A Data-driven Approach to Atrial Fibrillation Detection from Overnight Polysomnographic Sleep Recordings

Atrial fibrillation (AF) is the most prevalent arrhythmia. Despite the fact we spend on average about one third of our lives sleeping, very little research has focused on overnight AF analysis. We hypothesize that AF events may be identified during sleep from beat-to-beat interval time series. We further hypothesize that the identification of AF will not be impaired by the presence of sleep disordered breathing such as obstructive sleep apnea (OSA). A total of four databases totaling n=3,086 patients were used for this research. Among them, the Sleep Heart Health Study (SHHS, n=2966) database consists of polysomnographic (PSG) recordings, i.e. overnight recordings. The SHHS, considered as test set, was used to evaluate the feasibility to identify AF rhythm from PSG recordings. In the SHHS a total of 57 patients had AF rhythm and 2904 had non-AF rhythm. A random forest classifier was optimized using cross-validation to classify AF events, based on 8 features derived from the beat-to-beat time series. On the SHHS, we obtained overall statistics of Se=0.96, Sp=0.99, NPV=0.99, PPV=0.83 in classifying individuals with or without AF. We obtained similar performance for both OSA and non-OSA patients. We showed that AF individuals may be identified from their sleep PSG recordings, and that the accuracy of the detection is not affected by the presence of OSA. AF detection from overnight PSG recordings may enable better phenotyping of OSA by identifying individuals whom cardiac function may have been affected by OSA.