Northeastern Society of Plastic Surgeons

NESPS Home NESPS Home Past & Future Meetings Past & Future Meetings

Back to 2026 Abstracts

PRO-BREAST: Interpretable Machine Learning for Risk Prediction After Oncologic and Reconstructive Breast Surgery
Dany Y. Matar*1, John J. Lee2, Anthony Matar3, Arnav Khera1, Jonlin Chen1, Adriana C. Panayi4, Anahita Nimbalkar1, Jasleen Gill5, Charbel Saad1, William Huang6, Esther Jung1, Gabriel Hundeshagen4, Dennis P. Orgill5, Sashank Reddy1, Hooman Soltanian1, Salih Colakoglu1
1Department of Plastic and Reconstructive Surgery, Johns Hopkins Hospital, Johns Hopkins University School of Medicine, Baltimore, MD; 2University of Virginia School of Medicine, Charlottesville, VA; 3Machine Learning Core, National Institute of Mental Health, National Institutes of Health, Bethesda, MD; 4Department of Oral and Maxillofacial Surgery, Charité – Universitätsmedizin Berlin, Berlin, Germany; 5Division of Plastic Surgery, Department of Surgery, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA; 6George Washington University School of Medicine, Washington D.C.,

Background: Breast surgery spans heterogeneous oncologic and reconstructive procedures with widely variable risk. Existing calculators rely on linear models with limited procedural granularity, leaving clinicians without procedure-specific estimates for preoperative counseling.
Methods: We retrospectively analyzed adults undergoing mastectomy, breast reconstruction, or revision in ACS-NSQIP (2008-2024). Eighty-three perioperative variables, including reconstructive timing, laterality, and procedural complexity, were used to develop outcome-specific machine learning (ML) models for five 30-day endpoints: surgical complications, medical complications, unplanned reoperation, venous thromboembolism (VTE), and mortality. Logistic regression, XGBoost, LightGBM, neural network, and stacked architectures were trained on 70/15/15 splits and compared by AUROC, calibration (ICI), lift, and SHAP.
Results: Among 737,730 patients (mastectomy alone, 66.7%; reconstruction or revision, 15.7%; combined, 17.7%), XGBoost was selected for surgical, medical, reoperation, and VTE outcomes; LightGBM for mortality. Models showed strong performance for surgical complications (AUROC 0.75, lift 4.01), reoperation (0.77, 4.45), medical complications (0.73, 6.85), VTE (0.80, 8.66), and mortality (0.88, 29.46), with excellent calibration (ICI 0.002-0.016). SHAP revealed two risk profiles: surgical complications and reoperation were driven by procedural complexity, operative time, and resection extent, while medical complications, VTE, and mortality reflected patient frailty. Models were deployed as PRO-BREAST (https://pro-breast.streamlit.app/), an open-access calculator providing individualized estimates and patient-level drivers.
Conclusion: Interpretable, outcome-specific ML accurately predicts 30-day complications across breast surgery. PRO-BREAST translates these models into a scalable tool for preoperative counseling.
Back to 2026 Abstracts