A Data Multiple Imputation Method for Mortality Prediction in Patients with Heart Failure

Aiming at the problem that the imperfect data collection mechanism of diagnosis and treatment leads to data loss and the poor quality of existing network feature extraction leads to the limited performance of clinical prediction models, a data multiple imputation method (Self-attention and Generative adversarial network based Mortality Prediction, SGMP) for mortality prediction of patients with heart failure is proposed. Firstly, the self-attention mechanism is used to dynamically fuse multiple sets of candidate estimates in the potential space of the variational autoencoder (VAE), and the adversarial training strategy of the generative adversarial network (GAN) is introduced to optimize the representation learning ability. Then, the candidate estimation results are effectively obtained according to the mask matrix, and the multiple interpolation of missing data is realized. Finally, the synthetic minority over-sampling technique (SMOTE) is used to enhance the data, and the multilayer perceptron (MLP) is used to predict the mortality rate. Based on the diagnosis and treatment data of heart failure patients in a tertiary hospital in Xinjiang, the results show that: in the mortality prediction task, SGMP has significantly improved in multiple indicators compared with other models, and the area under the receiver operating characteristic curve reaches 0.902.