This paper presents an applications of RNNs to predict "clinical events", such as disease diagnosis and medication prescription and their timing.

The paper proposes/suggests:
1. Applying an RNN to disease diagnosis, medication prescription and timing prediction.

2. "Initializing" the neural net with skipgrams instead of one-hot vectors. However, it seems from the description that the authors are not "initializing", rather just feeding a different feature vector into the RNN.

3. Initializing a model that is to be trained on a small corpus from a model trained on a large corpus works. Concludes: information can be transferred between models (read across hospitals).