Spectral Attention Feature Selection for Patient Mortality Prediction

Patient representation learning for disease prediction is becoming increasingly important. EHRs as an crucial data source for patient representation learning. However, due to the large number of EHRs features, only some of them are relevant to disease prediction. Secondly, a large number of features will also bring redundant information, thereby reducing the accuracy of the model. To address this issue, we use the attention mechanism to learn the correlation of features in EHRs. We propose a spectral attention feature selection framework that not only reduces time overhead but also minimizes the impact of redundant features. We apply our method to existing healthcare models, and the experimental results verify the effectiveness of our method.