Truace
Subscribe

Improving Postpolypectomy Surveillance Decisions Using an Automated Analysis of Colonoscopy and Histopathology Reports: 2-Center Validation Study

Background: Postpolypectomy surveillance intervals derived from colonoscopy findings and histopathology should be guideline-based, yet these are frequently not adhered to, contributing to under- and oversurveillance. This motivates a more structured and automated approach to improve adherence. Objective: We developed and evaluated an automated, implementable, guideline-aligned decision support tool that extracts colorectal polyp features from free-text colonoscopy and histopathology reports and produces auditabl…

By the Truvace desk1 min read1 sourcePeer-reviewedTRV-2026-1640· cite & verify
Improving Postpolypectomy Surveillance Decisions Using an Automated Analysis of Colonoscopy and Histopathology Reports: 2-Center Validation Study
"Endoscopy" by Yuya Tamai is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

Researchers built a tool that reads free-text colonoscopy and histopathology reports from NHS trusts, pulls out structured polyp features using open-weight language models, and then applies encoded guideline rules to recommend follow-up intervals after polyp removal.

Guideline-based surveillance intervals are often not followed, leading to both too little and too much follow-up. A tool that is auditable and accurate could support safer scheduling, but the results come from retrospective samples of 100 cases per site, so real-world clinical impact is not yet shown.

Main points

  1. Extraction accuracy was 0.995 for specimen type, 0.981 for location and diagnosis, and 0.949 for excision completeness.
  2. Recommendation accuracy was 0.96 at Guy's and St Thomas' and 0.93 at Sandwell and West Birmingham.
  3. Cancer alert F1-score was 1.00 at both sites, while 3-year follow-up F1-scores were lower at 0.86 and 0.80.
  4. A deterministic rule engine, not the language model, computes the surveillance interval, making outputs auditable.

The gain

A hybrid pipeline of open-weight language models and a deterministic guideline rule engine reached 0.93 to 0.96 recommendation accuracy across two NHS trusts and would mainly reduce potential undersurveillance.

The rundown

The pipeline uses four prompting stages to split multispecimen histopathology reports into individual pots, align endoscopic findings with each specimen, and normalize each pot into a structured schema.

Extraction was tested on 100 procedures comprising 215 pots. Recommendation performance was tested on a same-site cohort of 100 patients and externally validated on 100 patients from a second trust.

A separate retrospective audit of 127 patients compared automated recommendations and observed routine scheduling with a clinician-derived reference based on full electronic health record review. The tool would improve adherence mainly by reducing potential undersurveillance.

Sources

  1. Peer-reviewedJournal of Medical Internet Research2026-10-16

The debate