This page is moving to a new website.

This xkcd cartoon by Scott Munro is open source, so I can hotlink the image directly. But if you go to the source, https://xkcd.com/327/, be sure to hover over the image for a second punch line.

I’m on a Clincal and Translational Science Award (CTSA) research grant (5UL1TR000001-05, formerly 1U54RR031295-01A1), which is pretty cool. My name is even mentioned a few times in the grant. I thought that as I plan what I would do for this grant, I would see what the grant promised and write down what, exactly, that those promises mean. As I talk with various people (especially Russ Waitman, who is supervising my work on this grant), I will revise and update my plans. Still, I thought it would be valuable to put some thoughts down now, both to help me focus on what I should be doing and to offer an early draft of those ideas to the various people that I will end up interacting with.

First, I should note that my name appears in the grant bibliography four times. I’m not first author on any of these, but I’m never first author. Two of the publications are with Janette Berkley-Patton, and related to some work I had done with her on community-based research. That’s going to be an important part of this CTSA grant, but one that I will not be as involved in as I might like. There’s only so much time in the day, and I have made an effort to focus on Informatics research. I’m also on a couple of publications with Byron Gajewski and Jiang Liu that talk about accrual models. I plan to work on these as well, but not for the CTSA grant.

What I will be working on is first described on page 35. Along with Mei Liu, I am developing

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

So let’s parse this statement carefully because it forms the foundation for most everything I talk about here.

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

First, i2b2 is an open source software system that serves as a repository for de-identified data from the Electronic Health Record (EHR) for various medical institutions. The acronym stands for Informatics for Integrating Biology and the Bedside. It is run by a consortium in Boston, and the software is very popular among medical researchers. A nice FAQ at Cincinnati Childrens Hospital gives a non-technical overview of i2b2.

Now i2b2 is not the same thing as Epic or Cerner. Those are the EHR systems used for day-to-day clinical operations. Even if they give me access to those systems, and they wouldn’t in a million years, I don’t want to touch the Cerner or Epic data. These systems are optimized to handle questions like “I’m about to discharge John Doe from the hospital. What do I need to know about his recent gall bladder surgery so I can decide what medications to send him home with?”

The i2b2 system imports data from Cerner or Epic, strips off any personal identifiers as best as it can, and does further actions like date shifting to better improve the confidentiality of the patients. They do such a good job of this, that most Institutional Review Boards (IRB) have designated any research studies involving i2b2 to be exempt from IRB review. You need to be cautious here. While all the obvious personal identifiers have been stripped away, a clever person could, if they wanted to, find a way to identify someone in the i2b2 system. So you have to sign a statement promising that you won’t try to re-identify anyone in this de-identified database.

You lose some information in the process of de-identification, but it is still a very rich data set, especially for longitudinal questions. But you should always keep in mind that when a health care professional enters information into Cerner or Epic, they don’t do it while thinking. “Gee, I better put this information in such a way that Steve Simon will be able to use it for some fascinating data analyses.” No, they’re entering the data to improve patient care. So i2b2 data has the hallmark characteristic of all secondary data analyses. When you use i2b2 data, you are taking information entered for one purpose and re-purposing it for your research needs.

Well, four paragraphs later, and we know understand just one word in that ten word quote. What’s next?

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

What is CDM? For the longest time, I thought that CDM and i2b2 were the same thing. They’re not.

CDM stands for Common Data Model. It is a standard developed by PCORnet (The National Patient-Centered Clinical Research Network) to facilitate sharing data across all the research sponsored by PCORNET. You can look at the descriptions of i2b2 and CDM and compare them yourself. They are similar in spirit, but it seems to me that CDM is more detailed and elaborate than i2b2.

BUT…

(and this is important to remember) the data that feeds into i2b2 and CDM is Epic data or Cerner data or data from some other EHR vendor is in a proprietary format and needs to be shoe-horned into the i2b2 or CDM format. This involves a lot of compromises beyond just the confidentiality issues. One thing I’d like to do, once I understand the EHR better, is to suggest improvements in the shoe horning process. I guess the technical term for this is ETL (Extract, Transform, and Load). This will always be an imperfect process because we are re-purposing, but there is some potential for improvement. That’s not part of my role in the CTSA grant, by the way, but one area I’m very interested in anyway.

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

The adjective “repeatable” is an important one. I want to develop a process that has value beyond the particular problem I am solving. I want a model that others can follow that answer other questions. Also, the group I am working with is very big into reproducible research and insists on things like version control for all their projects.

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

The ability to access the data directly is important to my work. The i2b2 system is very good for addressing focused research questions. It is fast and easy to use. It encourages exploration. But for the application I was most interested in, a data mining application, it is much better to have direct access to the underlying data. for data mining, you need to pull all of the information from the EHR for a well defined subset of patients. That means every diagnosis (ICD) code, every procedure (CPT) code, every lab test (LOINC) ordered, every drug (NDC, RxNorm) administered, and so forth. You also need access to the metadata. I had to learn a bit about SQL to do this, but SQL is a pretty easy language to pick up, at least for the sort of tasks that I was interested in.

“repeatable downstream analysis pipelines that directly access i2b2 and CDM.”

