Fractals and Surrogate Data Analysis for ECG Classification

Manouane Caza-Szoka, François Nougarou, Daniel Massicotte
Université du Québec à Trois-Rivières


This paper presents how the ideas in Fractals and Surrogate Data Analysis can be adapted to extract useful information for ECG classification. Fractals have been shown to be useful for the characterisation of the ECG. However, it is well-known that the nonlinear features such as the Fractal Dimension are very sensitive to autocorrelation. For this reason, the method of Surrogate Data is generally used to assess the nonlinearity presence in a signal. It allows to take into account the signal autocorrelation in nonlinear analysis. However, it is a very computationally intensive process. A lot of time series must be generated, and nonlinear features must be computed on every series to obtain the necessary distribution. For this reason, the use of a single surrogate time series representing the most nonlinear time series from a power spectrum is proposed. Although nonlinear analysis is generally applied to long time series, Fractal dimension has been shown to extract useful information in ECG context with data of a duration of a single heartbeat. Hence, Fractal analysis will be compared for full time series and around points of interest (e.g. around QRS complex), with or without the modified Surrogate Data. Preliminary results have been obtained for the Fractal Dimension applied to the whole time series, without the Surrogate Data with simple neural networks. Although cross-validation gave a success rate of about 50% (with hard decisions), the algorithm was not able to give significant results for F_2 Score (0.041), G_2 Score (0.012) and Geometric Mean Score (0.022).