Three ways of counting lymphocytes, and they do not agree

Design

Secondary biomarker analysis of a randomised phase 3 trial, with manual, digital and AI scoring and multivariable Cox models

Population

4,262 tumour samples from 4,805 patients with early HER2-positive breast cancer in APHINITY; median follow-up 74.1 months

Primary outcome

Invasive disease-free survival by stromal tumour-infiltrating lymphocyte level and scoring method

Effect

Manual interobserver intraclass correlation 0.84 (95% CI 0.79-0.88); AI reclassified 120/1,035 (11.6%) node-positive tumours from immune-low to immune-high; higher TILs associated with better invasive disease-free survival (hazard ratios 0.41-0.93); pertuzumab benefit at higher sTILs 0.36-0.48; largest six-year absolute gain 12.1 percentage points in node-positive disease with manual sTILs of 70% or more

Stromal tumour-infiltrating lymphocytes (sTILs) are established as prognostic in early HER2-positive breast cancer, but what they mean under dual HER2 blockade has been unclear. This secondary analysis rescored 4,262 haematoxylin and eosin slides from the APHINITY trial — 4,805 patients randomised to chemotherapy plus trastuzumab with either pertuzumab or placebo, median follow-up 74.1 months — by manual assessment, an automated digital method, AI-based lymphocyte quantification, and two AI-derived spatial features.

Manual scoring held up: five pathologists scoring 262 samples independently achieved an intraclass correlation of 0.84 (95% CI 0.79-0.88). Agreement between manual and automated methods, however, was only modest, and that disagreement is the finding. AI-based scoring moved 120 of 1,035 node-positive tumours (11.6%) from immune-low to immune-high, and that reclassified group showed greater separation between pertuzumab and placebo at five years than tumours both methods agreed were immune-low.

Higher lymphocyte levels went with better invasive disease-free survival on every measure (hazard ratios 0.41-0.93), and pertuzumab's benefit was larger at higher sTIL levels (0.36-0.48). The largest absolute six-year gain from pertuzumab was in node-positive disease with manual sTIL scores of 70% or above: 12.1 percentage points. AI spatial features, particularly immune hotspots, added information beyond density alone.

None of this is ready to select therapy. What it does is warn that 'immune-low' is not a fixed property of a tumour but an output of whichever method scored it.

• Do not use sTIL scores to withhold pertuzumab — this is prognostic and exploratory work
• Record which method produced a reported sTIL percentage; manual and AI scores are not interchangeable
• Note that manual scoring was reproducible; the disagreement is between methods, not between pathologists
• Node-positive, high-sTIL disease is where the absolute benefit of pertuzumab was largest
• Expect spatial AI metrics to appear in trial stratification before they appear in reports

Why it matters

It shows that an immune classification a clinician might act on depends on which method generated it, in one tumour in nine.

Don't overread it

This is a secondary, exploratory analysis: none of these measures has been shown to select patients for or against pertuzumab.

The statistics, in plain English

An intraclass correlation of 0.84 means pathologists agreed well with each other; it says nothing about whether they agreed with the machine, which they largely did not. The hazard ratio ranges quoted (0.41-0.93 and 0.36-0.48) span different measurement methods rather than describing uncertainty about one estimate. A 12.1 percentage point absolute gain is large, but it comes from a subgroup defined after the fact within a positive trial, and subgroup estimates in that position are systematically optimistic.

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.