Atrial fibrillation detection service validation tool

Abstract We developed a software tool to validate a deep learning algorithm for an atrial fibrillation detection service with heart rate data from a clinical study. The deep learning algorithm analyses the measurement data and establishes an estimated atrial fibrillation probability for each heartbeat. The software tool displays both data and deep learning analysis results. Furthermore, the graphical user interface can be used by medical experts to detect atrial fibrillation periods in the data and establish a reference result which will be treated as ground truth in subsequent result analysis steps. Once both deep learning and expert results are available, a confusion matrix is produced and the algorithm performance is validated by establishing accuracy, sensitivity, specificity, and f1-score. The software tool was created in Python and the software incorporated a graphical user interface as well as functional elements for data display and deep learning. To establish the required functionality, we used three different parallel processing methods for: (1) user interface processing, (2) data handling, and (3) deep learning. This highlights the need for parallel processing methods even for projects with a low or mid-range complexity. We have learned that the functionality of individual components can be expressed elegantly in Python. However, the lack of parallel debugging support makes it rather difficult to integrate functional components to establish a working solution.