Correlation Study Between Neoadjuvant Chemotherapy Response and Long-Term Prognosis in Breast Cancer Based on Deep Learning Models

TL;DR

This study uses deep learning to improve prediction of breast cancer recurrence after chemotherapy, showing that complete response does not guarantee good long-term outcomes.

Contribution

A novel interpretable deep learning model integrating multiple variables to predict recurrence and metastasis after neoadjuvant chemotherapy.

Findings

The MLP model achieved AUC values of 0.86 for HER2-positive, 0.82 for triple-negative, and 0.76 for HR+/HER2-negative breast cancer cases.

Post-NAC tumor size, Ki-67 index, and Miller–Payne grade were identified as key predictors by SHAP analysis.

Patients achieving pCR still had a 12% risk of recurrence, emphasizing the need for ongoing risk assessment.

Abstract

Background: The pathological response to neoadjuvant chemotherapy (NAC) is an established predictor of long-term outcomes in breast cancer. However, conventional binary assessment based solely on pathological complete response (pCR) fails to capture prognostic heterogeneity across molecular subtypes. This study aimed to develop an interpretable deep learning model that integrates multiple clinical and pathological variables to predict both recurrence and metastasis development following NAC treatment. Methods: We conducted a retrospective analysis of 832 breast cancer patients who received NAC between 2013 and 2022. The analysis incorporated five key variables: tumor size changes, nodal status, Ki-67 index, Miller–Payne grade, and molecular subtype. A Multi-Layer Perceptron (MLP) model was implemented on the PyTorch platform and systematically benchmarked against SVM, Random Forest, and XGBoost models using five-fold cross-validation. Model performance was assessed by calculating the area under the curve (AUC), accuracy, precision, recall, and F1-score, and by analyzing confusion matrices. Results: The MLP model achieved AUC values of 0.86 (95% CI: 0.82–0.93) for HER2-positive cases, 0.82 (95% CI: 0.70–0.92) for triple-negative cases, and 0.76 (95% CI: 0.66–0.82) for HR+/HER2-negative cases. SHAP analysis identified post-NAC tumor size, Ki-67 index, and Miller–Payne grade as the most influential predictors. Notably, patients who achieved pCR still had a 12% risk of developing recurrence, highlighting the necessity for ongoing risk assessment beyond binary response evaluation. Conclusions: The proposed deep learning system provides precise and interpretable risk assessment for NAC patients, facilitating individualized treatment approaches and post-treatment monitoring plans.

Figures