288P - External validation of an AI-enabled digital test (PreciseDx Breast) to predict breast cancer recurrence within 6-years exhibited good independent performance on a comparable early-stage cohort from the Netherlands

Background

We previously validated a clinical grade digital test (PreciseDx Breast, PDxBr) which combines image-derived Artificial Intelligent (AI)-grading features and clinical data to predict breast cancer (BC) recurrence within 6-years for patients with early-stage disease. In validation the test had a c-index (CI) of 0.75 and stratified patients into low vs high risk for recurrence, hazards ratio (HR) of 4.4, NPV of 0.94, PPV of 0.24 sensitivity of 0.60 and specificity 0.77. We sought to understand performance of the model in an independent, comparable cohort from the Netherlands.

Methods

An early stage cohort, median 6-year follow-up was identified at the Laboratory of Pathology, Dordrecht Albert Schweitzer Hospital, the Netherlands (NTH). H&E stained images (digitized at Philips, Eindhoven, NTH) with clinical data (athology and Dutch cancer registries) including demographics, pathology results, treatment type and recurrence events was obtained. Performance of the PDxBr validated model was evaluated using the AUC/concordance index, NPV, PPV, Hazards ratio (HR), sensitivity, and specificity. A MINDACT clinical model (ER/Her2/Grade/Nodal Status/Tumor Size) was utilized as a comparator.

Results

572 patients, median age 60 years, stage 1-IIIa, 100% HR+ve / Her2-ve, 80% LN-ve, and 55% grade 2. There were 89 events (16%: 58 deaths, 4 second primaries, 20 metastases and 7 local regional / nodal recurrences). PDxBR yielded a CI 0.70 (95% CI: 66,73), NPV 0.96, PPV 0.13, HR of 3.17, p-value <0.001, sensitivity 0.78, specificity 0.53. The model correctly classified 73% (65/89) events as high risk with only 27% (24/89) events as low risk (false negatives). The MINDACT model yielded a CI of 0.58 with a HR of 1.92.

Conclusions

The results demonstrate good performance of the PreciseDx Breast test on an independent low-risk early-stage cohort from the Netherlands. NPV of 0.96 (vs prior NPV of 0.94) with a sensitivity of 0.78 supports a potential role for enriching / triaging genomic assays and clinical models while guiding treatment / management decisions. Additional studies are underway to confirm these proposed clinical applications.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

PreciseDx.

Funding

PreciseDx.

Disclosure

M.J. Donovan; G. Fernandez; J. Zeineh; A. Sainath Madduri; A. Feliz; A. Shtabsky; X. zhang; B. Veremis; R. Deangel; J.C. Mejias: Financial Interests, Personal, Full or part-time Employment: PreciseDx. All other authors have declared no conflicts of interest.