Detection of Atrial Fibrillation Using Multi-Site Ballistocardiogram with Piezoelectric Rubber Sheet Sensors

Ballistocardiography (BCG) is a noninvasive modality for detecting cardiac activity. This study developed a robust atrial fibrillation (AF) detection algorithm using multiple BCG sensors at different locations and evaluated the improvement in accuracy by combining data from multiple sensors. We recorded the BCG using a piezoelectric rubber sheet sensor and an electrocardiogram in 84 participants (29 with AF and 55 without AF) in the supine position. Four BCGs (BCG1–4) were obtained using sensors placed from the head to the lumbar region (0, 25, 45, and 65 cm from the head). The BCG signals were divided into 32 s blocks and analyzed. After applying fast Fourier transform, we input the power spectrum, focusing on frequencies below 10 Hz, into machine learning (ML) classifiers to distinguish between AF and non-AF with parameter tuning. The AdaBoost classifier for BCG2 exhibited the highest accuracy (0.88) among the ML models for each sensor. When we applied the classifier to other BCGs, it achieved accuracies of 0.92, 0.73, and 0.78 for BCG1, 3, and 4, respectively. The combined model using multiple sensors exhibited an accuracy of 0.92. The optimized model for BCG2 was robust against shifts in the sensor toward the head and lumbar directions. A combined assessment using multiple sensors improved performance.