A prognostic prediction model for neoadjuvant chemotherapy in locally advanced breast cancer integrating deep learning and pathomics.

Abstract

e12564 Background: Breast cancer is the most prevalent malignant tumor among women worldwide, constituting a significant threat to women’s health. Neoadjuvant chemotherapy(NAC) has emerged as one of the standard treatment approaches for locally advanced breast cancer patients. However, individual responses to NAC vary considerably, leading to divergent treatment outcomes and complicating the timely adjustment of therapeutic strategies.Therefore, predicting the likelihood of achieving pathological complete response (pCR) and the risk of recurrence and metastasis based on pre-NAC biopsy findings is a critical clinical challenge for personalizing treatment plans. Methods: This study included 261 breast cancer patients who completed NAC between 2015 and 2022 in Cancer Hospital of Shantou University Medical College ,who were divided into a training group (n = 180) and a validation cohort (n = 81). Clinical and pathological data, including stage, and surgical details,were extracted from electronic medical records .The association between pCR and survival outcomes—disease-free survival (DFS), distant recurrence-free survival (DRFS), and overall survival (OS)—was analyzed using mixed-effects Cox proportional hazards models, with survival curves estimated by the Kaplan-Meier method. The ResNet50 deep learning architecture was employed to extract and learn histopathological features from whole-slide images (WSIs) of both pre-NAC biopsy tissues and post-NAC surgical specimens. Finally, an integrated model was developed to predict prognosis for patients with locally advanced breast cancer following NAC. Model performance was evaluated using the F1-score, time-dependent receiver operating characteristic (ROC) curves, and the concordance index (C-index). Results: Over a median follow-up of 5.6 years (IQR, 3.2-11.1），the 5-year DFS was 65.90%，and the 5-year OS rate was 81.61%.An integrated model combining clinical data with features from pre-NAC biopsy WSI achieved an area under the curve (AUC) of 0.72 for predicting recurrence and metastasis risk.To enhance predictive accuracy, features from post-NAC surgical specimen WSIs were incorporated, resulting in an improved integrated model with an AUC of 0.79. The F1-score, time-dependent ROC analysis, and C-index consistently supported the model's effectiveness in predicting both pCR and the risk of recurrence and metastasis. Conclusions: We developed an integrated model that synthesizes histopathological features from pre-NAC biopsy and post-NAC resectionspecimens with various key clinical parameters. This model holds promise for assisting clinicians in in accurately predicting patient prognosis prior to initiating neoadjuvant chemotherapy, thereby facilitating earlier and more timely formulation of personalized treatment strategies.