A Metric Learning Based Unsupervised Domain Adaptation Method with Its Application on Mortality Prediction

Deep learning models have been widely used in the field of healthcare prediction tasks and have achieved good results in recent ears. However, deep learning models often face the problems of insufficient labeled training data, the overall data distribution shift, and the category level data distribution shift, which leads to a decrease in the accuracy of the models. To solve the above problems, we propose an unsupervised domain adaptation method based on metric learning (additive margin softmax based adversarial domain adaptation, AMS-ADA). Firstly, this method uses the long short-term memory network with the attention mechanism to extract features. Secondly, this method introduces the idea of the generative adversarial network and reduces the overall data distribution shift via adversarial domain adaptation. Thirdly, this method introduces the idea of metric learning, which further reduces the category level data distribution shift by maximizing the decision boundary in the angular space. This method improves the effect of domain adaptation and the accuracy of the model. We perform the mortality prediction task of ICU patients in real-world healthcare datasets. The experimental results show that compared with other baseline models, our method can better solve the problem of data distribution shift and achieve better classification accuracy.