Real-world deployment of a multimodal artificial intelligence (AI) model for predicting response to neoadjuvant chemotherapy in early breast cancer.

Abstract

534 Background: Multimodal AI models integrate morphological information from H&E-stained images with clinical information, identifying subtle disease signatures beyond the perceptual limits of human experts. However, the clinical utility of AI tools in real-world settings remains underexplored. Here, we evaluated Ataraxis Breast (ATX), a multimodal AI test predicting response to neoadjuvant chemotherapy, in a prospective silent trial. Methods: We previously assembled an international dataset of whole slide images of H&E-stained biopsy specimens and clinical information to develop ATX as a biomarker of neoadjuvant response. To evaluate real-world clinical translation, we conducted a prospective silent trial of ATX in patients with early breast cancer treated with neoadjuvant therapy at our center. None of these patients were included in the original training, validation, or test datasets. The primary outcome was pathologic complete response (pCR). Area under the receiver operating curve (AUROC) analyses were conducted. Results: The study cohort included 50 female patients treated during 2025, with a median age of 53.8 years. Tumor subtypes comprised 30% (n = 15) triple-negative breast cancer (TNBC), 24% (n = 12) HER2-positive disease, and 46% (n = 23) hormone receptor–positive (HR+), HER2-negative tumors. Overall, 66% (n = 33) of patients had node-positive disease, and 60% (n = 30) had grade 3 tumors. Notably, all patients with TNBC received pembrolizumab, whereas patients with HER2-positive tumors were treated with chemotherapy combined with HER2-directed therapies. The overall pathological complete response (pCR) rate in this external validation cohort was 34%. ATX scores were significantly higher among patients who achieved pCR (p = 0.005). Moreover, when evaluated as a predictive biomarker across the entire cohort, ATX achieved an AUROC of 0.745 (95% CI, 0.58–0.89) for pCR. These findings were consistent across predefined subgroup analyses, including patients with HR+/HER2− disease (AUROC = 0.73) and TNBC (AUROC = 0.77). Conclusions: In this prospective deployment study, among patients with early breast cancer treated with neoadjuvant chemotherapy, higher AI-derived predictions were independently associated with improved neoadjuvant therapy response. Given the accessibility of H&E-slides, ATX may serve as a scalable predictive biomarker in real-world settings.