Recent research development has demonstrated the advantages of deep learning models in prediction tasks on
electronic health records (EHR) in the medical domain. However, the prediction results tend to be difficult to
explain due to the complex neuron structures. Without the explainability and transparency, deep learning
models are not trustworthy or reliable for making real world decisions, especially the high-stakes ones in the
healthcare domain. To improve the trustworthiness of the deep learning model, quantifying the uncertainty is
crucial.
In this dissertation work, we proposed several Bayesian Neural Network (BNN) structures to estimate the data
uncertainty and model uncertainty associated with the EHR data and deep learning models, respectively. We
also proposed Variational Neural Network (VNN) algorithms to estimate the uncertainty of the variables to
investigate the medical and temporal features that contribute the most to the patient-level uncertainty. In order
to verify the validity of the uncertainty estimations, we designed a series of experiments to examine the
computational results against widely accepted facts about uncertainty. We also conducted post-hoc analysis to
evaluate whether the proposed models tend to specialize in one or more patient subgroups, at the cost of
model performance on others, as well as whether the treatment (improving uncertainty in one subgroup) will
mitigate such performance cost. The experiment results have confirmed the validity of our computational
approaches. Finally, we conducted a user study to understand the clinicians' perception of the proposed
uncertainty models.