A Whole-Cohort Structural Atlas of Multi-Assignment Pathology Reports
Shaurya Bisht
Lay Summary:
A single pathology report can describe several tissue samples, but its billing record usually lists only totals, making it difficult to tell which part of the report belongs to each sample. By analyzing more than 300,000 de-identified reports, this project identified where that structure can be recovered automatically and where human review is still needed.
Abstract:
Background- Surgical pathology reports often describe several specimens within one accession, while linked billing records provide case-level quantities for CPT 88302, 88304, 88305, 88307, and 88309. This mismatch makes specimen-level segmentation a necessary but unmeasured step toward multi-CPT quantity prediction. We characterized recoverable report structure across the complete authorized corpus and identified where billing-derived counts cannot serve as specimen annotations. Methods- We analyzed 301,889 de-identified reports with linked billing records. After excluding 782 cases with disagreement between professional and technical component quantities, 301,107 reports remained. A conservative, deterministic parser detected specimen labels within recognized pathology sections, linked repeated labels across report fields, preserved exact character offsets, and emitted ambiguity flags. We summarized structure across all reports and predefined billing-complexity cohorts. CPT quantities were used only for descriptive comparisons, not as segmentation ground truth. Results- Among 301,107 eligible reports, 201,273 were target-positive, 77,314 had multiple assignments, and 11,751 had multiple target categories. Strict rules detected at least one aligned specimen in 55,235 reports (18.3%), 23,573 multi-assignment reports (30.5%), and 6,695 multiple-category reports (57.0%). Whole-report parsing identified 2,681 additional multi-assignment reports with structure and 3,443 additional reports with at least two specimens compared with diagnosis-only parsing. The detected-structure rate increased from 14.8% in 2018 to 21.8% in 2022, although no 2021 records were available. A CPT-blinded 96-report review sample and a predefined 24-report pathologist subset were prepared; human validation remains pending. Conclusions- Whole-report alignment recovers meaningful specimen structure in complex pathology reports, but case-level billing quantities cannot substitute for specimen-level annotations. This work establishes a reproducible foundation for expert-validated segmentation and subsequent CPT quantity prediction."
Q&A:
Bios: Shaurya Bisht
Program Track: Mentor
GitHub Username:
bshaurya -Shaurya Bisht
What was your favorite seminar? Why?
My favorite seminar was “Senescence, Aging & Cancer, and Bayesian ST Approaches.” I enjoyed how it connected cancer biology with computational methods and showed how spatial data can reveal patterns within tissue that traditional analysis might miss. -Shaurya Bisht
If you were to summarize your summer internship experience in one sentence, what would it be?
I enjoyed turning a broad clinical AI idea into a reproducible study by carefully auditing the data, defining what it could support, and building a foundation for expert-validated pathology report analysis. -Shaurya Bisht