This is the repo for the COLING 2022 paper Generation of Patient After-Visit Summaries to Support Physicians (Paper link will be available upon publication)

Note that the model & results reported in this paper were obtained based on patient data from University of Massachusetts, Chan Medical School. Due to privacy issue, we could not open source our datasets.

The good news is, we have recently found the MIMIC-III also contains data applicable to our task!

While MIMIC-III is a de-identified and publicly available dataset, you still need to apply to access the dataset on the website.

Our after-visit summary generation models pre-trained on MIMIC-III datasets is available HERE. To obtain the train/validation/test datasets, please refer to the instructions below.

Once you have been granted access to MIMIC-III, download NOTEEVENTS.csv, which contains a table of de-identified physician notes.

Run the following command:

python scripts/data_process/extract_mimic3.py \
    --input_file path/to/NOTEEVENTS.csv \
    --output_dir path/to/output/data/directory

This command will extract hospital course - AVS pairs from the downloaded file NOTEEVENTS.csv, and randomly split them into train/validation/test datasets in the output_dir as you specified.

Specifically, the train.json, validation.json and test.json should be dictionaries conforming to the following format:

{
    {
        "text": "Hospital Course Content 1"
        "summary": "After-Visit Summary Content 1"
    },
    {
        "text": "Hospital Course Content 2"
        "summary": "After-Visit Summary Content 2"
    },
    ...
}

To train / evaluate the summarization model, run the following commnand:

python run_summarization.py \
    --model_name_or_path allenai/led-base-16384 \
    --do_train \
    --do_eval \
    --do_predict \
    --train_file path/to/train.json \
    --validation_file .path/to/val.json \
    --test_file path/to/test.json \
    --output_dir path/to/output_dir \
    --overwrite_output_dir \
    --per_device_train_batch_size=8 \
    --per_device_eval_batch_size=8 \
    --predict_with_generate \
    --num_train_epochs 6 \
    --save_steps 3000 \
    --save_total_limit 3 \
    --max_source_length 1024 \
    --max_target_length 512 \

If you just wish to generate results using a pre-trained model, run the following command:

python run_summarization.py \
    --model_name_or_path path/to/pretrained/model \
    --do_predict \
    --test_file path/to/test.json \
    --output_dir path/to/output_dir \
    --overwrite_output_dir \
    --per_device_eval_batch_size=8 \
    --predict_with_generate \
    --num_train_epochs 6 \
    --save_steps 3000 \
    --save_total_limit 3 \
    --max_source_length 1024 \
    --max_target_length 512 \

The results will be generated in the output_dir

To train& apply our error detection model requires using MetaMap to extract UMLS concepts from hospital course notes and AVS notes.

Please install and set up MetaMap as instructed, and then prepare the following folders:

• train_src_txt/ contains hospital course notes, each hospital course note should be stored in a single txt file, e.g. src_txt_0.txt;
• train_src_ent/ contains UMLS concepts extracted by MetaMap from hospital course notes, the concepts extracted from a hospital course note txt file should be stored in a json file starting with the name of the txt file, e.g. src_txt_0.txt.json;
• train_tgt_txt/ contains AVS notes, similarly, each AVS note should be stored in a single txt file, e.g. tgt_txt_0.txt;
• train_tgt_ent/ contains UMLS concepts extracted by MetaMap from AVS notes, the naming criterion is similar to above , e.g. tgt_txt_0.txt.json;

Specifically, the json file in train_src_ent/ and train_tgt_ent/ should adopt the following format:

[
    {"begin": "61", "end": "80", "CUI": "C0037278", "orgi_term": "of a skin infection", "pref_term": "Skin Diseases, Infectious", "sem_type": "dsyn"}, 
    {"begin": "94", "end": "105", "CUI": "C0003237", "orgi_term": "antibiotics", "pref_term": "Antibiotics, Antitubercular", "sem_type": "antb"},
    ...
]

where begin and end refers to the position where the UMLS concept begins and ends in the text, CUI is the unique concept identifier in the UMLS, orgi_term and pref_term are the original and preferred term of the concept, sem_type is the semantic type. All this information could be provided by MetaMap.

Following the same manner, build the following folders: valid_src_txt/, valid_src_ent/, valid_tgt_txt/, valid_tgt_ent/, test_src_txt/, test_src_ent/, test_tgt_txt/, test_tgt_ent/.

To train / evaluate the hallucination detection model, you need to prepare the data through the following two steps:

1. To prepare the data for hallucination detection model, run the following command (You will need to revise a few hyper-parameters in the file before running it):

sh shell/prepare_data_err_detection.sh

2. To train the hallucination detection model, run the following commnand:

python ./run_err_det.py \
  --model_name_or_path google/bigbird-roberta-base \
  --train_file path/to/your_err_det_data_dir/train.json \
  --validation_file path/to/your_err_det_data_dir/val.json \
  --output_dir path/to/your_err_det_model_dump_dir \
  --save_steps 3000 \
  --do_train \
  --do_eval \
  --do_predict \
  --per_device_train_batch_size 8 \
  --per_device_eval_batch_size 8 \
  --num_train_epochs 3 \
  --hallucination_weight 3.0 

1. To prepare the data for key event identification model, run the following command: (Require your implementation of accessing to your own UMLS database)

python ./scripts/data_preprocess/miss_det-build_dataset.py \
    --input_src_txt_dir path/to/train_src_txt/ \
    --input_src_json_dir path/to/train_src_ent/ \
    --input_tgt_json_dir path/to/train_tgt_ent/ \
    --output_dir path/to/your_ke_idt_data_dir

2. To train the key event identification model, run the following commnand:

python ./run_err_det.py \
  --model_name_or_path google/bigbird-roberta-base \
  --train_file path/to/your_ke_idt_data_dir/train.json \
  --validation_file path/to/your_ke_idt_data_dir/val.json \
  --output_dir path/to/your_ke_idt_model_dump_dir \
  --save_steps 3000 \
  --do_train \
  --do_eval \
  --do_predict \
  --per_device_train_batch_size 8 \
  --per_device_eval_batch_size 8 \
  --num_train_epochs 3