MOLSTM for Survival Prediction in Breast Cancer: A Dynamic Feature Aggregation Approach

Inspired by the MA.5 dataset on early - stage breast cancer, this study compares relapse - free survival rates of patients on novel (CEF) and standard (CMF) chemotherapy regimens and evaluates their quality - of - life (QoL) via the BCQ emotional sub - scale. The data presents challenges such as ceiling/floor effects in BCQ scores and non-linear QoL-time relationships. To address these, we apply data transformations, generate derivative QoL variables, use a change-point detection model with feature construction, and develop a hierarchical MOE + bidirectional LSTM model for survival prediction. The approach achieves a C-index of 0.66 and AUC over 0.69, outperforming baseline models. SHAP analysis aids interpretation. Clinically, this model provides individualized survival predictions and QoL assessments, supporting treatment decision-making and enhancing understanding of factors impacting prognosis.