[7A3] Validating a deep-learning model for atrial fibrillation detection
M Kareem and O Faust
Sheffield Hallam University, UK
This study validates a deep learning model that was designed to detect atrial fibrillation (AF) episodes in heart rate (HR) signals. The model is based on a recurrent neural network (RNN) with long short-term memory (LSTM). It was trained and tested with data from 23 subjects. 10-fold cross validation resulted in 98.51% accuracy for AF detection. We recognise that 10-fold cross validation means to test the model with unknown signal segments from a known subject. Such a test does not reflect a use case scenario where the deep learning model must classify HR signals from unknown subjects. In this study we address this problem by using the data from 18 subjects, different from the patients with which the model was trained, for blind-fold validation. Based on this dataset the model achieved 97% accuracy. This result indicates that the deep learning model could extract relevant feature maps from unseen data and thereby AF signs could be detected. The outcome is important for the practical realisation of computer aided diagnosis (CAD) systems, because it shows that knowledge extracted from a small training dataset was successfully applied to HR traces from unknown subjects. With that, we are one step closer to mimicking human decision-making, which is also based on knowledge extracted from a limited known dataset.
