Introduction

This page describes files which compose the research implementation on supporting medical diagnosis under incomplete data. The approach includes interval modeling of incomplete data, uncertaintification of classical models and aggregation of incomplete results. The evaluation of the approach uses medical data for ovarian tumor diagnosis, where the problem of missing data is commonly encountered.

Technical details

All scripts are written in R 3.1.2. The RStudio project is supervised by packrat software to maintain compatibility of R packages. Documents are generated with use of knitr.

Experiment at a glance

The research consists of 3 steps:

1. making analytic datasets,
2. training and evaluation,
3. visualizing results.

--< datasets/db-2015-04-30.csv
|
|              STEP 1                                  STEP 2                                STEP 3
|
|  ##############################     #########################################     #########################
|  #  make-datasets.Rmd         #     #  training-and-evaluation.Rmd          #     #  results-overview.Rmd #
|  ##############################     #########################################     #########################
|  #                            #     #                                       #     #                       #
----> make-datasets.R           #  ----> training-and-evaluation.R            #  ----> results-overview.R   #
   #   |                        #  |  #   |                                   #  |  #                       #
   #   -> datasets/training.csv >---  #   -> datasets/evaluation-output.RData >---  #                       #
   #   -> datasets/test.csv     >---  #                                       #     #                       #
   #                            #     #                                       #     #                       #
   ##############################     #########################################     #########################
     |                                  |                                             |
     -> make-datasets.html              -> training-and-evaluation.html               -> results-overview.html

Downloading the results

To view outputs of the experiment, run download-data.R script. It will download CSV datasets and binary RData output:

• datasets/training.csv,
• datasets/test.csv,
• datasets/evaluation-output.Rdata.

Reproducing the research

To prepare the software environment to the experiment, open ovarian-tumor-aggregation.Rproj file in RStudio in order to launch packrat and download necessary libraries. The installation process may take from a few to several minutes.

Due to legal restrictions, the initial database datasets/db-2015-04-30.csv can not be published. Therefore, the first step is not reproducible. The remaining steps can be reproduced in two ways (A or B, see sections below). To reflect whole experiment, non-reproducible steps also will be mentioned.

A. Creating datasets and final results

To create only datasets and results, which can be further investigated, execute the following scripts:

1. make-datasets.R (not reproducible),
2. training-and-evaluation.R.

Caution: running training-and-evaluation.R is very time-consuming and extensively absorbs computational resources; it is recommended to run it in environment with 32 x 2.0 GHz cores and at least 200 GB RAM in such setting; the calculation process should take approximately 18 hours.

B. Creating datasets, final results and documents

To create the datasets, the results and additionaly generate the documentation (which explain the implementation of the experiment and the results) launch in knitr following .Rmd files:

1. make-datasets.Rmd (not reproducible),
2. training-and-evaluation.Rmd,
3. results-overview.Rmd.