Unsupervised Disease Progression

This model contains a reimplementation of the unsupervised disease progression model by Wang et al. link.
The code is heavily based on a previous reimplementation found here.

Installation

The code requires pymc3 (>3.0), please install it using pip install pymc3==3.0. We include testdata from the original implementation.
To check whether the code works for you, please execute python main.py -t 3 -n 5.
This will run the model for only very few data points and iterations.

How to use the code

This code requires a CSV file with three columns with the names PID (patient ID), time_delta, and Code sorted by PID and time_delta.
The patient ID should be an increasing unique integer for each patient.
time_delta describes the time interval in which an observation was made. The original paper describes a window of 90 days.
Code is a task-dependend observation. While the original paper (and the test data) uses ICD9 as observations, the code will run as long as observations are represented as unique strings.

Preprocess

Given the CSV file described above, you can preprocess by using preprocess.py. The execution requires a number of parameters described in the following list:

• The path to the CSV file (not optional)
• Outdir Name of the directory to store preprocessed files in
• Paramdir You can use pre-initialized parameters, stored in .txt files in this folder. Look at data/param_init_small for examples
• anchorsdir The original paper enables the use of anchors for comorbidities. Those are specified in a file named anchor_icd9.csv in this directory.
• maxclaims If set, the preprocessing drops all codes that are not in the top n occurring ones
• minsteps The minimum number of time windows a patients needs to have data for
• seed Random seed
• kcomorbid Number of comorbidities
• mstates Number of states

Running the preprocessing will create a number of .pkl files that are used in further steps.

Running a model

The code to run the model is both in Model.ipynb and in main.py. We recommend the ipython notebook to become familiar with the code.
Individual steps of the model are commented.

Evaluating a model

The resulting parameters of the model are found in the trace, namely in the last and final step of the trace.
Model.ipynb and generate chf patients.ipynb have the necessary code to generate the figures from the original paper.