Design: Multicentre retrospective cohort with machine-learning model and external validation
Population: 754 patients, 11,678 lymph node whole-slide images
Primary outcome: Overall survival discrimination
Effect: Interobserver κ 0.879; C-index 0.744 temporal and 0.748 geographic validation

This multicentre study took whole-slide images from 754 patients having curative colorectal cancer surgery across three cohorts, covering 10,410 non-metastatic and 1,268 metastatic nodes. Non-metastatic nodes were classified on H&E as stimulated, unstimulated or depleted according to proliferative and depletion-related features.

Agreement between observers was high (Fleiss' κ 0.879, 95% CI 0.810 to 0.938). Stimulated nodes showed immune activation, and depleted nodes showed stromal remodelling and weaker adaptive immune signalling, matching previously described molecular node subtypes. A random survival forest built on each patient's mix of node subtypes reached C-indices of 0.744 and 0.748 in temporal and geographic validation, adding prognostic information beyond conventional staging.

This is research-stage. The classification is reproducible and uses a stain every laboratory already has, which makes it attractive, but it has not entered any reporting dataset and would add time to node assessment. It is worth knowing about as a possible future addition to staging rather than something to start reporting now.

Continue to report node counts and metastatic involvement according to current datasets.
Treat node subtype classification as research-stage; it is not part of any reporting standard.
Note that the scheme uses routine H&E and showed high interobserver agreement.
Watch for prospective validation before adding node morphology to reports.

Why it matters

Nodes reported only as negative may carry information about the host immune response that staging ignores.

The statistics, in plain English

A C-index of about 0.75 in external cohorts means the model ranked patients' survival correctly about three times in four. The training C-index of 0.877 was higher, as expected, and the drop on validation is the realistic estimate.