Cardiac Arrhythmias Classification Using Machine Learning and Single-Lead ECG

Cardiac arrhythmias are associated with an elevated risk of myocardial infarction, stroke, and other abnormalities. Accurate and timely identification of the type of arrhythmia enables cardiologists to provide prompt and appropriate treatment. This paper presents an accurate approach to detect 6 types of heart rhythms (Sinus Bradycardia (SB), Sinus ventricular Tachycardia (SVT), Atrial Fibrillation (AFIB), Sinus Tachycardia (ST), Sinus Irregularity (SI), and Sinus Rhythm (SR)) from single-lead ECG. ECG recordings are de-noised by the Butterworth filter, Non-Local Mean (NLM) smoothing, and robust LOESS. A wide range of amplitude, spectral, nonlinear, entropy, and heart rate variability features are selected based on feature permutation importance and fed to various machine learning models to classify the type of rhythm. The single-lead models are compared to the 12-lead and evaluated using 5-fold cross-validation and testing on 2277 patients. The Histogram Gradient Boosting (HGB) model trained on lead II ECG data offers the best results of 97.6
^%
accuracy among single lead-trained models with fewer features. The highest accuracy of 98.4% is also obtained from the HGB model trained on 12-lead ECG data. The proposed single-lead detection with a smaller set of features, i.e. 35, is a promising solution for low-power and low-resource environments.