Compact Representations of Patient States using Unsupervised Deep Learning Techniques - overview


Abstract:

The application of deep learning in healthcare has become popular in recent years and early research has indicated that the use of Electronic Health Records (EHR) can help derive patient representations for improved clinical predictions. Common architectures of Artificial Neural Networks, including supervised Feed-forward Neural Network (MLP), CNN, RNN and unsupervised Autoencoders (AE, RBM), have been applied to the problem of efficient EHR representation for outcome prediction. The objective is to investigate unsupervised deep learning techniques in producing compact patient embeddings using EHR data including model explainability.

Requirements:

  • Knowledge of artificial neural networks
  • Preferred but not essential:  experience with Python, Tensorflow and Keras