AI-enabled digital test to predict disease recurrence for patients with early-stage invasive breast cancer and performance in a MammaPrint low-risk cohort from the Netherlands with a median 6-year follow-up.

Abstract

1569 Background: Traditional invasive breast cancer (IBC) grading, although useful, remains limited due to diagnostic subjectivity and absence of phenotypic diversity including the recently observed importance of tumor epithelial – stromal interactions and lymphocyte content-distribution. We developed and validated a clinical grade digital test (PreciseDx Breast, PDxBr) which combines image-derived Artificial Intelligent (AI)-grading features and clinical data (i.e. age, stage, tumor size, LN status) to predict recurrence in early-stage IBC and sought to understand performance in a MammaPrint cohort with outcome data. Methods: A MammaPrint cohort with median 6-year follow-up was identified at the Laboratory of Pathology, Dordrecht, the Netherlands (NTH). H&E stained images (digitized at Philips, Eindhoven, NTH) with clinical data (from the pathology and Dutch cancer registries) including demographics, pathology results, MammaPrint risk classification, treatment type and recurrence events were obtained. Performance of the PDxBr validated model (AI-grade + clinical) on the MammaPrint cohort was evaluated using the AUC/concordance index, along with NPV, PPV, Hazards ratio (HR), sensitivity, and specificity. Results: 250 patients, median age 57 years, majority stage 1-IIIa, 100% HR+ve, Her2-ve, 84% LN-ve, 67% grade 2 and 66% MammaPrint low risk. There were 15 events (6%: 7 deaths, 3 second primaries,4 metastases and 1 local regional). PDxBR model classified 134 patients as high risk and 116 as low with 10 of 15 (67%) events identified as high risk. The NPV for PDxBr was 96% with a HR 1.63, Se 70%, Sp 46% vs. MammaPrint with an NPV of 94%, HR of 0.96, Se 40%, Sp 66%; identifying only 5 of the 15 events (33%) as high risk while missing 10. Of note, the AI- grade/imaging model (without clinical features) yielded an NPV of 95% with a HR of 1.46, Se 70%, Sp 42% and correctly identified 10 of 15 events. Conclusions: The results from this observational study suggest that triaging patients with the PDxBr test could potentially be adjunctive to the management decision process of patients with early-stage IBC including the use and subsequent interpretation of genomic tests such as MammaPrint. Additional studies are underway to confirm these initial findings.