Interpretable and continuous AKI prediction in the intensive care

This is an implementation of an interpretable and continuous Long-Short term Memory (LSTM) network for the task of Acute Kidney Injury (AKI) prediction in ICU settings.

Running

The LSTM model can be run via the LSTM.ipynb; the other baseline models can be run using the LR_XGB_RF.ipynb notebook.

To run the models some preliminary steps are needed to set-up the dataset and extract the data (see Section Data below).

Dependencies

• torch
• numpy
• sklearn
• pandas
• captum
• matplotlib
• seaborn

Data

The MIMIC III dataset was used. The expected data files are:

1. kdigo_stages_measured.csv containing time-series measurements of creatinine, urine output for the last six, 12 and 24 hours and the respective labels.
2. icustay_detail-kdigo_stages_measured.csv containing non-temporal variables of patient demographics such as: age (numerical), gender (binary), ethnicity group (categorical) and type of admission (categorical).
3. labs-kdigo_stages_measured.csv containing time-series data of the laboratory tests.
4. vitals-kdigo_stages_measured.csv containing time-series data of the measurements of vital signs.
5. vents-vasopressor-sedatives-kdigo_stages_measured.csv containing temporal information on whether mechanical ventilation, vasopressor or sedative medications were applied.

To generate such data files some preliminary step are needed:

1. Set-up MIMIC III
2. Run the following SQL scripts from the MIMIC code repository:
   • mimic-iii/concepts/echo-data.sql
   • mimic-iii/concepts/demographics/icustay_detail.sql
   • mimic-iii/concepts/durations/weight-durations.sql
   • mimic-iii/concepts/durations/vasopressor-durations.sql
   • mimic-iii/concepts/durations/ventilation-durations.sql
   • mimic-iii/concepts/fluid-balance/urine-output.sql
   • mimic-iii/concepts/organfailure/kdigo-creatinine.sql
   • mimic-iii/concepts/organfailure/kdigo-stages-48hr.sql
   • mimic-iii/concepts/organfailure/kdigo-stages-7day.sql
   • mimic-iii/concepts/organfailure/kdigo-stages.sql
   • mimic-iii/concepts/organfailure/kdigo-uo.sql
3. Run the SQL scripts in the sql folder. The extract_data.sql should be run after all the other scripts. These scripts builds on and extends the scripts from the MIMIC code repository mentioned at point 2.

Models

Continuous AKI prediction

• LSTM

Prediction 48 hours before the onset of AKI

• Logistic regression
• XGBoost
• Random forest

Cite

If you use our code in your own work please cite our paper: Interpretable and Continuous Prediction of Acute Kidney Injury in the Intensive Care.

Bibtex:

@inproceedings{Vagliano:2021,
     author = {Vagliano, Iacopo and Lvova, Oleksandra and Schut, Martijn C},
     title = {Interpretable and Continuous Prediction of Acute Kidney Injury in the Intensive Care},
     booktitle = {Public Health and Informatics},
     series = {Studies in health technology and informatics},
     pages = {103—107},
     DOI = {10.3233/shti210129},
     volume = {281},
     month = {May},
     year = {2021},
     URL = {https://doi.org/10.3233/SHTI210129},
     ISSN = {0926-9630},
}