A feasibility study on the automatic detection of atrial fibrillation using an unobtrusive bed-mounted sensor

We present a feasibility study on the automatic detection of atrial fibrillation (AF) from a cardiac vibration signal (ballistocardiogram). Signals were recorded by means of an electromechanical foil attached to a bed's mattress. A clinical study with 10 AF patients was conducted to assess whether ballistocardiograms (BCG) provide sufficient information to automatically distinguish atrial fibrillations from normal sinus rhythms. For this purpose, the BCGs were split into 30 s long epochs which were manually labelled as AF or sinus rhythm. Using features extracted from an autoregressive time-frequency representation of the BCG, a support vector machine classifier was trained to detect AF epochs. The classifier was evaluated on a set of 245 epochs by means of leave-one-out cross-validation. Our results (sensitivity: 96.2% / specificity: 91.9%) suggest that it is indeed feasible to use bed-mounted BCG sensors to screen for atrial fibrillations.