RGD-VNet: Raw, Generative, and Discriminant Views Network for Boosting Postoperative Complication Prediction

Dapeng Tao, Chunxiao Quan, Yaosheng Hu, Yiqiang Wu, Yibing Zhan, Hua Jin

Abstract: Postoperative complications are adverse reactions caused by anesthesia and surgical trauma and severely affect patients’ recovery and life. To reduce the risk of postoperative complications, postoperative complication prediction (PCP) has been proposed to predict the probability of various postoperative complications and help physicians to take interventions in advance. However, most existing work still depends on expensive labeled data and scarcely utilizes unlabeled data, resulting in limited prediction performance. In this paper, we propose a novel framework named Raw, Generative, and Discriminant Views Network (RGD-VNet) for PCP that can better mine the internal connections between labeled and unlabeled data. Specifically, RGD-VNet contains two modules: an unsupervised view encoder and a multi-view consistency regularizer. In the unsupervised view encoder, we construct a hybrid and unsupervised representation encoder to encode raw, generative, and discriminant views (RGD-Views), into a unified representation. In the multi-view consistency regularizer, due to the distribution difference among varying views, we design a view consistency loss function to strengthen the relationship among RGD-Views. In the experiment, we adopt four common postoperative complications, including pain, dizziness, nausea, and vomiting, to show the effectiveness of the hybrid RGD-Views. The experimental results also demonstrate the superiority of RGD-VNet and achieve the SoTA performance.

External IDs:doi:10.1145/3594315.3594649