Case Study: Cutting Breast Cancer Pathology Abstraction Time by 80%
May 7, 2026
The Challenge: Scaling Pathology Abstraction for a Large Epidemiologic Cohort
At a large academic cancer center, a research team running one of the country's most established breast cancer cohort studies faced a familiar problem: their data was only as good as their ability to abstract it.
Research at this scale depends on structured, accurate tumor data abstracted from hundreds of pathology reports across decades of patient follow-up. That abstraction work was entirely manual, including fields like receptor status, tumor size, and histologic grade. A trained registrar averaged 7 to 15 minutes per primary breast pathology report, which is a pace difficult to sustain at cohort scale and nearly impossible to replicate quickly when a study expands, a new cohort is added, or a retrospective analysis requires re-abstraction under updated variable definitions.
The variables involved aren't simple lookups. Breast pathology reports encode receptor status in multiple ways: as numeric percentages, categorical text, intensity scores, and narrative qualifiers. A single patient may have multiple reports that must be reconciled into a coherent patient-level summary. And the information researchers need is often present in one report but absent from another, requiring judgment about what's truly missing versus what simply wasn't documented in a given note.
The team's goal was straightforward: evaluate whether an AI-guided approach could handle this complexity accurately and complete the abstraction faster.
The Approach: Human-in-the-Loop Abstraction with Brim
The research team partnered with Brim to evaluate AI-assisted abstraction across 829 pathology reports from 588 study participants, with approximately 400 reports having gold-standard manual abstractions available for comparison.
The workflow followed Brim's standard human-in-the-loop pipeline:
- PDF to text conversion: reports were converted from source documents into processable text
- Variable definition: the research team described each abstraction target in Brim's variable specification framework
- Prompt tuning: variable instructions were refined based on initial performance
- Machine run: Brim processed the full report set against finalized variable definitions
- AI-assisted review: researchers reviewed Brim's outputs, focusing time on flagged or uncertain cases rather than reading every report from scratch
The variables targeted covered the core tumor characterization data needed for breast cancer epidemiologic research:
- Estrogen receptor (ER) status:: percent positive, category/text, and intensity
- Progesterone receptor (PR) status: percent positive, category/text, and intensity
- HER2 status: IHC, FISH, and HER2/CEP ratio
- Tumor size (numeric)
- Tumor grade
The Results: 80% Reduction in Abstraction Time
The headline result was a reduction in total abstraction time of approximately 80% compared to manual review alone. The bulk of time in the manual workflow was consumed by reading and extracting from individual reports; with Brim, that step was replaced by AI-assisted review of outputs, a fundamentally different and faster activity.
Accuracy on Non-Missing Data
The most meaningful frame for evaluating Brim's extraction quality is accuracy among cases where the relevant information was actually present in the pathology report. In that subset, Brim achieved:
ER and tumor grade hit 97–98% accuracy. They were also the variables with the largest sample sizes and the most standardized documentation conventions. HER2 IHC and FISH were similarly high. PR and HER2/CEP ratio were slightly lower but still strong, with PR reflecting the more variable ways progesterone receptor status is documented across institutions and time periods.
Understanding the Missingness Story
Overall accuracy figures were calculated across all reports including those where a variable simply wasn't documented, and thus were lower, ranging from 55–89% depending on the variable. This is expected and not a reflection of errors: many pathology reports in a longitudinal cohort study genuinely don't contain every variable for every patient. Some reports are outside pathology, addenda, or interim notes that weren't the primary surgical specimen.
Importantly, the research team's manual review confirmed that Brim's non-missing outputs were not hallucinations. In many cases, Brim was correctly pulling values from addenda or cross-references to other reports in a patient's record. This is information a human reader would also use, but that requires knowing where to look. This is a property of Brim's transparent, evidence-linked outputs: every extracted value is traceable back to the source text that informed it, making it straightforward for a researcher to verify or override any individual result.
Why This Matters for Epidemiologic Research
Cancer epidemiology depends on large, well-characterized cohorts, but building them is slow. A study spanning decades of patient follow-up accumulates research value only if the underlying tumor data is structured, accurate, and available for analysis. Manual abstraction has long been the bottleneck.
The 80% time reduction demonstrated here isn't just an efficiency metric. It means:
- Larger effective cohorts: teams can abstract records that would have been deprioritized under manual constraints
- Faster study setup: new research questions can be operationalized without months of re-abstraction
- More consistent data: Brim applies the same variable definitions uniformly across every report, eliminating the inter-rater variability that accumulates across large manual abstraction projects
- Researcher focus on judgment, not retrieval: the human remains in the loop for every abstraction, but spends time reviewing and adjudicating, not reading to find.
This case highlights a pattern we see across healthcare research: critical decisions depend on information buried in clinical notes, the work is repetitive but requires clinical expertise, and manual review does not scale with increasing study volume. By combining automated abstraction with transparent, evidence-linked outputs, Brim enables teams to uphold data quality standards while reallocating time to higher-value scientific work.
Looking Ahead
The research team identified several directions for future work: improving handling of complex multi-report scenarios including multifocality and second primaries, and continuing to refine variable definitions to more closely mirror the full registrar workflow. These are precisely the kinds of iterative improvements that Brim's specification-driven architecture is designed to support. Variable definitions can be updated and re-run without any model retraining.
If your research team is spending hours on pathology report abstraction that could be partially automated, Brim can help.
Interested in learning more? Contact us to see how Brim can support your research workflow.