Overview

Sources for CS234 Project - Estimation of the Warfarin Dose.

Paper

FILES

• /data - contains provided warfarin.csv dataset and support files
  • appx.pdf: appendix file providing the context
  • metadata.xls: meta data describing the columns of the data set
  • warfarin.csv: original dataset with 5700 patient records
• results - where run results and plots are stored
• clinical_dose.py - subclass of Recommender with implementation of Warfarin Clinical Dosing Algorithm
• config.py - configuration classes for algorithms.
• constant.py - define constants.
• ensemble_majority3.py - subclass of Recommeder with implementation of majority vote ensemble algorithm
• evaluation.py - utilities for evaluating algorithms.
• feature.py - define all enums for features
• fixed_dose.py - subclass of Recommender with implementation of fixed dose algorithm
• lasso_bandit.py - subclass of Recommeder with implementation of Lasso bandit algorithm
• lin_ucb.py - subclass of Recommender with implementation of LinUCB algorithm (disjoint).
• patient.py - encapsulate all info about a patient
• plotting.ipynb - jupyter notebook to generate plots from a result set from a previous run
• preprocess.py - handles all pre-processing of the patient data
• recommender.py - abstract Recommender class to represent a recommendation model
• requirements.txt - for installing required libraries
• tree_heuristic.py - subclass of Recommender with implementation of DTree algorithm.
• util.py - utilities to load and preprocess warfarin dataset
• warfarin.py - the main program to run Warfarin dosage recommendations

USAGE

Run Warfarin Dose Recommender

General command template:

python warfarin.py --algo=[algo_names] --iter=[iterations] --train_ratio=[training set ratio]

• [algo_names]: all for running all models OR one of fixed_dose, clinical_dose, linucb_disjoint. Default is fixed_dose.
• [iterations]: Number of iterations to run the experiments through the entire data set. Each iteration will run on a randomly shuffled permutation of the dataset. Default is 1 for single iteration.
• [training set ratio]: Ratio of the data set used for training. The rest of the data set is used for test set. Default is 0.8 for an 80-20 training/testing split.

Examples:

• Run Fixed Dose recommendation (baseline 1) for 1 (default) iteration with 80/20 (default) training/testing split:

$ python warfarin.py --algo=fixed_dose

• Run Warfarin Clinical Dosing Algorithm recommendation (baseline 2) for 5 iterations with 80/20 (default) training/testing split::

$ python warfarin.py --algo=clinical_dose --iter=5

• Run LinUCB Disjoint recommendation for 1 (default) iteration with 50/50 training/testing split:

$ python warfarin.py --algo=linucb_disjoint --train_ratio=.5

• Run ALL models for 10 iterations with 70/30 training/testing split:

$ python warfarin.py --algo=all --iter=10 --train_ratio=0.7

Generate Plots from Result Set

jupyter notebook

• Launch Jupyter notebook
• open plotting.ipynb and run the cells

RESOURCES

1. Schedule
2. Project Page
3. Default Project Page

REFERENCES

1. A Contextual-Bandit Approach to Personalized News Article Recommendation
2. Contextual Bandits with Linear Payoff Functions
3. Online Decision-Making with High-Dimensional Covariates
4. A Practical Method for Solving Contextual Bandit Problems Using Decision Trees