Augmenting Knowledge Discovery in PubMed with R

Michael J. Kane

Phronesis LLC and Yale University

Acknowledgements

Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of The Bill and Melinda Gates Foundation or AT&T Labs Research. 

What's the point of this talk?

>=

WRT text:

Data we

can collect

Data we

can search

>=

Data we

can understand

What's the point of this talk?

>=

WRT text:

Data we

can collect

Data we

can search

Data we

can understand

>=

What's the point of this talk?

Data we

can collect

>=

Data we

can search

>=

Data we

can analyze

Data we can understand

Who cares about your stupid Facebook posts?

Why are they being stored?

Do they have monetary value?

How are they used?

The Cohen Conjecture

The Significance of Search

"The Information Age, also commonly known as the Computer Age or Information Era, is an idea that the current age will be characterized by the ability of individuals to transfer information freely, and to have instant access to knowledge that would have been difficult or impossible to find previously" - Wikipedia

Data + Search

How much data?

Google had indexed about 100 PB as of 2013

"Processing" about 20 PB per day

NSA "touches" 29.21 PB which is 1.6 % of the 1.8 EB transferred over the internet daily.

We can search through PB's of data

What if we don't know what we're looking for?

What if I'm not looking for what everyone else has already found?

The problem with PageRank for research

PageRank works by counting the number and quality of links to a page to determine a rough estimate of how important the website is. The underlying assumption is that more important websites are likely to receive more links from other websites.

Articles in PubMed (http://dan.corlan.net/medline-trend.html)

# Articles in PubMed by year
require(ggplot2)
x <- scan("http://dan.corlan.net/cgi-bin/medline-trend?Q=",
  what=character(), sep="\n", skip=1)
x <- x[-length(x)]
x <- gsub("^[ ]+", "", x)
x <- gsub("[ ]+", " ", x)
pm_pubs <- as.data.frame(matrix(
  as.numeric(unlist(strsplit(x[-1], " "))), ncol=3,    
  byrow=TRUE))
names(pm_pubs) <- c("Number", "Year", "Percent")
qplot(Year, Number, data=pm_pubs, geom="line")

Why PubMed?

Huge repository of knowledge in health science

Programmatic accessibility

Enormous demand for interdisciplinary research between data scientists and clinical investigators

What do you mean by "augmented knowledge discovery" and what does it have to to with PubMed?

Computers don't do research (yet)

Individually, humans don't understand a lot of the available research

Can we use machine learning (statistics) to help human discovery in scientific literature?

Use search to create a research context

Rather than searching for specific terms define the domain of interest and explore the landscape of articles returned

That's a nice narrative but what the hell does it mean and how do you implement it?

Document processing

1. Retrieve documents from PubMed
2. Create term-document matrices
3. Apply a learner to the TDM
4. Visualize the results
5. Iterate
   

Retrieving documents from PubMed programmatically

Code is available at https://github.com/kaneplusplus/akde-pubmed

> source("entrez.r")
> query <- 'brucellosis OR "machine learning"'
> max_ids <- 2000
> doc <- pm_doc_info(query, max_ids)
> names(doc)
[1] "title" "author" "date"
[4] "journal" "publication_type" "url"
[7] "title_and_abstract" 

> plot_article_activity(query, max_ids)
> journal_hist(query, max_ids)

brucellosis OR "machine learning"

brucellosis OR "machine learning"

Term document matrices

Document 1: He walks the dog up the hill                

Document 2: She sees the cat at the top of the hill

A document view

A topic view  (Ken Shirley and Carson Sievert)

User defined topics

1. Define a subset documents of interest within the context
   
2. Mark those documents as "interesting"
3. Use a classifier to find other interesting documents
4. Accuracy improves with the size of the "interesting" document set

Type 1 error (false positives) tends to be small

Type 2 error (false negatives) can be high

What have we learned?

Document clustering lets us understand the spatial relationships among documents

Document classification looks promising

We need better features

We need different perspectives on a context to understand its constituent documents

What else are we missing?

There are a lot of ways to look at a set of documents

We'd like an integrated research and collaboration platform

Future work

Topic evolution

Integration and user experience

Where's the code?

https://github.com/kaneplusplus/akde-pubmed

https://github.com/cpsievert/LDAvis

http://rcharts.io/

http://www.rstudio.com/shiny/