECG biomarkers for accurate and interpretable cardiac arrhythmia detection.
• To compare the proposed feature selection technique against established methods like LASSO, F-test, and mRMR.
• To enhance the performance of a multi-class classifier for inter-patient arrhythmia classification.

Main Methods:

• A novel feature selection algorithm was developed to identify significant ECG biomarkers.
• Binary (One-vs-One) random forest (RF) classifiers were employed for feature discernment.
• A multi-class inter-patient RF classifier was trained using the selected features and compared against a Residual Network.

Main Results:

• The proposed feature selection algorithm demonstrated superior performance compared to LASSO, F-test, and mRMR.
• The selected significant features enabled the RF classifier to achieve a high average F1 score of 0.98 on the test dataset.
• The RF classifier outperformed a state-of-the-art Residual Network trained on the same dataset.

Conclusions:

• The developed feature selection algorithm effectively identifies key ECG biomarkers for enhanced arrhythmia detection.
• The approach offers a clinically interpretable and accurate tool for automated cardiac arrhythmia diagnosis.
• This method has the potential to significantly assist clinicians in diagnosing cardiac arrhythmias.