Detecting atrial fibrillation from short single lead ECGs using statistical and morphological features.

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Point-of-care ECG devices offer potential for early atrial fibrillation (AF) detection.
• Algorithm efficiency relies on accurate AF detection from short, single-lead ECGs amidst noise and artifacts.
• Distinguishing AF from normal sinus rhythm and other arrhythmias in noisy conditions remains a challenge.

Purpose of the Study:

• Develop an algorithm to classify short single-lead ECGs into 'Normal', 'AF', 'Other', and 'Noisy' categories.
• Identify challenges in algorithm development for point-of-care ECG analysis.
• Propose mitigation strategies for issues like lead inversion, low signal amplitude, noise, and artifacts.

Main Methods:

• Rule-based identification for lead inversion and noisy records.
• Extraction of statistical and morphological features.
• Support vector machine (SVM) classifiers for rhythm classification.
• Training and testing on 12,186 short single-lead ECGs from a point-of-care device.

Main Results:

• Achieved 77.5% sensitivity and 97.9% specificity for AF detection against non-AF rhythms.
• Overall accuracy of 96.1% in cross-validation.
• F1 measures of 89% ('Normal'), 78% ('AF'), and 67% ('Other') on a hidden test set.
• Attained a 78% overall score in the Computing in Cardiology Challenge 2017.

Conclusions:

• The developed algorithm effectively classifies short single-lead ECGs from point-of-care devices into four categories.
• Addresses challenges unique to point-of-care ECG analysis, including noise and artifacts.
• Demonstrates a viable computational approach for improving automated arrhythmia detection in resource-limited settings.