7/17/24 - Episode 2 - Cohort Builder

Eric Snyder: Welcome to our second webinar in our monthly series! Today we're going to present a live demonstration of the Cohort Builder. I'm Eric Snyder, the executive director of the Technology and Innovation Group here at the University of Rochester Medical Center and the Wilmot Cancer Institute. Leading today's demo is our lead engineer and principal developer, Lisa Smith. Lisa has quite a number of apps, including our award-winning CANVAS snapshot app. Within the team, we handle projects in a couple of different ways: some projects go through our typical product management process, and some that we call “Innovation Projects,” come from the developers themselves. These are usually derived from ideas where people are asking for things over and over again. Cohort Builder was what we would consider one of these internal Innovation Projects and the app is already getting quite a bit of use; we expect it to be one of the more utilized apps in the Hyperion Suite. Lastly, as a reminder, we are going to keeping these webinars to around 15 minutes with some questions at the end in order to keep this lunch break length, but without further ado, I'll hand it over to Lisa.

Lisa Smith: Thank you, Eric! Today I'm going to be talking about Cohort Builder. As Eric said, I'm Lisa Smith, and I'll be walking you through this new app, which is our brand-new, self-service data analytics platform. We'll start today with the motivation behind the tool, followed by an overview of its functionality and key features. Next, I'll demonstrate Cohort Builder in action and go over ideas for future enhancements to the product, then I'll introduce you to the development team and provide contact information for the Technology and Innovation Group here at Wilmot, and lastly, we'll have time for questions at the end.

What was the motivation behind building Cohort Builder?

The motivation for cohort Builder came, as Eric said, from the increased need to both understand our patient population and the patterns that I saw emerging. We needed to identify patient subgroups to support the activities of those we serve. For example, while a clinician needs to identify a patient cohort to support opening a new clinical trial, hospital leadership needs to know where best to open a new clinic. Meanwhile, a researcher might be more interested in identifying disease clusters in a certain region. All of these needs come at a cost: non-discrete data takes time to aggregate, decisions often require multiple views of the patient population, and we are a small team working beyond capacity. A self-service solution focused on aggregate counts would allow for greater data access and autonomy of our users while maintaining data security and reducing the burden on our staff. As I said, Cohort Builder is a self-service analytics platform that provides users with immediate access to aggregate data, which allows clinicians, researchers, admins, and others to explore real-time, validated data on an as-needed basis with almost zero barrier to entry. As per its name, Cohort Builder enables users to create patient cohorts based on a multitude of factors, each derived from the details of one aspect of the patient Cancer Center relationship, and to compare those groups seamlessly. The tool also simplifies the process for users to request more information on the cohort they determine and to edit and download the details for future use.

Key Features

Let's go over the key features of this app. The user interface consists of two major sections: the filter area where users select the factors that identify their cohort, and the results where cascading, decreasing counts will indicate how each set of factors has limited their cohort. The factors fall into five major categories:

• The point of contact, including where and when the patient was seen, whether their appointment is completed or simply scheduled, and their insurance information.
• The diagnosis, which includes over 100 specific diseases, when they were diagnosed, and the age of the patient at diagnosis.
• Additional constraints with respect to the stage of the disease and the treatments the patient is receiving.
• Patient demographics, which cover current age, sex, race, ethnicity, vital status, and auditory disability.
• Regional information with respect to the boundaries at the county and zip code levels as well as from a social perspective. Do the patients live in a rural or urban area? What is their national area deprivation index score?

By selecting from as few or as many of these factors, users can build and compare countless cohorts with great ease. After the filter choices are selected, cascading counts are displayed with one click to inform the user again on how their choices progressively limited the cohort size. Once a user has identified their cohort of interest, they can request more information via our ticketing system with one click. From the resulting model, the user can edit, copy, and save the text describing their cohort and simply click open a ticket to request the patients described.

Live Demo

Let's take a walk through a live demo of cohort Builder. So, as you can see here, these are all my filters on the right hand side, and I'm going to just start looking at a retrospective. Let's say that we wanted to look at patients who completed their appointments last year, who were maybe in one of our more vulnerable groups—let’s say ages 65 and over—and from rural communities. This is a live data stream, so when we click retrieve patient counts, this is doing a live data pull and only takes a few seconds. Now, we can see that roughly 48,300 patients were seen last year, and over half of those patients (roughly 25,000) were in our older patient category, with approximately 8,800 of them being from rural zip codes—so roughly 1/6 of those patients seen were from rural zip codes. But we also might want to look at more current data. If we look at patients with completed visits so far this year, we might be interested in looking at the patients who were seen for treatment for CLL. We can see that so far this year we have seen 1,116 patients who were diagnosed with CLL. Maybe we don't want to know just who was seen but who was newly diagnosed this year, so again we can filter this and see a result of 93. If I come back and use this tool tomorrow, these numbers could very well change because there is no preloaded data. This is all live.

Future Enhancements 

Let's talk about the future enhancements coming up for this app. I'm currently in the process of creating admin features to streamline fulfilling requests submitted by our users as well as simplify testing and increase transparency. We hope to integrate Cohort Builder with existing applications, including MARTHA (so users can filter by gene alteration) and CANVAS snapshot (because Cohort Builder could improve our current geospatial analysis). We hope to expand the functionality of the tool itself to include more factors like labs, procedures, or inpatient stays and include a full suite of analytics tools so users can view data graphically or look at trends. As Eric said, the idea for this app came from my experience delivering data, so I was responsible for most aspects of this project, but I collaborated with JC Conrad, who designed an amazing user interface. To learn more about the Technology and Innovation group here at Wilmot, please feel free to visit our website, where you can learn more about our innovations and our team members. Thank you so much for your time, and I'll open it up to questions.

Q&A Session

Q: How modular is this, and is it easy to add in additional data?

A: The structure of the app, from the backend code to the user interface, is completely modular so we can add search factors and expand its capabilities very easily.

Q: Would this be extensible to other groups like primary care or cardiology?

A: Yes, definitely. Its modular nature makes it easily adaptable to other use cases.

Q: What is the update process for this? Does it take a long time to load data in?

A: The app pulls the data in real time, so there's nothing static or pre-loaded, so the load time is what you saw in the demonstration (a couple seconds). It's almost instantaneous, and the data used is continually being updated.

Thank you so much, Lisa, for showing everything and walking through this!