lpanda2 / term-project

Term Project

This data science project predicts an Acute Kidney Injury (AKI) in patients admitted to hospitals. As training data, we are using the MIMIC-III Critical Care Database which has over 40,000 patients and 58k hospital admissions between 2001 and 2012. The objective is to find a parsimonious model which can help predict the likelihood of AKI in a setting where rapid delivery of information can mean rapid treatment. We will reduce features to the few elements of information that are sufficiently useful to predict likelihood of AKI with a decent degree of accuracy.

File Structure

The code and the data are located in the same directory. The demo data is in a demo-data/ folder on my local and the full data files are located there as well. The following ipython notebooks do a bunch of different things, splitting some data which was too big for memory and feature engineering/exploration. Each jupyter notebook has its own section below.

term-project
    ├── data/   (not committed)
    │   ├── data-file-1...
    │   └── data-file-2...
    ├── demo-data/   (not committed)
    │   ├── demo-data-file-1...
    │   └── demo-data-file-n...
    ├── split-data/   (not committed)
    │   └── chart-events/
    │       └── each-bin-of-chart-data...
    ├── dev_feature_engineering.ipynb (1)
    ├── feature_engineering_utilities.py (2)
    ├── split_charts_data.ipynb (3)
    ├── prod_feature_engineering.ipynb (4)
    └── readme.md

Splitting Data (split_charts_data.ipynb)

The chartevents table was too big for memory. I could have used the AWS Environment (with Athena and the Sagemaker notebooks but wouldn't have been able to preserve my work easily here). I also could have used bigquery and the python pandas API for it, but it felt unnecessary. Instead, I took the largest table I needed to parse and split the data by a md5 hash of the hadm_id.

Features

Feature Investigation List

• chronic liver disease
• sepsis
• anemia
• potassium
• sodium
• medical imaging that uses contrast dyes

Done

• prescriptions which can be nephrotoxins
• mechanical ventilation
• low blood ph
• hypertension
• hematocrit
• gender
• age
• ethnicity
• creatinine increases
• urine color
• urine appearance
• prior admission in 30, 60, 90, 120 days
• prior micu icustay in 30 days
• prior ccu icustay in 30 days

References

Mimic Data Dictionary Mimic Table Schema CMS Risk Adjustment HCC descriptions Medicines that can cause AKI Physician Logic Contrast Dye/Kidney Link