Auto-Narratives

I had recently learned about patient safety narratives, a key deliverable in clinical study reporting that is prepared on a per-patient basis. At the same time, I was also looking to extend my Rmarkdown knowledge to create automated workflows or “pipelines”. You can see where this is going right?

My aim was pretty clear: I wanted to create an Rmarkdown-based reporting engine, capable of automatically creating safety narratives for all participants in a clinical trial based on a single template.

For this example, I borrowed the clinical trial data from the PHUSE Github Repository, openly available here

Next, I got to work in seeing what data elements I would want to pick out for my narratives. I was less concerned with what’s actually in a narrative and more interested in the process of automating their creation. Nevertheless, I picked some elements that I think are relevant such as:

• Patient Age/Sex/Race
• Dates of first and last treatment
• Adverse Events, including reported term, severity, outcome and associated dates

The technical part of this project was interesting to think about and research. I ultimately decided with the following approach:

1. Create a template or skeleton structure of a narrative that will be applied to each patients data
2. The template includes data wrangling steps that will be repeated for each subject. In here, I also include a table object just to test out variety.
3. Using Rmarkdown’s parameters (here each subject’s unique identifier), loop through the unique identifier list and have the corresponding report created.

This was great and I was able to get it to work almost immediately.

However…

I wanted extend the current approach by providing a user interface to the engine. This is especially useful for non-programming colleagues (or colleagues not familiar with R or Rmarkdown). Enter Shiny.

Briefly, the user experience can be described as:

1. A list of patient identifiers are populated in a DataTable, and the user is instructed to select all or some of the patients they wish to generate a report for.
2. The user then has the option to select the format of their narratives: .docx, .html, or .pdf
3. Once selections are made, from here the process works similar to the above, but within the Shiny server context.
4. The last step involves collecting all generated reports, bundling them into a single zip file, and returning them to the user as a download. While this process happens in real time, a progress bar is provided to the user as feedback.

That last part still amazes me.

I’ve since taken this a few steps further, investigating even more approaches (e.g. parent-child reports) and conditional execution of narrative text based on values of users data and including graphics. I will likely update this repo to include those enhancements, but for now enjoy as is!