This is the last phrase, but it is the one that gives the best hint at what I am hoping to achieve. The noun pipeline and adjective downstream means that the information is flowing from one location to another. I’m not interested in building something to replace i2b2 or CDM but rather something that sits on top of these systems and pulls information needed for analysis. In my case, it is a particular type of analysis, data mining. I believe that with the right set of software tools, I could make mining of the electronic health record a lot easier and more efficient.

So what, exactly, would this downstream analysis pipeline do? It is way to early to say for sure but I want something that will work not just with the i2b2 database at KUMC, but with any database conforming to the i2b2 or CDM standards. I want something that can move data into R, and once it gets into R from there to other packages like SAS, Stata, and SPSS. I want to optimize the process of pulling data, because the current SQL code that I use relies on some fairly fragile self-joins. I want to utilize the sparse matrix format that is available in R. I want to be able to examine hierarchical relationships within various diagnostic, procedure, and drug codes using information in the metadata of the i2b2 standard.

In order to develop a downstream analysis pipeline, I need to talk to some people who are familiar with i2b2 databases at places like Children’s Mercy Hospital, Saint Luke’s Hospital, and Truman Medical Center. I need to understand the types of research they are interested in, how they interact with i2b2, and what enhancements they want.

There are other things I’m being asked to do for the CTSA grant, and I’ll parse out their meanings in future blog posts. But I’ll put them here just so I don’t forget about them. These are on page 36 of the grant.

• data quality and profiling; sharing quality metrics (where appropriate) and providing capacity to develop added data quality measures for investigators on a fee for service or co-investigator basis;
• analytic methods to determine “fit for use” for trial and studies as data sources increase; and
• machine learning and knowledge discovery application and development.

This last bullet is a task that I share with Dr. Liu.

This is a page outlining several related efforts at RStudio to make it seaier for you to work with data stored in various relational databases.

RStudio. Databases using R. Available at http://db.rstudio.com/.

This is a series of videos and homework exercises that you can work on at your own pace. I have only viewed the outline for this, but anything from DataCamp comes highly recommended.

Nick Carchedi. Intro to SQL for Data Science. Available at https://www.datacamp.com/courses/intro-to-sql-for-data-science.

This is the github repository of Ben Baumer. He is one of the co-authors of “Modern Data Science with R” and the data and code from that book is available here. He also provides code and data for OpenWAR, an open source method for calculating a baseball statistic, Wins Above Replacement. Finally, there is an R library for extracting, transforming, and loading “medium” sized datasets into SQL. Medium here means multi-gigabyte sized files. Related to this are a couple of “medium” sized data sets from the Internet Movie Database and from the NYC CitiBike dataset.

Baumer B. beanumber repository. Available at https://github.com/beanumber.

I have not viewed this video yet, but it comes from a good friend. There is a substantial effort at Microsoft to better integrate the R programming language and their flagship database produce, SQL Server.

Nagaraj Venkatesan. R and SQL Server 2016. Recorded and posted on the SQL Presented December 2016 and posted on the SQL Server Central blog. Available at: http://www.sqlservercentral.com/blogs/sql-and-sql-only/2016/12/22/r-and-sql-server-2016/.

I was running some SQL code within R, and I could not get what I wanted.

> ms <- dbGetQuery(c_connect,
+ "select patient_num start_date from blueherondata.observation_fact
+ where concept_cd LIKE 'CPT:1930%'")

was giving me just a single column for start_date, and the start_date values looked nothing like a date. Was I spelling a variable wrong? Did I get the wrong capitalization?

> ms <- dbGetQuery(c_connect,
+ "select patient_num from blueherondata.observation_fact
+ where concept_cd LIKE 'CPT:1930%'")

worked just fine, and

> ms <- dbGetQuery(c_connect,
+ "select start_date from blueherondata.observation_fact
+ where concept_cd LIKE 'CPT:1930%'")

but when you ask for more than one field, you need to separate those fields with a comma. If you don’t then this version of SQL considers

> ms <- dbGetQuery(c_connect,
+ "select patient_num start_date from blueherondata.observation_fact
+ where concept_cd LIKE 'CPT:1930%'")

identical to

> ms <- dbGetQuery(c_connect,
+ "select patient_num as start_date from blueherondata.observation_fact
+ where concept_cd LIKE 'CPT:1930%'")

So it was extracting the patient_num field and changing its name to start_date. Boy did I feel stupid when I finally realized my mistake. Those pesky commas are important.

I’m hoping that if I write down my mistake and publish it widely, I won’t forget and make the same mistake again next week.

Steve Jones. Where Do You Run Your R Scripts? SQL Server Central Blog. 2016-10-20. Available at http://www.sqlservercentral.com/articles/Editorial/147834/.

Kristian Widjaja, Jun Yang, Jeff Ullman. Oracle Dates and Times. Available at http://infolab.stanford.edu/~ullman/fcdb/oracle/or-time.html.

This page has moved to a new website.

Lukas Eder. 10 Easy Steps to a Complete Understanding of SQL. Created March 17, 2016. Available at https://blog.jooq.org/2016/03/17/10-easy-steps-to-a-complete-understanding-of-sql/