Automatic Atrial Fibrillation Detection Based on BCG Using Signal Transformation and Random Kernels Convolution

Atrial fibrillation (AF) is one of the most common cardiac arrhythmias worldwide, posing a significant threat to patients’ health and a heavy burden to national economy. Developing early screening methods for AF is crucial. Thus, an automatic AF detection method based on noncontact ballistocardiogram (BCG) signals collected from unobtrusive bed-mounted sensors was proposed. BCG signals from 50 subjects during their sleep were collected, which were divided into nonoverlapping 30-s segments, including 12311 AF segments and 9940 sinus rhythm (SR) segments. After signal preprocessing, the autocorrelation transformation and energy signal transformation for data augmentation were implemented, followed by random kernels convolution. Dual pooling methods, including maximum value and proportion of positive value (PPV), were employed to extract features afterward. A logistic regression classifier distinguished AF from SR, with model generalizability rigorously validated through fivefold cross-validation and intersubject testing. The AF detection achieved the average precision, recall, ${F}1$ -score, and accuracy of 96.5%, 97.3%, 96.9%, and 96.1%, respectively.