Design: independent external validation study pooling seven randomised adjuvant trials, with masked deployment of locked AI models (CATALINA)
Population: 1,356 patients with early triple-negative breast cancer with complete clinicopathological, pathologist and computational TIL data
Primary outcome: invasive disease-free survival, distant disease-free survival and overall survival; 5-year time-dependent AUC
Effect: pathologist stromal TILs HR 0.73 (95% CI 0.66-0.82) and computational TILs HR 0.80 (0.73-0.89) for invasive disease-free survival; computational score lost significance when added to a model containing the pathologist score

CATALINA pooled long-term outcome data from seven randomised adjuvant trials and deployed two previously validated AI pipelines, locked and without retraining, to score tumour-infiltrating lymphocytes on slides from 1,356 patients with early triple-negative breast cancer.

The computational scores were independently prognostic: hazard ratios of 0.80 for invasive disease-free survival, 0.77 for distant disease-free survival and 0.79 for overall survival after adjustment for clinicopathological factors. They also improved five-year discrimination over clinicopathological variables alone.

The negative finding is the more useful one. Once pathologist-scored stromal TILs were in the model, the AI score no longer held a significant prognostic association and did not further improve the area under the curve. Correlation between the two was only modest, at 0.375 to 0.473 — they are measuring overlapping but not identical things, and the pathologist's version carries the information.

The authors' conclusion is appropriately narrow and, for Indian practice, the relevant one: the case for computational TIL scoring is strongest where routine pathologist assessment is not available. In a centre with a breast pathologist who scores TILs, this adds nothing; in a network where slides are read by generalists under volume pressure, a reproducible automated score may be worth more than the validation statistics suggest.

Do not replace pathologist TIL scoring with an automated score where the pathologist assessment exists
Consider computational scoring where TILs are not being scored at all, which is the realistic alternative in many centres
Report TILs as stromal TILs using the standard method, so results remain comparable across centres
Treat TILs as prognostic, not predictive of benefit from a specific therapy
Note that the AI models were deployed locked and unmodified — local retraining would invalidate this validation

Don't overread it

Prognostic is not predictive — nothing here shows that a TIL score, by either method, should change which treatment a patient receives.

The statistics, in plain English

The hazard ratios below 1.0 mean higher TIL scores predicted better outcomes, and all were statistically robust with q-values under 0.0001 — the q-value is a p-value adjusted for testing many outcomes at once. The important result is the one with no number attached: after adjusting for clinicopathological variables and the pathologist score, the AI score's association was no longer significant. That is what 'adds no incremental information' means statistically. The modest correlation of 0.375 to 0.473 between the two scoring methods means they agree far less than their similar prognostic performance implies.