Overview

In this assignment you will explore deep learning applied to electronic health records using the MIMIC-III dataset.

In part 1, you'll use structured clinical measurements to predict daily sepsis, myocardial infarction (MI), and vancomycin antibiotic administration over two week patient ICU courses.

In parts 2 and 3, you will work with clincial text embeddings to predict ICU patient readmissions.

Assignment Setup Instructions

1. Make sure you use a server with at least 8 CPUs, 30GB memory, and 100 GB disk space.
2. Like with the previous assignment, download the assignment files on your VM instance and then unzip them:

   
                       wget https://biods220.stanford.edu/notebooks/assign2.zip
                       unzip assign2.zip
                     

Note 1 Each notebook has downloading and preprocessing code in the first few cells. This will take about 60mins for part 1, 10mins for part 2, and 60mins for part 3, so plan accordingly. We recommend running all the preprocessings first because it only requires one pass.

Note 2 When you finish one assignment part and start another, remember to restart the finished notebook by clicking the button at the top to release all the resources, including CPU, GPU, memory.

Note 3 The files in this assignment need less than 100GB of disk space. If you run out of disk space, you can delete the datasets from assignment1, or create a new VM. You can always check disk usage by running this in the terminal: du -sh *

What to hand in for this assignment: Submit one PDF file containing all your notebook solutions (code + written). To convert your notebook to PDF, go to File > Download as > PDF via LaTeX. Alternatively, if you encounter issues with converting to PDF using LaTeX, you can also convert to HTML (File > Download as > HTML) and then "Save to PDF" by printing. Concatenate the PDFs from both parts into one and make sure that both the coding sections and the written sections are visible in the PDFs prior to submission.