Recent years have witnessed significant interest in exploiting machine learning models for clinical decision support. However, due to the high dimensionality and noisiness of clinical data, it is often challenging to develop accurate, robust and interpretable representation of clinical data for predictive models. This thesis work investigates representation learning based on variational autoencoder (VAE) in the context of clinical applications.  Specifically, this thesis makes three contributions to the state of the art of representation learning for clinical data: 1) explicit expression of prediction and disentangled latent space; 2) continuous update of predictions with the inflow of dynamic time series; 3) integration of static and time-series data in deep latent variable models. This research is evaluated using real-world preoperative datasets and decision support applications.

Recent years have witnessed significant interest in exploiting machine learning models for clinical decision support. However, due to the high dimensionality and noisiness of clinical data, it is often challenging to develop accurate, robust and interpretable representation of clinical data for predictive models. This thesis work investigates representation learning based on variational autoencoder (VAE) in the context of clinical applications.  Specifically, this thesis makes three contributions to the state of the art of representation learning for clinical data: 1) explicit expression of prediction and disentangled latent space; 2) continuous update of predictions with the inflow of dynamic time series; 3) integration of static and time-series data in deep latent variable models. This research is evaluated using real-world preoperative datasets and decision support applications.