DL4H_final

Code

• Citation to the original paper

  Fatemeh Amrollahi et al. “Contextual embed- dings from clinical notes improves prediction of sepsis”. In: AMIA Annual Symposium Pro- ceedings. Vol. 2020. American Medical Infor- matics Association. 2020, p. 197

• Dependencies

  1. Text preprocessing: text/preprocessing_notes.ipynb to clean the notes, and then use text/to_embeddings.ipynb to convert the BERT embeddings. text/to_tfidf.ipynb to calculate the TF-IDF
  2. data to tensor: produce_tensor/Check features in chartevents.ipynb and store_data_in_tensor_5-8-22.ipynb.
  3. dataloader, training model: training/training_BERT_structured.ipynb, training/training_structured.ipynb

• Data download instruction

  Data is available in the website (see https://physionet.org/ content/mimiciii/1.4/) once you pass the train- ing courses (see https://eicu-crd.mit.edu/ gettingstarted/access/).

• Table of results

  In this project, we implemented model 2 (structured features only), and model 4 (structured features + ClinicalBERT). We have construct TF-IDF embeddings, but the tensor is too large to be fit into RAM, so we didn't train model 3 (TF-IDF + structured features)

  We got AUC = 0.875 for model 2.

  For model 4, we only have time to train several epochs and didn't tune the learning rate properly, so we got AUC = 0.5.