Chemo-praidict eBC: A deep learning solution for predicting pathological complete response on biopsies of breast cancer patients treated with neoadjuvant chemotherapy.

Abstract

e15156 Background: In precision medicine, the prediction of tumor chemosensitivity is of major importance to offer cancer patients the best possible treatment from the outset. In this study, we introduce Chemo-praidict eBC, a deep learning model designed to predict the occurrence of pathological complete response (pCR) in early breast cancer (eBC) patients treated with standard neoadjuvant chemotherapy (NAC). This prediction is based on an analysis of the initial tumor diagnostic biopsy. Methods: We used two extensive cohorts (total n = 1140 patients) spanning various molecular subtypes of eBC (HER2-amplified (HER2+), estrogen-receptor positive/HER2 non amplified (ER+/HER2-), and triple-negative (TN) tumors): the PRIMUNEO prospective cohort (n = 500) for training and internal validation and the CGFL Breast Cancer Neoadjuvant database (n = 640) for external validation. Results: Chemo-praidict eBC demonstrated good performance on the external validation dataset for HER2+ tumors (Area Under the Curve (AUC): 0.652 (P = 0.001), Odds Ratio (OR): 2.42 (P = 0.0131)), ER+/HER2- tumors (AUC: 0.814 (P = 0.003), OR: 20.56 (P = 0.00413)) and TN tumors (AUC: 0.677 (P = 0.001), OR: 3.44 (P = 0.00373)) compared to standard clinicopathological features. We also evaluated the robustness of our algorithm through testing on several scanned sections per patient. Chemo-praidict eBC exhibited strong consistency in the external validation cohort, with a Pearson correlation coefficient of 0.933 (P < 0.001) for HER2+, 0.932 (P < 0.001) for ER+/HER2- tumors, and 0.939 (P < 0.001) for TN. Conclusions: Chemo-praidict eBC is a new tool for identifying eBC that are differentially sensitive to standard NAC, and could help to select the most appropriate treatment strategy in HER2+, ER+/HER2- and TN eBC.