Predicting 30-day ICU readmissions from the MIMIC-III database

This project was developed as the Capstone project for the Udacity Machine Learning Engineer Nanodegree. Full description and discussion can be found in Report.pdf.

Project Overview

This project describes the development of a model for predicting whether a patient discharged from an intensive care unit (ICU) is likely to be readmitted within 30 days. To do this, I used the MIMIC-III database which contains details records from ~60,000 ICU admissions for ~40,000 patients over a period of 10 years. Using these records, patient records for those readmitted within 30 days of discharge were isolated and used to train a generalized classifier to predict readmission. The project also includes parameter optimization and in depth performance analysis for the classifier.

Data

This project uses the MIMIC-III database, accessible via: https://mimic.physionet.org/

To properly run the project from the included files, a local postgres SQL server must be installed and the MIMIC-III database must be set up as described in https://github.com/MIT-LCP/mimic-code/tree/master/buildmimic.

An SQL materialized view was extracted from the database as defined in all_data.sql.

python libraries used:

• numpy - numerical operation
• pandas - dataframe handling
• os - general operating system operations
• psycopg2 - Used to access a locally installed postgresql server and perform sql queries.
• xgboost - eXtreme Gradient Boosted trees. Classifier implementation.
• scikit-learn - Used for hyperparameter optimization and performance metrics evaluation.
• scipy - interp function used during plotting of the ROC curve.
• matplotlib - visualization
• seaborn - visualization

Main results:

Raw numerical data scatter matrix. A: Scatter matrix of all numerical features. B: zoomed-in view on the first 6 features. Each square shows the scatter plot corresponding to the two features defined by the row and column. The diagonal shows the kernel density estimation plots.

Preprocessed numerical data scatter matrix. A: Scatter matrix of all numerical features. B: zoomed-in view on the first 6 features. Each square shows the scatter plot corresponding to the two features defined by the row and column. The diagonal shows the kernel density estimation plots.

Receiver Operating Characteristic curve for the optimized model, using 5-fold cross validation.

Important features. (A) shows the features sorted by “weight”. (B) shows kde plots for the top 9 most important features, separated by label.