Cardiac anomaly detection based on time and frequency domain features using tree-based classifiers.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Mobile health (mHealth) devices have increased the need for accurate, automated cardiac anomaly detection algorithms.
• Electrocardiogram (ECG) analysis is crucial for diagnosing various heart conditions, including arrhythmias.

Purpose of the Study:

• To develop and evaluate a combined classical signal analysis and machine learning algorithm for high-accuracy cardiac anomaly detection.
• To improve classification accuracy for arrhythmias using the Computing in Cardiology Challenge (CinC) 2017 dataset.

Main Methods:

• Developed a method combining classical signal analysis with machine learning, utilizing nearly 400 hand-crafted features.
• Compared gradient boosted trees and random forests, assessing the impact of different training annotations and feature subsets.
• Introduced a web-based ECG viewer for reviewing and correcting training labels.

Main Results:

• The gradient boosted tree model outperformed random forests in classification accuracy.
• Achieved a mean F1-score of 83.2% on the CinC 2017 dataset, placing second among 80 algorithms.
• Specific class accuracies included 90.8% for Normal and 84.1% for Atrial Fibrillation (AF).

Conclusions:

• The developed algorithm demonstrates strong performance in detecting cardiac arrhythmias, suitable for mHealth applications.
• The combined approach of signal analysis and machine learning offers a robust solution for automated ECG interpretation.
• Feature importance analysis and optimized model selection are key to achieving high accuracy in arrhythmia detection.