Work

45. Machine Learning Accurately Predicts Patient Reported Outcomes 1 Year After Breast Reconstruction

Year: 2025

Type: article

Abstract: PURPOSE: Patient reported outcomes (PROs) are of great importance in breast reconstruction, which aims to restore a patient’s appearance, function, and wellbeing. Currently, there are no reliable predictive tools to determine PROs after reconstruction. With machine learning, automated individualized preoperative risk assessments and predictors have become a reality. This study aims to develop a machine learning approach to predict improved or not improved 1-year follow-up BREAST-Q scores for patients undergoing breast reconstruction. METHODS: We collected data on patent characteristics, medical history, reconstructive techniques, and BREAST-Q scores for women undergoing implant and autologous reconstruction at Memorial Sloan Kettering Cancer Center between January 2010 and March 2024. Supervised machine learning algorithms were developed to predict if patients would experience improved versus not improved 1-year post-reconstruction BREAST-Q scores. The dataset was divided into training, testing, and validation sets in a ratio of 60:20:20. We employed fivefold cross-validation to tune model hyperparameters and externally validated models using separate data from the international Mastectomy Reconstruction Outcomes Consortium (MROC). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Brier score. Shapley Additive explanations (SHAP) were calculated for model interpretability. Top-performing machine learning models were integrated into an online clinical prediction application. RESULTS: A total of 2,687 patients were included in model development. All five machine learning algorithms demonstrated excellent predictive performance: AUCs were 0.97 (95% CI: 0.90 - 0.99) for physical wellbeing of the abdomen, 0.86 (95% CI: 0.80 - 0.87) for satisfaction with breast, 0.79 (95% CI: 0.72 - 0.80) for sexual wellbeing, 0.78 (95% CI: 0.72 - 0.79) for physical wellbeing of the chest, and 0.74 (95% CI: 0.66 - 0.74) for psychosocial wellbeing. Upon external validation on 2,089 patients from separate multi-institutional data, models similarly achieved high AUCs of 0.86 (95% CI: 0.81 - 0.90) for satisfaction with breast, 0.85 (95% CI: 0.80 - 0.94) for physical wellbeing of the abdomen, 0.78 (95% CI: 0.69 - 0.80) for physical wellbeing of the chest, 0.76 (95% CI: 0.69 - 0.80) for psychosocial wellbeing, and 0.73 (95% CI: 0.65 - 0.77) for sexual wellbeing. Based on SHAP values, variables that contributed the most to model predictions across all domains were preoperative BREAST-Q scores, timing of radiation, body mass index, age, and reconstructive technique. CONCLUSIONS: Using a large cohort of BREAST-Q patient data, we developed five machine learning models and a web-based application to predict BREAST-Q scores after reconstructive surgery. Our findings provide surgeons an accessible resource to input patient-specific data and predict 1-year satisfaction and quality-of-life outcomes. Ultimately, this data-driven approach can streamline decision-making and enhance patient-centered care in breast reconstruction.

Source: Plastic & Reconstructive Surgery Global Open

DOI: 10.1097/01.gox.0001112132.58213.03