PRESCIENTai, an AI-based digital histopathological image signature for risk of late distant recurrence and extended endocrine therapy (EET) benefit in hormone receptor–positive breast cancer.

Abstract

1556 Background: A subset of patients (pts) with hormone receptor-positive (HR+) breast cancer (BC) experiences late distant recurrence (DR) and is more likely to benefit from EET. Clinical practice guidelines recommend use of genomic assays such as Breast Cancer Index (BCI) to identify these pts. We developed an updated AI-based digital histopathological risk score model to predict risk of late DR and extended letrozole therapy (ELT) benefit in this population. Methods: The AI model, PRESCIENTai, was trained on eligible samples (N = 2,271) from the National Surgical Adjuvant Breast and Bowel Project (NSABP) B-42 cohort, which randomized postmenopausal women with HR+ BC who were disease-free after 5 yrs of endocrine therapy (aromatase inhibitor (AI) or tamoxifen followed by AI) to either 5 yrs of letrozole or placebo. A transformer-based end-to-end deep learning model predicted risk score from H&E whole-slide images (WSI) in conjunction with clinical information (age at randomization, surgery type, node status, prior use of tamoxifen, race, lowest bone mineral density T-score, HER2 status). CTransPath was used for feature extraction from WSI tiles. 5-fold cross validation was performed with data split into training, validation, and test sets (60:20:20). The risk score threshold was defined by the 50% quantile of the training set for each fold. Cox regression and Kaplan-Meier analysis evaluated late DR and ELT benefit in high- and low-risk pts. Results: Hazard ratio (HR) was computed for DR in low- vs. high-risk pts [HR = 0.198 (95% CI: 0.124, 0.317); p < 0.001], with absolute difference of 7.61% in 10-yr DR (1.84% vs. 9.46%). High-risk pts experienced greater ELT benefit over placebo (HR = 0.622; 95% CI: 0.416–0.929; p = 0.02) than low-risk pts (HR = 0.727; 95% CI: 0.305–1.733; p = 0.471), with 10-yr absolute benefit of 3.74% vs. 0.66%. Even among node(+) pts, PRESCIENTai identified greater ELT benefit for high-risk pts (HR = 0.521; 95% CI: 0.329–0.827; p = 0.006) than low-risk pts (HR = 0.53; 95% CI: 0.048-5.905, p = 0.606), with 10-yr absolute benefit of 6.66% vs. 1.89%. ELT benefit was also observed for high-risk pts in other clinical subgroups such as age ≤60 years and prior tamoxifen. However, p-interaction for ELT benefit in high- vs. low-risk groups was not significant for all pts ( p = 0.791) or node(+) pts ( p = 0.889). Conclusions: This novel digital signature predicts risk of late DR in pts with HR+ BC. Although absolute ELT benefit was greater in high- vs. low-risk pts, the treatment by risk score interaction was not statistically significant. This is, to our knowledge, the first AI model to predict long-term outcomes in pts with HR+, early BC using a single slide image and clinical information. Successful validation in additional pt cohorts will confirm the clinical utility of PRESCIENTai for prediction of late DR risk and EET benefit.