A Novel Kernel-Based Method for Atrial Fibrillation Diagnosis

Zouhair Haddi1, Bouchra Ananou2, Stéphane Delliaux3, Mustapha Ouladsine4
1Aix-Marseille Université, Laboratoire d’Informatique et Systèmes,, 2Aix Marseille Univ., Université de Toulon, CNRS, LIS, Marseille, France, 3Aix-Marseille Univ., IRBA, DS-ACI, Marseille, France, 4Aix-Marseille Université, Université de Toulon, CNRS, LIS, Marseille, France


Abstract

Aim: This study aimed to develop an efficient diagnosis method for atrial fibrillation (AF) arrhythmia based on inter-beat interval time series analysis and relevance vector machine (RVM) classifier. Motivation: Automatic and fast AF diagnosis are still a major concern for the healthcare professional. Several algorithms based on univariate and multivariate analysis have been developed to detect AF. The published results do not show satisfactory detection accuracy especially for brief duration as short as one minute. Although RVM has been applied on tasks such as computer vision, natural language processing, speech recognition etc., this is the first attempt to adopt RVM for AF diagnosis. Methods: Four publicly-accessible sets of clinical data (AF Termination Challenge Database, MIT-BIH AF, Normal Sinus Rhythm RR Interval Database, and MIT-BIH Normal Sinus Rhythm Databases) were used for assessment. All time series were segmented in 1 min RR interval window and then three specific features were calculated, namely, Vector Angular Index, Vector Length Index and Dispersion of points along the perpendicular to the diagonal line. The features of the four databases were merged in order to give rise huge variability and therefore to better characterize AF rhythm. The RVM classifier was trained on 2000 randomly selected samples from the merged database. Several kernel functions have been explored and the Gaussian radial basis function has been selected to build the proposed method. Results: The results showed that the RVM model performed better than do existing algorithms, with 99.20% for both sensitivity and specificity. Of the 1000 AF segments of 1 min included in this study, 992 AF segments have been recognized as AF. Conclusion: The RVM diagnosis method holds several interesting properties and can be implemented with few arithmetical operations which makes it a suitable choice for real-time medical-decision-making operations.