Interpretable Machine Learning for Predicting Neoadjuvant Chemotherapy Response in Breast Cancer Using the Baseline Clinical and Pathological Characteristics

Background

The pathological response to neoadjuvant chemotherapy (NAC) has become a vital prognostic indicator for patients with breast cancer (BC). The newly generated models depended on rather basic imaging and pathology characteristics and did not sufficiently elucidate the importance of the incorporated data. The purpose of this study is to establish and authenticate a machine learning model for predicting the pathological complete response to NAC using baseline clinical and pathological features in BC patients.

Methods

Data were collected from hospitalized BC patients treated with NAC at Zhejiang Provincial People's Hospital between January 2014 and August 2023. The dataset was randomly split, with 70% allocated for model training and 30% for validation. LASSO regression was used to select predictive features. Six ML models—XGBoost, LightGBM, CatBoost, logistic regression, random forest (RF), and support vector machine (SVM)—were developed, with performance assessed using the area under the curve (AUC) and accuracy, precision, recall, F1 score, and Brier score. Clinical benefits were evaluated using decision curve analysis (DCA), and SHapley Additive exPlanation (SHAP) was applied to interpret the features of the optimal ML model.

Results

A total of 303 bc patients treated with NAC were included, with a pCR rate of 29.37% (89/303). Twelve features, such as age, menopausal status, PR, HER2 status, Ki-67 expression, stromal tumor-infiltrating lymphocytes (sTILs) et al., were selected for model construction. Among the six models, the CatBoost model demonstrated the best predictive performance, achieving an AUC of 0.853 after Bayesian hyperparameter tuning. SHAP analysis ranked sTILs as the most critical predictive feature. In fivefold cross-validation, the CatBoost model incorporating sTILs achieved an average AUC of 0.83.

Conclusions

The ML-based pCR prediction model enables more accurate pCR prediction for BC patients at baseline, aiding in optimizing treatment strategies. Additionally, the interpretable SHAP framework enhances model transparency, fostering clinical trust, and understanding among doctors.