A Novel Method to Identify Relevant Features for Automatic Detection of Atrial Fibrillation

The selection of an appropriate subset of predictors from a large set of features is a major concern in clinical diagnosis research. The purpose of this study is to demonstrate that the Multiple Kernel Learning (MKL) approach could be successfully applied as a feature selection process for machine learning pipelines. Furthermore, we suggest a multi-dynamic analysis of heartbeat signal to characterize the most common sustained arrhythmia, Atrial Fibrillation (AF). Indeed, we have targeted six different dynamics of QRS time series, where each one will be associated with 12 linear and nonlinear functions to yield a set of 72 features. Afterward, a feature selection process is implemented using the MKL to evaluate the relevant features allowing AF diagnosis. Hence, a subset of only 13 features has been selected. To demonstrate the effectiveness of the proposed approach, Support Vector Classification (SVC) model has been conducted, first, on all features, and then on the features issued from the MKL selection feature process. The obtained results showed that the SVC model trained by 13 features outperformed the one trained by 72 features. This approach has reached 99.77% of success rate in the discrimination between Normal Sinus Rhythm (NSR) and AF. The proposed selection feature method holds several interesting properties in dimensionality reduction which makes it a suitable choice for several applications.