https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset/
Our understanding of cardiovascular diseases is still relatively primitive, and only a relatively limited amount of prevention and intervention methods are available.
The conventional risk factors associated with cardiovascular diseases includes biographical data(age, gender, height, weight), high systolic blood pressure, high diastolic blood pressure, high cholesterol, high glucose and many other chronic diseases. The chosen dataset(https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset/) contains not only many of these risk factors, but also contains some novel attributes, like physical activity status, smoking status, and alcohol intake from 70000 patiences. It is interesting to explore if these unconventional attributes would have any contribution to predicting one's likelihood of getting a cardiovascular disease.
A data cleaning step and explorative data analysis (EDA) will be performed first to see the structure of the dataset, the levels of the discrete (categorical) variables, and the distribution of the continuous variables. Also, data visualization could help to identify strong relationship between the covariates, if any. After performing an initial explorative data analysis and visualization, the dataset is analyzed with Linear Discriminant Analysis, Logistic regression, and AdaBoost. The fit is then compared and discussed. We can then see if the interesting novel attributes truly have an influence or not.
In summary, it would be an interesting exercise to explore the potential covariates that can predict the likelihood of getting a cardiovascular disease.
This repository is best used via Docker although you may be able to consult the Dockerfile to understand what requirements are appropriate to run the code.
Docker is a tool from software engineering (really, deployment) which is nevertheless of great use to the data scientist. Docker builds an environment (think of it as a light weight virtual computer) which contains all the software needed for the project. This allows any user with Docker (or a compatible system) to run the code without bothering with the often complex task of installing all the required libraries.
Build the docker image by typing in the terminal:
docker build . -t 611
Then start an RStudio server by running:
docker run -v $(pwd):/home/rstudio/work\
-p 8787:8787\
-e PASSWORD="611"\
-it 611
You then visit http://localhost:8787 via a browser on your machine with username rstudio and password 611 to access the machine and development environment.
The best way to understand what this project does is to examine the Makefile.
A Makefile is a textual description of the relationships between artifacts (like data, figures, source files, etc). In particular, it documents for each artifact of interest in the project:
- what is needed to construct that artifact
- how to construct it
But a Makefile is more than documentation. Using the tool make (included in the Docker container), the Makefile allows for the automatic reproduction of an artifact (and all the artifacts which it depends on) by simply issuing the command to make it.
Consider this snippet from the Makefile included in this project:
#clean the dataset
clean_data.csv: .created-dirs data_clean.R \
data/cardio_train.csv
Rscript data_clean.R
The lines with # are comments which just describe the target. Here
we describe an artifact (clean_data.csv), its
dependencies (.created-dirs, data_clean.R,
data/cardio_train.csv) and how to build it Rscript data_explore_clean.R. If we invoke Make like so:
make clean_data.csv
Make will construct this artifact for us. If any dependency doesn't exist for some reason it will also construct that artifact on the way. This greatly simplifies the reproducibility of builds and also documents dependencies.
The report/result of this analysis can be produced by invoking the following through the terminal in RStudio:
cd work
make report.html
And this will build the report and any missing dependencies required on the way.
For our dataset, two of the novel attributes (smoking and alcohol intake) seem to have a significant correlation with the dichotomous cardiovascular disease outcome during the explorative data analysis stage. The variable selection procedure in logistic regression most often picks some of the conventional variables as important predictors: systolic blood pressure, diastolic blood pressure, height, glucose, weight, cholesterol and age. However, for some seeds (seed=359), physical activity, smoking and alcohol intake are also being selected，which indicates these novel attributes may indeed have some importance in the cardiovascular disease outcome. This is an interesting result and worth further looking into with larger and more balanced datasets. Also, it is worth noting that for analysis of this dataset, the three methods (LDA, Logistic regression, and Adaboost) have similar power in prediction.