Using Machine Learning Models to Predict Pathologic Complete Response to Neoadjuvant Chemotherapy in Breast Cancer

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Neoadjuvant chemotherapy (NAC) is a standard treatment for breast cancer.
• Predictive modeling aids in forecasting treatment response, optimizing patient management.
• Accurate prediction of pathologic complete response (pCR) to NAC is crucial for treatment stratification.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting pCR in breast cancer patients undergoing NAC.
• To compare different methods for handling missing data in predictive models.
• To identify key clinical features influencing pCR prediction.

Main Methods:

• Utilized data from 499 breast cancer patients treated with NAC across two Singaporean centers.
• Trained and evaluated five ML models using eleven clinical features.
• Assessed listwise deletion and imputation techniques for missing data.
• Model performance was measured using Area Under the Curve (AUC) and Brier score.
• Feature importance was determined using Shapley additive explanations (SHAP).

Main Results:

• Random Forest (RF) models demonstrated high predictive accuracy (AUC ~0.79) and good calibration in external datasets.
• Imputing missing data slightly improved AUC and significantly enhanced Positive Predictive Value (PPV) and Negative Predictive Value (NPV).
• Estrogen receptor intensity, HER2 intensity, and age at diagnosis were identified as the most significant predictors of pCR.

Conclusions:

• ML, especially RF, offers a viable approach for predicting pCR to NAC in breast cancer.
• Data imputation is a valuable strategy to improve the performance and reliability of pCR prediction models.
• Key clinical factors like ER, HER2 status, and age are critical for accurate pCR prediction.