Scattering Transform of Heart Rate Variability for the Prediction of Ischemic Stroke in Patients with Atrial Fibrillation

Abstract

Summary Background: Atrial fibrillation (AF) is an identified risk factor for ischemic strokes (IS). AF causes a loss in atrial contractile function that favors the formation of thrombi, and thus increases the risk of stroke. Also, AF produces highly irregular and complex temporal dynamics in ventricular response RR intervals. Thus, it is hypothesized that the analysis of RR dynamics could provide predictors for IS. However, these complex and nonlinear dynamics call for the use of advanced multiscale nonlinear signal processing tools. Objectives: The global aim is to investigate the performance of a recently-proposed multiscale and nonlinear signal processing tool, the scattering transform, in predicting IS for patients suffering from AF. Methods: The heart rate of a cohort of 173 patients from Fujita Health University Hospital in Japan was analyzed with the scattering transform. First, p-values of Wilcoxon rank sum tests were used to identify scattering coefficients achieving significant (univariate) discrimination between patients with and without IS. Second, a multivariate procedure for feature selection and classification, the Sparse Support Vector Machine (S-SVM), was applied to predict IS. Results: Groups of scattering coefficients, located at several time-scales, were identified as significantly higher (p-value < 0.05) in patients who developed IS than in those who did not. Though the overall predictive power of these indices remained moderate (around 60 %), it was found to be much higher when analysis was restricted to patients not taking antithrombotic treatment (around 80 %). Further, S-SVM showed that multivariate classification improves IS prediction, and also indicated that coefficients involved in classification differ for patients with and without antithrombotic treatment. Conclusions: Scattering coefficients were found to play a significant role in predicting IS, notably for patients not receiving antithrombotic treatment. S-SVM improves IS detection performance and also provides insight on which features are important. Notably, it shows that AF patients not taking antithrombotic treatment are characterized by a slow modulation of RR dynamics in the ULF range and a faster modulation in the HF range. These modulations are significantly decreased in patients with IS, and hence have a good discriminant ability.