Started working on a pipeline using all data. Issues with memory (RAM) while generating one-hot label encodings. Solutions: use data generators to train via fit_generator. Examples:
https://github.com/keras-team/keras/issues/1627
https://www.kaggle.com/ezietsman/simple-keras-model-with-data-generator
https://stanford.edu/~shervine/blog/keras-how-to-generate-data-on-the-fly.html

Final model with balanced s2s training data and ICD10 without additional data points from CC/CB combination.

Added ICD 10 classification with attention (best results), attention with context (worst results) and char level classification (running atm).