Predicting chemotherapy benefit in premenopausal women with intermediate genomic scores using deep learning.

Abstract

602 Background: The TAILORx trial demonstrated that adjuvant chemotherapy can be safely omitted in postmenopausal women with node-negative HR+/HER2- breast cancer and a 21-gene Recurrence Score (RS) of 11–25. However, chemotherapy benefit could not be excluded for pre/peri-menopausal women with RS 16–25, creating a clinical dilemma that may lead to overtreatment. Using TAILORx data, we have shown that deep learning (DL) applied to hematoxylin and eosin (H&E) can accurately identify women with low or high RS (Shamai et al., Lancet Oncol, 2026). We hypothesized that DL applied to H&E could be used to predict chemotherapy benefit in pre/peri-menopausal women with RS 16–25, thereby facilitating precise treatment de-escalation. Methods: We trained a DL model using H&E slides of both post- and pre/peri-menopausal women from TAILORx, with low (RS<16) and high (RS>25) genomic scores (N=5811), to predict distant recurrence-free interval (DRFI), while excluding patients with RS 16–25 from the training process. The test cohort consisted of women from TAILORx with RS 16–25. Within TAILORx, this test group was originally randomized to either endocrine therapy alone or chemo-endocrine therapy, enabling a direct evaluation of the model’s ability to predict chemotherapy benefit. For comparison, RSCLIN was also computed. Results: When testing on pre/peri-menopausal women with RS 16–25 (N=1234), the model classified 942 (76%) and 292 (24%) as low and high risk. For patients receiving no chemotherapy, the model demonstrated high prognostic performance for DRFI (C-index=0.75; 95% CI: 0.67–0.82), and strong risk stratification (HR=5.43; 95% CI: 2.87–10.29, p<0.001). In the low-risk group, no benefit from chemotherapy was observed (HR=1.07, 95% CI: 0.54–2.15, p=0.84). In the high-risk group, chemotherapy was associated with a significant improvement in DRFI (HR=3.57, 95% CI:1.54–8.30, p=0.003). A significant interaction between treatment and risk group was observed (p=0.024). The model outperformed RSCLIN (p-interaction=0.314). When testing on postmenopausal patients with RS 16–25 (N=2249), the model could not predict chemotherapy benefit (p-interaction=0.926). Conclusions: This digital pathology signature provides prognostic information for distant recurrence and predictive information for chemotherapy benefit in pre/peri-menopausal women with HR+/HER2-, node-negative early breast cancer with RS 16–25. These findings support the potential use of histopathology-based DL to guide chemotherapy de-escalation in this clinically challenging subgroup. The inability of the model to predict chemotherapy benefit in postmenopausal women suggests that the deep learning signature may be identifying tumors sensitive to chemotherapy-induced ovarian suppression, highlighting a subset who could benefit from OFS as an alternative to chemotherapy.