An H&E-derived score refined pancreatic cancer prognosis, but its cut-off did not travel

Design

Multicohort development and validation of a deep-learning prognostic score

Population

644 training cases plus six retrospective cohorts and a prospective external cohort of 155 resected PDAC

Primary outcome

Overall survival

Effect

C-index gain 0.015–0.047; prospective cohort mortality 47.5% vs 22.7%, adjusted HR 4.45 (2.28–8.70)

Researchers built TIGERS, a nine-feature score from routine H&E whole-slide images of resected pancreatic ductal adenocarcinoma, measuring tumour burden, gland and nuclear shape, chromatin texture and tumour-stroma arrangement. It was trained on 644 treatment-naive cases and tested in six retrospective cohorts and one prospectively collected external cohort of 155.

Adding the score to a standard clinicopathological model improved prognostic discrimination in every retrospective cohort, though modestly. In the prospective cohort, mortality was 47.5% in the high-risk group against 22.7% in the low-risk group, and high risk remained linked to worse survival after adjustment. The score moved 43% of patients into a different risk group from stage alone.

In one external cohort, the fixed cut-off labelled 96% of patients high risk, showing that thresholds set in one laboratory may not hold in another with different staining or scanning.

This is prognostic, not predictive: it does not yet tell anyone what to do differently, and the authors call for interventional studies first.

• Treat H&E-based AI risk scores as research tools until prospective interventional data exist.
• Expect thresholds to need local recalibration when staining or scanners differ.
• Keep TNM stage, margin and nodal status as the reported prognostic basis.
• Validate any AI tool on local cases before using it in reports.

Why it matters

It shows routine slides carry prognostic signal beyond stage, and also how fragile fixed AI thresholds are across laboratories.

Don't overread it

Prognostic only; it has not been shown to change treatment decisions or outcomes.

The statistics, in plain English

A C-index rises from 0.5 (chance) to 1.0 (perfect); gains of 0.015 to 0.047 are real but small. A hazard ratio of 4.45 in the prospective cohort comes from 155 patients, so its interval (2.28 to 8.70) is wide.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.