Application of Machine Learning Techniques for Testing Sensitivity of Cardiac Dynamics to Arrhythmia Vulnerability Parameters

Abstract

In this study, we developed a new approach to address the importance of arrhythmia vulnerability parameters for cardiac dynamics prediction in response to autonomic stimulation. We generated various case studies using a series of complex electrophysiology protocols for control versus stimulated single cardiac myocytes by autonomic nervous system (ANS). We collected the “TRiAD” related parameters (triangulation (APD TRI), reverse use dependence (RUD), beat‐to‐beat (bTb) instability of action potential duration and spatial APD dispersion (T‐wave area)). We used collected data to train different machine learning algorithms for classifying control versus ANS case studies. We used artificial neural network (ANN), support vector machine (SVM) with Gaussian kernels and random forest (RF) algorithms and observed that RF outperformed the other two algorithms with higher model accuracy for output prediction. Our results illustrated that bTb instability, APD TRI, RUD and T‐wave area are respectively the most important parameters for our classification problem. We also studied the sensitivity of cardiac dynamics to the listed TRiAD parameters. We observed that increase in bTb instability values leads to higher probability of case study being classified into control group while increase in RUD and TRI values result into higher probability of case study being classified into ANS group. Whereas T‐wave area values illustrated different trend compared to three other parameters, where low values could be misleading in decision making about the group that the case study will belong to and higher T‐wave values mostly belong to ANS classified case studies. Therefore, our study suggests the advantages of data analytics techniques for discoveries that may not have been evident.Support or Funding InformationThis study was funded by National Institutes of Health (Grant OT2OD026580).