This is a project to reproduce the authors' claims in the paper mentioned above. We will be implementing techniques such as Latent Semantic Analysis, Self-attention, Long Short Term Memory, bi-LSTM, Average pooling, SAPS-II as mentioned in the paper.
pip3 install -r requirements.txt
You must obtain access to the MIMIC-III Clinical Database.
Once data is acquired, follow instructions in mimic-code to load the dataset into a MySQL database (This process may take multiple hours to complete). This database will be used in the data aggregation process.
Open a shell in your MySQL database. Run source ./preprocess/extract_mimic_data.sql to extract relevant data out of the raw MIMIC-III dataset (This may take over 30 minutes). Note: innodb_buffer_pool_size must be increased for your MySQL instance. 2147483648 is sufficient.
Export the KY_MIMIC_EVENTS_V4 and KY_ADM_LENGTH tables as CSVs with column headers to ./raw_data/MIMIC_FULL_BATCH.csv and ./raw_data/MIMIC_ADM_INFO.csv respectively.
Run ./aggregate_data.sh to further aggregate the MIMIC-III data into a format that will be usable by the model.
Run ./run_pipeline_module.sh to train and evaluate the model. Results and status updates will be put in ./logs/exp_name_train.log. Training and evaluation may take over an hour.
Yu, K., Zhang, M., Cui, T., & Hauskrecht, M. (2020). Monitoring ICU Mortality Risk with A Long Short-Term Memory Recurrent Neural Network. Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing, 25, 103–114.