TensorFlow application: LSTM with Word2Vec on a medical database.
- Remove the patientnumbers and days from the states to get more general events
- Represent the days as differences between the events -> Better representation of the problem BUT still keep the date also to find relations with certain time periods as winter
- Make the LSTM network
- Look into DeepLearning4J as it offers a lot more documentation and feels a lot more natural to program in. It also offers LSTM and Sequence2Vec!
- Check graph links between diseases after word2vec (like in paper) (with TSNE)
- Check if preprocessing has effect on model with cross-validation
- Check how well word2vec works with complete new examples with cross-validation
- General accuracy of model of course