Atrial fibrillation (AF) is one of the most common heart diseases in population. Timely diagnosis of AF is challenging due to the asymptomatic and episodic nature of the disease. It is therefore necessary to develop methods that can identify patients with AF using electrocardiographic data of normal sinus rhythm when there's no abnormal rhythm present in recordings. Models based on convolutional neural networks have been successful using 12-lead ECGs with high sampling rates. We believe it is possible to solve the problem with more generalised biomarker data of heart rate variability. We consider recurrent neural networks for this task. In this paper, we consider a convolutional recurrent neural network (CRNN) model with LSTM layers and compare its classification performance to a convolutional neural network (CNN) with a global average pooling operator. In addition, we generate attention maps of the CRNN model and demonstrate the peculiarities of its decision mechanism. Open PhysioNet data is used for model training; data from a local clinical hospital is used for model validation. In general, the CRNN model shows better patient classification results and provides interpretable attention maps with GradCAM++ method.