Back to 2026 Abstracts
Temporally Validated Machine Learning for 30-Day Outcomes After Mastectomy and Reconstruction for Breast Cancer
Dany Y. Matar, B.A.1,2; Anthony Y. Matar, B.A.3; John J. Lee, B.A.2,4; Anahita Nimbalkar, B.S.2; Arnav Khera, B.S.2; Jonlin Chen, M.D.2; Adriana C. Panayi, M.D., Ph.D. 5: ; Jasleen Gill, B.S.1; Carla M. Hamwi, B.S.6; Charbel Saad, M.D.2; Esther Jung, D.O.2; Gabriel Hundeshagen, M.D.7; Dennis P. Orgill, M.D., Ph.D.1; Sashank K. Reddy, M.D., Ph.D.2; Hooman Soltanian, M.D.2; Salih Colakoglu, M.D., Ph.D.2. 1 Division of Plastic Surgery, Department of Surgery, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA 2 Department of Plastic and Reconstructive Surgery, Johns Hopkins Hospital, Johns Hopkins University School of Medicine, Baltimore, MD, USA 3 Machine Learning Core, National Institute of Mental Health, National Institutes of Health, Bethesda, MD, USA 4 University of Virginia School of Medicine, Charlottesville, VA, USA 5 Department of Cranio-Maxillofacial and Oral Surgery, University Hospital Zurich, University of Zurich, Rämistrasse 100, Zurich, 8091, Switzerland 6 Georgetown University School of Medicine, Washington DC, USA 7 Department of Hand, Plastic, and Reconstructive Surgery, BG Klinik Ludwigshafen, Ludwigshafen, Germany
Background: Patients undergoing oncologic and reconstructive breast surgery have heterogeneous baseline and procedure-specific risks that limit conventional perioperative tools. We developed temporally validated, interpretable machine-learning models to predict 30-day adverse events after cancer-directed breast surgery. Methods: ACS-NSQIP data (2014–2024) identified 366,992 adults undergoing mastectomy alone, immediate reconstruction, or delayed reconstruction. Logistic regression, XGBoost, LightGBM, neural network, and stacked ensemble models were trained (2014–2021), validated (2022–2023), and tested (2024) using preoperative variables. Outcomes included any or serious complication, mortality, surgical site infection, unplanned reoperation, and major medical complications. Performance was assessed by discrimination, calibration, and risk stratification. Results: Among 366,992 patients, most underwent mastectomy alone (70.6%), followed by immediate (23.4%) and delayed reconstruction (6.0%). Thirty-day event rates were 6.8% (any complication), 4.8% (serious complication), 3.4% (surgical site infection), 3.3% (unplanned reoperation), and less than 0.1% (mortality). XGBoost was selected for any complication, serious complication, mortality, and surgical site infection, while LightGBM was selected for unplanned reoperation. In the held-out 2024 cohort, observed event rates increased monotonically across risk strata, with high-risk groups comprising 1.4% to 8.1% of patients but capturing disproportionate adverse events, including 46.7% of deaths and 20.8% of surgical site infections. Selected models achieved AUROCs of 0.670 to 0.892 and integrated calibration indices of 0.004 to 0.032. Models were deployed as a publicly available risk calculator (https://pro-breast.streamlit.app/). Conclusions: Interpretable, outcome-specific models provided clinically meaningful risk stratification and may support counseling, shared decision-making, and perioperative planning, pending external validation.
Back to 2026 Abstracts
