Updates

Clone the repo

git clone https://github.com/AI-in-Hospitals/Patient-Instructions.git# clone the following repo to calculate automatic metricscd Patient-Instructiongit clone https://github.com/ruotianluo/coco-caption.git

Environment

conda create -n pi python==3.9conda activate pipip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.htmlpip install transformers==4.10.0pip install pytorch-lightning==1.5.1pip install pandas rouge scipy# if you want to re-produce our data preparation processpip install scikit-learn plotly

Download the data

Patient-Instructions/ # the root of the repo    data    ├── README.md    ├── prepare_dataset.ipynb    ├── prepare_subtasks.ipynb    ├── diagnose-procedure-medication    │   ├── admDxMap_mimic3.pk      # Source: D_ICD_DIAGNOSES.csv    │   ├── admMedMap_mimic3.pk     # Source: prescriptions.csv    │   ├── admPrMap_mimic3.pk      # Source: D_ICD_PROCEDURES.csv    │   └── readme.txt    ├── splits                      # Source: NOTEEVENTS.csv    │   ├── train.csv               # obtained by data/prepare_dataset.ipynb    │   ├── val.csv                 # obtained by data/prepare_dataset.ipynb    │   ├── test.csv                # obtained by data/prepare_dataset.ipynb    │   └── subtasks                    │       ├── age                 # Source: NOTEEVENTS.csv    │       │   └── ...             # obtained by data/prepare_subtasks.ipynb    │       ├── sex                 # Source: NOTEEVENTS.csv    │       │   └── ...             # obtained by data/prepare_subtasks.ipynb    │       └── diseases            # Source: NOTEEVENTS.csv, D_ICD_DIAGNOSES.csv    │           └── ...             # obtained by data/prepare_subtasks.ipynb    └── vocab                      ├── special_tokens_map.json # obtained by data/prepare_dataset.ipynb       ├── tokenizer_config.json   # obtained by data/prepare_dataset.ipynb       └── vocab.txt               # obtained by data/prepare_dataset.ipynb

Pretreatments

# Generate the adjacent matrix of all unique digonosis, medication, and procedure codes# This is essential if we want to use knowledge graph to assist PI-Writerpython pretreatments/prepare_codes_adjacent_matrix.py# Get top-300 most similar admission records from the training set for each query hospital admission# This is essential if we want to retrieve historical PIs to assist PI-Writerpython pretreatments/prepare_relevant_info.py# Use bert-base-uncased to extract sentence-level embeddings of PIs# We apply max pooling on the word embs of the last layerpython pretreatments/extract_instruction_embs.py

Training

python train.py --gpus 8 --batch_size 8 --arch base --setup transformerpython train.py --gpus 8 --batch_size 8 --arch base --setup transformer_Fullpython train.py --gpus 8 --batch_size 4 --accumulate_grad_batches 2 --arch base --setup transformer_Fullpython train.py --gpus 8 --batch_size 8 --arch small --setup lstmpython train.py --gpus 8 --batch_size 8 --arch small --setup lstm_Full

Evaluation

python translate.py $path_to_model

# The ouput file will be saved to `./inference_results/preds_and_scores.json` in this casepython translate.py $path_to_model --save_json --save_base_path ./inference_results --save_folder "" --json_file_name preds_and_scores.json

python translate.py $path_to_model --subtask_type agepython translate.py $path_to_model --subtask_type sexpython translate.py $path_to_model --subtask_type disease

Bugs or Questions?

Citation

@inproceedings{liu2022retrieve,title={Retrieve, Reason, and Refine: Generating Accurate and Faithful Patient Instructions},author={Liu, Fenglin and Yang, Bang and You, Chenyu and Wu, Xian and Ge, Shen and Liu, Zhangdaihong and Sun, Xu and Yang, Yang and Clifton, David A.},booktitle={Advances in Neural Information Processing Systems},year={2022}}

Acknowledgements