SP40. Machine Learning Prediction Tool for Patient Reported Outcomes in Breast Cancer Patients

PURPOSE: Our group previously showed that patients undergoing breast-conserving therapy (BCT) report higher quality of life and satisfaction BREAST-Q scores compared to those receiving postmastectomy breast reconstruction (PMBR). However, no tool exists to predict these scores based on individualized data and patient preferences. This study aimed to use machine learning to predict BREAST-Q scores for BCT and PMBR patients and develop an online tool for predicting one-year postoperative improvements. METHODS: Female patients over 18 who underwent BCT or PMBR at Memorial Sloan Kettering Cancer Center from 2017-2022 and completed the BREAST-Q before and after surgery were included. Exclusion criteria were prophylactic surgery or stage IV breast cancer. Model variables included demographics (i.e. age, BMI), surgery type, baseline BREAST-Q scores, chemotherapy, radiation, cancer stage, 90-day reoperation and readmission, and psychiatric history. Five machine learning models were developed to predict improvements in satisfaction with breasts, physical, sexual, and psychosocial well-being. Data were split into training (60%), validation (20%), and test (20%) sets. Model performance was assessed using Area Under the Curve (AUC) and model interpretability assessed using Shapley Additive exPlanations (SHAP). RESULTS: We analyzed 6,215 patients (2,501 BCT; 3,714 PMBR). At one year, 49.7% reported improved satisfaction with breasts, 32.4% physical well-being, 47.4% psychosocial well-being, and 15.5% sexual well-being. All models demonstrated high predictive performance: AUCs were 0.898 for satisfaction with breasts, 0.816 for physical well-being, 0.762 for psychosocial well-being, and 0.71 for sexual well-being. LightGBM outperformed AdaBoost for psychosocial well-being (p < 0.05), and XGBoost outperformed Random Forest for satisfaction (p < 0.05). Key predictors included baseline BREAST-Q scores, BMI, age, and surgery type. CONCLUSION: We developed five machine learning models that accurately predict one-year postoperative BREAST-Q scores for BCT and PMBR patients. These models will be incorporated into an online tool for clinical decision-making, allowing surgeons to compare predicted PROs for patients undergoing either BCT or PMBR.