Back to 2026 Display Posters
Interpretable Machine-Learning Risk Prediction for 30-Day Outcomes After Panniculectomy
John J. Lee B.A.1,2, Bashar M. Alsamman B.A.2, Dany Y. Matar B.A.2, Anahita Nimbalkar B.S.2, Dennis P. Orgill M.D., Ph.D.3, Gabriel Hundeshagen M.D., M.M.S.4, Adriana C. Panayi M.D., Ph.D.5 1University of Virginia School of Medicine, Charlottesville, VA, USA 2Department of Plastic and Reconstructive Surgery, Johns Hopkins Hospital, Johns Hopkins University School of Medicine, Baltimore, MD, USA 3Division of Plastic Surgery, Department of Surgery, Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA 4Department of Hand, Plastic and Reconstructive Surgery, Burn Center, BG Trauma Center Ludwigshafen, University of Heidelberg, Ludwigshafen 67071, Germany 5Department of Cranio-Maxillofacial and Oral Surgery, University Hospital Zurich, University of Zurich, Rämistrasse 100, Zurich, 8091, Switzerland
Background: Panniculectomy carries substantial wound-dominated morbidity, but existing risk tools are non-procedure-specific, single-outcome, or poorly calibrated in body contouring. We developed and validated interpretable machine-learning models for 30-day outcomes after panniculectomy. Methods: All ACS-NSQIP panniculectomy records from 2014-2024 (CPT 15830, n=31,334) were split by year: training 2014-2021, validation 2022-2023, held-out testing 2024 (n=2,829). Forty features were modeling, including nine concurrent-procedure flags. Six primary outcomes (wound, any-complication, and serious-complication composites; reoperation; readmission; transfusion) and eight secondary (superficial, deep incisional, and organ/space SSI; dehiscence; venous thromboembolism; sepsis; non-home discharge; mortality) were modeled. Logistic regression, LightGBM, XGBoost, neural-network, and stacked-ensemble models were Optuna-tuned, calibrated on validation, and assessed by AUROC, calibration index, four-bin lift, and SHAP values. Results: Patients were 88.7% female, median age 47 years, median BMI 30.5 kg/m². Event rates were 8.2% for wound complications, 13.3% any complication, 5.8% serious complication, 3.9% reoperation, 5.7% readmission, and 0.2% mortality. In the 2024 test cohort, deployed models achieved AUROC 0.656-0.725 across the six primary outcomes and 0.79-0.96 for organ/space SSI, sepsis, and non-home discharge, with close calibration (ICI 0.016 or lower). Observed event rates rose monotonically across bins, with high- versus very-low-bin lift of 3.1-7.7 (wound complication, 27.0% vs 5.6%). BMI, ASA class, albumin, and concurrent ventral hernia repair were leading contributors. Conclusion: Machine learning models provide calibrated, interpretable 30-day risk estimates under strict temporal validation. Discrimination is modest for common wound outcomes, but reliable absolute risk and stratification support counseling and perioperative planning. External validation is warranted.
Back to 2026 Display Posters