SHAP analysis of predictive factors for pathologic complete response in breast cancer patients undergoing neoadjuvant chemotherapy.

Abstract

e12527 Background: Pathologic complete response (pCR) to neoadjuvant chemotherapy (NAC) is a key predictor of long-term outcomes in breast cancer. This study employs SHAP (SHapley Additive exPlanations) to interpret a Random Forest model predicting pCR based on clinical and treatment factors, providing insights for personalized treatment. Methods: A dataset of 158 patients was split into 80% for training and 20% for testing. SHAP analysis evaluated feature impacts on pCR prediction (pCR=1). Key features were analyzed, and unexpected findings explored for future investigation. A final predictive model will be presented in subsequent publications. Results: Key features, ranked by importance, included Ki67, Trastuzumab, Total RDI (Relative Dose Intensity), Subtype, Age, cT (tumor size), and cN (lymph node involvement). • Ki67: Higher values strongly correlated with pCR, reflecting its role in tumor proliferation. • Trastuzumab: Showed strong positive contributions, particularly in HER2-positive subtypes. • Total RDI: Lower values unexpectedly correlated with pCR, challenging the paradigm that higher RDI improves outcomes. This highlights the need to explore confounding factors such as dose reductions, patient tolerance, or regimen effects. • Age: Higher values correlated positively with pCR, possibly reflecting favorable tumor biology or effective treatments in older patients. • Subtypes Luminal B and HER2-Enriched: Displayed distinct and sometimes counterintuitive contributions, emphasizing tumor biology complexity. • Tumor size (cT): Larger tumors correlated positively with pCR, likely due to subtype-specific responses. Conclusions: SHAP analysis revealed key factors influencing pCR in NAC-treated breast cancer patients. The inverse relationship between TotalRDI and pCR warrants further investigation, challenging reliance on dose intensity as a predictor. These findings demonstrate machine learning’s potential to uncover nuanced relationships and improve treatment personalization. Key predictive features and their contribution to Pathologic Complete Response (pCR) prediction in breast cancer patients. Predictive Feature Average Contribution Impact on Prediction 1 Ki67 0.094592 Higher values increase pCR 2 Trastuzumab 0.091794 (same) 3 TotalRDI 0.061163 (same) 4 Subtype_Luminal B 0.051591 (same) 5 Subtype_HER2_Enriched 0.037895 (same) 6 Age 0.032980 (same) 7 Subtype_Luminal A 0.016545 (same) 8 cT 0.013637 (same) 9 cN 0.011557 (same) 10 Regimen_TC4 0.010680 Higher values increase pCR SHAP (SHapley Additive exPlanations) analysis was used to evaluate the impact of individual features on pCR prediction. The “Average Contribution” reflects the mean SHAP value for each feature, and “Impact on Prediction” indicates whether higher feature values were associated with an increased or decreased likelihood of achieving pCR.