Comparing ECG Lead Subsets for Heart Arrhythmia/ECG Pattern Classification: Convolutional Neural Networks and Random...

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) and conventional machine learning (CML) are increasingly used for electrocardiography (ECG) pattern classification.
• Limited research compares DL and CML performance on reduced ECG lead subsets for arrhythmia classification.
• The optimal number and selection of ECG leads for these methods remain unclear.

Purpose of the Study:

• To compare the accuracy of DL (CNN) and CML (RF) models in classifying heart arrhythmias using reduced ECG lead subsets.
• To identify optimal ECG lead subsets for both DL and CML methods.
• To assess the impact of lead reduction on classification performance.

Main Methods:

• Utilized the PhysioNet Cardiology Challenge 2020 dataset.
• Developed a CNN classifier for DL and a random forest classifier for CML.
• Employed recursive feature elimination to determine optimal ECG lead subsets for both methods.

Main Results:

• The CML method required approximately 2 more leads than the DL method.
• Four common leads (I, II, V5, V6) were identified for CML, while no common leads were consistent for DL.
• Average macro F1 scores were 0.761 for DL and 0.759 for CML.

Conclusions:

• Optimal ECG lead subsets yield classification accuracy comparable to using all 12 leads for both DL and CML.
• DL methods demonstrated slightly superior accuracy on larger datasets and required fewer leads than CML.
• Reduced lead sets are effective for arrhythmia classification using both DL and CML approaches.