Shapelet Discovery for Atrial Fibrillation Detection

The main goal of this study is to develop an automatic classification algorithm for normal sinus rhythm (NSR) versus Atrial Fibrillation (AF) from a single channel short ECG segment. For this purpose, AF and NSR from PhysioNet/Computing in Cardiology Challenge 2017 training dataset were used in this study. RR intervals were extracted using GQRS algorithm, and RR time series with less than 30 beats were excluded from the analysis. The stratified split was applied to create a training set (NSR: 1521 and AF: 239) and test set (NSR: 1527 and AF:234). Shapelets were extracted by scanning RR time series in the training set and identifying statistically significant patterns. For classification, an XGBoost model was trained using the presence or absence of the top 100 shapelets. Using the top 100 significant shapelets (Shapelet length between 5 and 29), we achieved the area under the ROC curve (AUC) and the area under the precession recall curve (AUPRC) of 0.94 and 0.77 in discrimination between AF and NSR. Among the top 100 significant shapelets, we used all shapelets with length no greater than a certain threshold (maximum acceptable Shapelet length) for training different models. Increasing the number of shapelet features by varying the threshold from 5 to 30 in the model training improved AUC/AUPRC 0.91/0.68 to 0.94/0.77