Expert-enhanced machine learning for cardiac arrhythmia classification.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Distinguishing atrial fibrillation (AFib) from atrial flutter (AFlu) using surface electrocardiograms (ECGs) is challenging.
• Existing automatic cardiac arrhythmia classification methods often fail to differentiate between AFib and AFlu.
• Deep learning approaches may not provide satisfactory results for this specific classification task.

Purpose of the Study:

• To develop a novel method for classifying AFib versus AFlu from ECGs.
• To address limitations of current deep learning and traditional methods in differentiating these arrhythmias.
• To incorporate expert knowledge and clinical interpretability into an automated classification system.

Main Methods:

• Generation of novel, clinically interpretable features using mathematical optimization and solving a regression problem.
• Development of a machine learning model integrating expert knowledge on atrial flutter pathophysiology.
• Implementation of a tailored Branch-and-Bound algorithm for domain knowledge and standard algorithms like Adam for training.

Main Results:

• Achieved an accuracy of 82.84% and an Area Under the ROC Curve (AUC) of 0.9, classified as 'excellent'.
• The method provides inherent interpretability, offering insights into potential multilevel atrioventricular blocking mechanisms.
• Successfully classifies the difficult AFib↔AFlu distinction, complementing existing methods.

Conclusions:

• The proposed method offers a significant advancement in classifying AFib versus AFlu, improving upon existing techniques.
• The generated features and model provide valuable clinical insights, potentially aiding treatment decisions.
• This optimization-driven approach enhances machine learning for challenging cardiac arrhythmia classification.