Case Study: Discovering a Rare Disease Cohort in Minutes with Brim

The Challenge: Discovery in Unstructured Data

‍

Large research consortia often sit on an incredible asset: millions of pages of rich clinical notes collected across institutions over many years. These notes contain critical details about diagnoses, disease progression, and outcomes, but almost all of that information is unstructured.

For one multi-institutional academic research consortium focused on pediatric disease, this created a familiar problem:

• Clinical notes were abundant, but manual review did not scale
• Existing structured datasets only covered predefined variables
• Exploring new or rare conditions required months of ad-hoc chart review
• There was no practical way to ask: “Do we have patients who might meet this emerging clinical definition?”

In short, discovery itself was the bottleneck.

‍

The Goal: Identify a Rare, Uncurated Condition

‍

Researchers wanted to explore a rare condition that was not already curated in their existing data collection instruments. There were no REDCap fields, no abstraction guidelines, and no pre-labeled cohorts to start from.

The team wanted to answer a simple but powerful question:

“Do we have patients with evidence of this condition?”

But in order to do that, they needed to scan thousands of operative, radiology, and pathology reports across a large patient population.

Doing this manually would have been impractical.

‍

The Brim Approach: Exploratory Discovery at Scale

‍

Using Brim, the team set up an exploratory chart abstraction workflow designed specifically for discovery rather than validation.

‍

Step 1: Load unstructured clinical notes

The project included:

• 2,600+ patients
• Nearly 10,000 clinical notes spanning operative reports, radiology, and pathology

‍

Step 2: Define an exploratory abstraction task

Instead of relying on an existing schema, the team used Brim to:

• Ask high-level clinical questions of the notes
• Extract candidate signals suggesting the rare condition
• Propose values based on patterns observed in the text

This was intentionally not a finalized abstraction; it was designed to surface possibilities.

‍

Step 3: Generate results in minutes

Brim scanned the full dataset of over 2,600 patients and 10,000 notes in about 15 minutes, producing:

• A short list of 16 candidate patients
• Supporting evidence snippets for each candidate

‍

Step 4: Human Review

With Brim narrowing the search space from thousands of patients to a focused set, clinical curators were able to:

• Review candidates efficiently
• Confirm which patients truly met the emerging definition
• Refine and formalize a new curation schema after discovery

This flipped the traditional workflow on its head, so researchers could begin with discovery, use that discovery to refine a schema, and then scale that schema.

‍

The Impact: Making the Previously Impossible Practical

‍

By using Brim for exploratory discovery, the consortium was able to:

• Identify a rare patient cohort that hadn't been curated before
• Reduce weeks or months of manual review to minutes of compute and a few hours of targeted validation
• Create a reusable abstraction framework grounded in real-world data
• Expand the scope of what questions could reasonably be asked of their dataset

Most importantly, Brim enabled researchers to ask new questions of existing data without waiting for new instruments, new funding cycles, or large manual abstraction efforts.

‍

Why This Matters

‍

Exploratory discovery is where many research ideas stall. If finding the right patients is too slow or expensive, promising hypotheses don't get tested.

This case study highlights a different model:

• Use AI to explore broadly
• Use humans to validate precisely
• Build structure after insight emerges

Brim makes that workflow possible, turning unstructured clinical notes into a substrate for discovery, not just downstream analysis.

‍

Interested in Exploratory Discovery with Brim?

‍

If your team is sitting on large volumes of unstructured clinical data and wondering what else might be hiding inside, Brim can help you move from “we think this might be there” to “here’s the cohort”.

Learn more and sign up for a demo at brimanalytics.com.

‍