A hybrid two-stage approach for paroxysmal atrial fibrillation prognosis problem

We develop a hybrid two-stage approach for paroxysmal atrial fibrillation (PAF) prognosis based on features extracted from short-term heart rate variability (HRV) sequences. At the first stage, a data-mining-based approach is used to identify crucial medical-oriented features that can distinguish PAF HRV sequences from non-PAF HRV ones. However, PAF patients can experience PAF without exhibiting the medical-oriented features. To detect this type of patients, at the second stage, we employ a machine-learning-based approach to select certain nonlinear features that can classify HRV sequences into classes of PAF or non-PAF The developed approach was trained on the PAF Prediction Challenge Database and was tested on the dataset consisting of minute HRV episodes extracted from MIT-BIH Atrial Fibrillation Database and the MIT-BIH Normal Sinus Rhythm Database. It was obtained from the numerical evaluation that the developed approach achieved about 85% of accuracy in short-term prognosis of PAF by using the first stage approach alone and around 90% of accuracy with the combination of both stages. Furthermore, the developed medical-oriented features can be clinically valuable to the cardiologists for providing insights to the initiation of PAF.