MemGCN: memory-augmented graph neural network for predict conduction disturbance after transcatheter aortic valve replacement

Transcatheter aortic valve replacement (TAVR) has recently emerged as an effective treatment for patients with severe symptomatic aortic stenosis. Accurate prediction of postoperative conduction disturbance after TAVR is crucial for successful surgery planning. Most current risk prediction models for conduction disturbance rely on monomodal data methods, such as images. However, predicting conduction disturbance based on multimodal features becomes a challenge. We propose a new memory-augmented graph neural network (MemGCN) to deal with this challenge. The GCN in the proposed network is constructed by taking the patient’s image features as nodes’ features and creating the edges based on the patient’s clinical features, thus effectively fusing the patients’ images and clinical feature information. Moreover, the traditional GCN model trains feature extractors and GCNs independently due to limited computing resources, which can negatively impact the feature learning process. We introduce the memory bank and propose a local update learning algorithm to address this issue. Specifically, the proposed method connects the feature extractors and GCN through a memory bank. The CNN module extracts patients’ features from their ECG images and updates the memory bank. The GCN constructs the graph using the patients’ characteristics in the memory bank as node features. In this way, the node features are cached in the memory bank and gradually updated at each optimization iteration so that the feature extractors and GCN are optimized jointly. The experiments on a real-world TAVR dataset show the proposed method’s superior performance. It improves the prediction accuracy by over 4% compared to the traditional GCN model. Furthermore, to demonstrate the applicability of our proposed method to other multimodal data analysis tasks, we evaluate the performance of the proposed method on an ocular disease intelligent recognition benchmark. The proposed method achieves competitive prediction performance.