jayaos / rnn-covid Goto Github PK

This repository contains source codes for implementing models and reproducing the results in the following paper:

Severity Prediction for COVID-19 Patients via Recurrent Neural Networks
    Junghwan Lee, Casey Ta, Jae Hyun Kim, Cong Liu*, Chunhua Weng*
    AMIA Informatics Summit 2021, https://doi.org/10.1101/2020.08.28.20184200

The source codes were written in Python 3.7.1.

1. Clone this repository.

2. Move to the directory where you clone this repository and install requirements

    pip install -r requirements.txt
   

The RNN model and baselines require patient records, demographic information of the patients, patient labels, and pre-trained embedding (optional). Example format of the data is available at /data.

• Patient records: medical records of the patients. List of lists saved in pickle file format. For example, there are two patients who made two visits, then corresponding patient records are [[[1,2] , [3,4,5]], [[1,3], [4,5,6]]]. This record can be interpreted as the first patient has concept 1 and 2 at the first visit and concept 3, 4, and 5 at the second visit; and the second patient has concept 1 and 3 at the first visit and concept 4, 5, and 6 at the second visit. For efficient handling of the data, all medical concepts are required to be mapped to integer and a mapping dictionary has to be provided to implement the models. For example, {“concept 1” : 0, “concept 2” : 1} represents “concept 1” and “concept 2” are mapped to integer code 0 and 1. Those mapped integer codes are used in representing patient records, not real value of the concepts. The mapping dictionary is also needed to be saved in pickle file.
• Demographic information of the patents: List of demographic information of the patients saved in pickle file format. You can create your own demographic information vector. We used sex and age information in the paper (e.g., [1, 0, 0.32] for a male patient whose normalized age is 0.32 and [0, 1, 0.53] for a female patient whose normalized age is 0.53.
• Patient labels: List of binary labels saved in pickle file format indicating whether a patient was progressed into a critical status or not. 1 indicates progression of critical status and 0 indicates moderate patient.
• Pre-trained embedding: Pre-trained embedding of medical concepts appeared in the data. This should be saved in numpy data format (.npy). We used GloVe for pre-trained embedding. You can implement models without preparing pre-trained embedding, but pre-trained embedding can improve models performance as described in the paper.

An example command to implement the RNN model evaluated by 5-fold cross validation AUC is:

    python rnn.py --input_record "input_path" --input_demo "input_path" --input_label "input_path" --output_path "output_path" --max_epoch 20 --batch_size 2 --gru_units 128 --hidden_units 100 --embedding_dim 128 --input_vocabsize 6280 --demo_vocabsize 3 --l2_reg 0.01 --learning_rate 0.01 --dropout_rate 0.3 --k 5

You can change hyperparameter setting (see the descriptions using help command!). The output is the .txt file that contains a list of calculated AUC of each fold. If you want to look up detailed output (e.g., patient vector or predicted risk score) use functions in rnn_experimental.py with your own modifications.

Example commands to implement the baselines evaluated by 5-fold cross validation AUC are:

    python logistic_reg.py --input_record "input_path" --input_demo "input_path" --input_label "input_path" --output_path "output_path" --max_epoch 20 --batch_size 2 --input_vocabsize 6280 --demo_vocabsize 3 --l2_reg 0.01 --learning_rate 0.01 --k 5

    python mlp.py --input_record "input_path" --input_demo "input_path" --input_label "input_path"--output_path "output_path" --max_epoch 20 --batch_size 2 --input_vocabsize 6280 --demo_vocabsize 3 --hidden_units 1000 --l2_reg 0.01 --learning_rate 0.01 --k 5

To implement the baselines (with embedding layer) evaluated by 5-fold cross validation AUC are:

    python logistic_reg_emb.py --input_record "input_path" --input_demo "input_path" --input_label "input_path" --output_path "output_path" --max_epoch 20 --batch_size 2 --input_vocabsize 6280 --demo_vocabsize 3 --embedding_dim 128 --l2_reg 0.01 --learning_rate 0.01 --k 5 --pretrained_embedding "path of pretrained embedding if provided"

    python mlp_emb.py --input_record "input_path" --input_demo "input_path" --input_label "input_path"--output_path "output_path" --max_epoch 20 --batch_size 2 --input_vocabsize 6280 --demo_vocabsize 3 --embedding_dim 128 --hidden_units 1000 --l2_reg 0.01 --learning_rate 0.01 --k 5 --pretrained_embedding "path of pretrained embedding if provided"

You need patient record and the mapping dictionary (i.e. concept2id) to obtain pre-trained embedding using GloVe. See hyperparameters using help.

   python GloVe --help

Example command:

   python GloVe.py --input_record "input_path" --input_concept2id "input_path" --output_path "output_path" --num_epochs 30 --batch_size 51200 --max_vocab 100 --scaling_factor 0.75 --learning_rate 0.01 --dim 128 --use_gpu True

Supplementary material for the paper is available at the archived paper.