A filter approach for feature selection in classification: application to automatic atrial fibrillation detection in electrocardiogram recordings

International audience ; Background: In high-dimensional data analysis, the complexity of predictive models can be reduced by selecting the most relevant features, which is crucial to reduce data noise and increase model accuracy and interpretability. Thus, in the field of clinical decision making, only the most relevant features from a set of medical descriptors should be considered when determining whether a patient is healthy or not. This statistical approach known as feature selection can be performed through regression or classification, in a supervised or unsupervised manner. Several feature selection approaches using different mathematical concepts have been described in the literature. In the field of classification, a new approach has recently been proposed that uses the γ-metric, an index measuring separability between different classes in heart rhythm characterization. The present study proposes a filter approach for feature selection in classification using this γ-metric, and evaluates its application to automatic atrial fibrillation detection. Methods: The stability and prediction performance of the γ-metric feature selection approach was evaluated using the support vector machine model on two heart rhythm datasets, one extracted from the PhysioNet database and the other from the database of Marseille University Hospital Center, France (Timone Hospital). Both datasets contained electrocardiogram recordings grouped into two classes: normal sinus rhythm and atrial fibrillation. The performance of this feature selection approach was compared to that of three other approaches, with the first two based on the Random Forest technique and the other on receiver operating characteristic curve analysis. Results: The γ-metric approach showed satisfactory results, especially for models with a smaller number of features. For the training dataset, all prediction indicators were higher for our approach (accuracy greater than 99% for models with 5 to 17 features), as was stability (greater than 0.925 regardless of ...