Abstract: Paroxysmal atrial fibrillation(PAF)is characterized by sudden onset and short duration， at present， electrocardiogram(ECG)is applied as the clinical diagnosis method， which is inconvenient for daily monitoring. Therefore， a noninvasive atrial fibrillation automatic detection method based on ballistocardiogram(BCG)was proposed. The optimal structure of one-dimensional convolutional neural network(CNN)for detecting PAF is achieved via matching different input data lengths and network depths. Through the test of 2000 sets of data， the best performances of the model proposed are: a test accuracy rate of 94.8%， a sensitivity of 97.2%， and a specificity of 92.7%， which provides the possibility of arrhythmia detection and remote daily home monitoring from BCG.

Key words: ballistocardiogram(BCG), electrocardiogram, convolutional neural network(CNN), paroxysmal atrial fibrillation(PAF), daily home monitoring