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).

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

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