Cardiac Arrhythmia Detection by Involving Multiscale Multifractal Analysis and Deep Learning

The multifractal nature of physiological responses varies across different time scales, influencing the accuracy of multifractal analysis. Traditional studies on heart rate variability for arrhythmia classification often use a fixed time scale, leading to limitations such as user dependence, sensitivity to scale range, fitting procedure artifacts, and failure to capture the crossover effect. Multiscale Multifractal Analysis (MMA) addresses these issues by handling varying scaling behaviors, offering more reliable results. The Generalized Hurst Surface <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$h$</tex>(q,s), a 3D representation from MMA, shows how local Hurst exponents vary with moments (q) and scales (s), providing comprehensive multifractal insights. While interactive exploration allows for an in-depth examination of intricate patterns, there is a need for a concise and static representation of h( q,s) to use as an input feature for classification models. To achieve this, two innovative approaches are introduced. The first approach projects the contour of h(q, s), used as input for AlexNet-based classifiers, while the second organizes Hurst surface values into a multi-variate series (h(q, s), q, s) for ResNet-based classification. Using datasets from Shaoxing People's Hospital, PhysioNet databases, and the 2017 PhysioNet/CinC Challenge, the models achieved over 99% accuracy, demonstrating significant efficacy in cardiac arrhythmia classification.