The work proposed to apply dropout and l2 weight decay [7] on the model, and add input dependent noise on stochastic attention model. The evaluation compared the results on electronic health record datasets with conventional approaches that use soft attention and stochastic hard attention (which use input-independent noise). The work also evaluated the quality of feature selection by the attention mechanism through comparing it with human selected features, and shown improvement compared to the soft attention. The authors show that the resulting uncertainty prediction allows their models to filter out more false negative. Being able to provide uncertainty of model prediction is a critical criterion for medical domain, and the proposed approach provide reasonable improvement over existing approaches.