Breast cancer survival prediction using machine learning and multimodal data for personalized care plan

This paper presents a time-stratified breast cancer survival analysis that incorporates tumor characteristics, disease stage, and patient features, using machine learning (ML) algorithms to support personalized care planning. Considering nonlinear relationships between input data and the risk of death, as well as the dynamics of patient characteristics over time, we utilize 6 ML algorithms, including XGBoost, Logistic Regression, Random Forest, Decision Tree, AdaBoost, and Multilayer Perceptron, for less than 6 months, 6 months, 1-, 2-, 3-, 5-, and 10-year survival rates prediction. We utilize multimodal data from 3,476 breast cancer patients, comprising 43 features. XGBoost outperforms the rest, achieving an accuracy of over 90%. The top five predictors, using Shapley Additive exPlanations (SHAP), include DCIS, Ki-67, age, feeding, and lymph node grade. SHAP analysis indicated that lower values of DCIS and Ki-67, lower lymph node and tumor grade, less sentinel lymph node involvement, higher age, longer duration of breastfeeding, lower tumor size, medullary tumor type, PR+, ER+, less positive node, more pregnancy, DCIS pathology class, BCS surgery type for early-stage, lower tumor stage, no necrosis, HER2-, no family history, and no calcification are result in longer breast cancer survival. Furthermore, we validate the model's performance using 5-fold cross-validation (CV) and nested CV. We also evaluate model performance using accuracy, precision, F1 score, and recall. All results show the proposed method outperforms. This paper presents clinical, molecular, and demographic insights into survival analysis, utilizing explainable ML techniques to support personalized treatment decisions.