Abstract: Based on the arrhythmia characteristics of atrial fibrillation(AF)， phase space reconstruction(PSR)is applied to extract the 2D rhythmic feature of the ballistocardiogram(BCG)signal. In the process of reconstruction， the optimal embedding dimensions and time delay parameters are discussed. Firstly， the heart beat is regarded as a non-linear dynamic system， and the 1D time series is mapped to the high-dimensional phase space， based on the phase space reconstruction theory， to obtain the 2D trajectory， which describes the abnormal rhythm of AF in BCG signal. Secondly， the optimal embedding dimension and time delay parameters of the reconstruction procedure for AF diagnosis are discussed， and the convolutional neural network(CNN) is applied to identify AF automatically. Finally， 2000 BCG segments from 59 subjects are used to validate the classification performance. The accuracy reaches 91.00% by means of the tenfold cross-validation. Compared to the machine learning method based on the classical time-frequency features， the accuracy is improved， which verifies the superiority of the proposed method.

Key words: ballistocardiogram(BCG) signal; electrocardiographic(ECG) signal; atrial fibrillation(AF) detection; phase space reconstruction(PSR); convolutional neural network(CNN)