Feasibility Study on the Use of Heart Rate Variability Parameters for Detection of Atrial Fibrillation with Machine Learning Techniques

The paper is devoted to development and studies on atrial fibrillation (AFib) detection in electrocardiogram (ECG) using digital signal processing (DSP) and machine learning (ML). The goal of this pilot study was to find the DSP and ML methods suitable for the AF detection in real-time in short single-lead ECGs containing 32 consecutive cardiac cycles. Three simple Heart Rate Variability (HRV) parameters from the time domain analysis were calculated and used as features for ML algorithms. Binary decision tree and shallow neural network were used for classification, and the impact of metaparameters on the performance of the AFib detection algorithms was investigated to determine the lower limit of their required complexity. In the neural network, different numbers of hidden neurons and different activation functions were examined. In the decision tree, different limits on the maximum number of splits were set. For both AFib detection algorithms, various sets of HRV-based features were tested. With neural network (two features, ten hidden neurons), 98.3% accuracy, 97.1% sensitivity and 99.1% specificity were obtained. With decision tree (two features, seven splits), 96.9% accuracy, 96.3% sensitivity and 97.4% specificity were reached. This study shows the usefulness of neural network and decision tree algorithms for the detection of atrial fibrillation using the simplest HRV parameters. The use of more complex HRV parameters in AFib detection with the proposed ML algorithms requires further investigation.