Comprehensive Summary
In this study, researchers trained and validated a neural network to accurately identify the manufacturer and model group of cardiac rhythm devices from radiographic images. 1676 images were sourced from the Imperial College Healthcare NHS Trust and divided into training (1451 images) and testing (45 images) sets. During the development phase, five different convolutional network architectures (DenseNet, Inception V3, VGGNet, ResNet, and Xception) were evaluated in a 4-fold cross validation using the training set. Networks were tasked with identifying the device’s manufacturer (from five possible manufacturers) and model group (from 45 possible groups). Out of the five networks, Xception was found to be the network with the lowest loss (0.34) and highest accuracy (91.1%). Following training, the neural network was assessed with the test set images. Accuracy for identification of the manufacturer was 99.6% while accuracy for model group was 96.4%. This accuracy was significantly higher than the median (72.0%) of five human cardiologists’ accuracies and the highest single human accuracy (88.9%).
Outcomes and Implications
Identification of cardiac rhythm device manufacturers is essential for hospital staff to be able to interact with and program these devices during emergencies. However, human identification of these devices from radiographic images is time-consuming and prone to error, both factors which can lead to adverse health outcomes. This freely available neural network can identify both manufacturer and model type with a higher accuracy than experienced clinicians, which can possibly save vital time and reduce identification error in a cardiac emergency. Additionally, this neural network creates a saliency map identifying the features that had factored most into its decision, which can help medical staff learn what features to look for during human determination.