Back to 2026 Abstracts
Machine Learning-Based Prediction of Tissue Expander Complications: A Multi-Institutional Analysis
Ariel Gabay
*1, Timothy C. Olsen
2, Archana Babu
3, Jonlin Chen
4, Abbas Hassan
5, Robyn B. Broach
2, Babak J. Mehrara
1, Robert J. Allen
1, Said Azoury
2, Alexander F. Mericli
3, Jonas Nelson
11Plastic and Reconstructive Surgery, Memorial Sloan Kettering Cancer Center, New York , NY; 2Plastic and Reconstructive Surgery, University of Pennsylvania, Philadelphia, PA; 3Plastic and Reconstructive Surgery, M.D. Anderson Cancer Center, Houston, TX; 4Plastic and Reconstructive Surgery, Johns Hopkins University, Baltimore, MD; 5Plastic and Reconstructive Surgery, Indiana University, Indianapolis, IN
Background: Our group previously developed machine learning (ML) and nomogram-based risk prediction tools for individualized preoperative assessment of tissue expander (TE) complications, including TE loss, infection, and seroma. Building on this work, we aimed to further develop and evaluate ML models for predicting TE complications across a large multi-institutional cohort of patients undergoing immediate TE-based breast reconstruction for external validation.
Methods: Patient characteristics, surgical techniques, and TE complications were collected for patients undergoing immediate TE placement from 2007-2025 at Memorial Sloan Kettering Cancer Center (MSKCC), MD Anderson Cancer Center (MDACC), and the University of Pennsylvania (UPenn). ML models (XGBoost, LightGBM, Random Forest, AdaBoost, logistic regression, and neural networks) were trained to predict TE loss, seroma, and infection. Model performance was assessed using area under the curve (AUC), sensitivity, specificity, and Brier score. SHapley Additive exPlanations (SHAP) values ranked predictors influencing model outcomes.
Results: In total, 10,589 patients were included (MSKCC: n=7,956; MDACC: n=1,711; UPenn: n=922), with a mean age of 49.4 years (SD 10.9) and mean BMI of 24.9 kg/m
2 (SD 5.2). TE loss occurred in 4.8% of patients, infection in 6.3% and seroma in 8.7%. Top performing ML models achieved moderate performance, with AUCs of 0.792 (95% CI: 0.719-0.817) for TE loss (XGBoost), 0.731 (95% CI: 0.618-0.731) for infection (LightGBM), and 0.705 (95% CI: 0.62-0.712) for seroma (AdaBoost). Key predictors included dissection plane, age, BMI, chemotherapy timing, reconstruction laterality, and smoking status.
Conclusions: ML models demonstrated moderate performance for predicting TE complications across a large, multi-institutional cohort. These findings support the generalizability of our prior work, and future studies will focus on the development of real-time clinical decision-aid tools to facilitate personalized preoperative counseling in alloplastic breast reconstruction.
Models adjusted for BMI, tissue expander size, and axillary irradiation. OR > 1 indicates higher odds with early exchange.
Back to 2026 Abstracts
Models adjusted for BMI, tissue expander size, and axillary irradiation. OR > 1 indicates higher odds with early exchange.
