Artificial intelligence may soon change how oncologists decide which metastatic breast cancer patients should receive expensive immunotherapy drugs. A retrospective study published in Breast Cancer Research found that AI-based digital pathology assessment of PD-L1 expression identified patients with significantly longer progression-free survival on pembrolizumab, while traditional pathologist evaluation failed to reach statistical significance.

The study analyzed 49 patients with metastatic triple-negative breast cancer treated with pembrolizumab between 2021 and 2024. All patients had already met the standard clinical threshold for treatment—a PD-L1 combined positive score of at least 10—yet outcomes varied considerably. The question was whether AI could identify which patients within this approved group would actually benefit.

AI Detects a Signal Human Assessment Missed

Using Lunit's SCOPE platforms to analyze both standard H&E slides and PD-L1 immunohistochemistry, the AI identified 51.5% of samples as PD-L1 positive, compared to 74.2% flagged by pathologists. This more stringent AI classification proved clinically meaningful: patients classified as AI-positive had a median progression-free survival of 8.8 months versus 6.7 months for AI-negative patients (hazard ratio 0.50, p=0.031).

By contrast, pathologist assessment showed a trend toward longer survival in PD-L1-positive patients (7.9 versus 7.2 months) but did not reach statistical significance (HR 0.63, p=0.072).

The predictive power was strongest in pre-treatment samples, where AI-positive patients survived 7.2 months without progression compared to just 4.4 months for AI-negative patients (HR 0.29, p=0.006). Post-treatment samples showed no significant association, and the researchers noted that survival benefits emerged primarily after six months of treatment.

Macrophages and Treatment Timing

The study revealed an intriguing biological pattern: PD-L1-positive macrophages comprised 34% of positive immune cells in post-treatment samples versus only 16% in pre-treatment samples (p=0.0044). This shift in immune cell composition may explain why post-treatment biopsies were less predictive of pembrolizumab response.

The 49-patient cohort included 23 pre-treatment specimens and 24 post-treatment specimens, with two patients contributing both types (though not as serially paired samples). Median follow-up was 13.2 months.

Implications for Clinical Practice

The findings suggest that AI-based PD-L1 quantification could refine patient selection for pembrolizumab beyond current clinical cutoffs, potentially identifying responders more accurately than human pathologists. For Lunit, the study provides clinical validation evidence for its SCOPE platform in the competitive digital pathology AI market.

For health systems, automated digital pathology platforms could standardize PD-L1 assessment, reduce inter-observer variability, and potentially improve cost-effectiveness of immunotherapy—pembrolizumab costs approximately $150,000-200,000 annually. More precise biomarker assessment could spare non-responders from ineffective treatment toxicity while ensuring true responders receive therapy.

However, the retrospective design and modest sample size require cautious interpretation. Prospective validation will be necessary before AI-based assessment can be adopted into clinical practice for immunotherapy patient selection.

Sources

  1. Artificial intelligence-based tumor microenvironment and PD-L1 analysis to predict pembrolizumab response in metastatic triple-negative breast cancer — Breast Cancer Research
  2. AI pathology scoring flags longer survival on pembrolizumab in metastatic triple-negative breast cancer — Scienmag