Artificial Intelligence Enabled Prediction of 30-day complications after Breast Surgery: A preliminary analysis of 737,730 patients

Le résumé fourni par la source

PURPOSE: Disease-directed breast surgery spans diverse resection and reconstructive procedures with substantial postoperative risk, yet existing calculators lack procedural breadth, calibration, and individualized prediction. We developed, validated, and deployed machine learning models to predict major 30-day postoperative complications. METHODS: A retrospective cohort study was conducted using the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) database from 2008-2024. Adult patients undergoing disease-directed breast surgery were included (i.e., mastectomy, mastectomy with immediate reconstruction, or reconstruction/revision alone). Eighty-three perioperative variables (demographics, comorbidities, surgical indication, preoperative lab values, concurrent procedures, and operative characteristics) were used to develop five ML architectures: tuned logistic regression, XGBoost, LightGBM, neural network, and stacked generalization. Each model was independently developed to predict five major 30-day postoperative outcomes: surgical complications, unplanned reoperation, medical complications, venous thromboembolism (VTE), and mortality. Data were randomly partitioned into training (75%), validation (15%), and held-out test (15%) cohorts using stratified sampling. Primary evaluation emphasized clinically oriented risk stratification in the test cohort. Secondary performance metrics included AUROC, calibration measures, confusion matrices, F1 score, accuracy, precision, and recall. Model interpretability was assessed using SHAP analyses. RESULTS: Using data from 737,730 adults undergoing disease-directed breast surgery, we developed PRO-BREAST ( P ost-resection and R econstruction O utcome prediction for BREAST surgery), a publicly available outcome-specific risk prediction calculator ( https://pro-breast.streamlit.app/ ) that provides individualized and interpretable risk estimates. Thirty-day event rates for predicted complications in the study cohort were 3.9% for surgical complications, 3.3% for unplanned reoperation, 0.7% for medical complications, 0.2% for VTE, and 0.1% for mortality. The cohort included mastectomy alone (66.7%), reconstruction or revision alone (15.7%), and mastectomy with immediate reconstruction (17.7%). Outcome-specific models were selected from five candidate architectures based on clinical risk-stratification performance in the test set, defined by the ability to concentrate observed complications in the high-risk category relative to the overall cohort (lift). XGBoost and LightGBM were selected for PRO-BREAST , with XGBoost exhibiting highest lift for all primary outcomes except mortality. Observed complication rates increased monotonically from very low- to high-risk strata for selected models, ranging from 0.43%-4.61% for medical complications (lift up to 6.85), 1.45%-14.74% for unplanned reoperation (lift 4.45), 1.16%-15.49% for surgical complications (lift 4.01), 0.11% to 1.76% for VTE (lift 8.66), and 0.02%-1.62% for mortality (lift 29.46). High-risk groups comprised approximately 1-7% of patients yet captured a disproportionate share of events, including 42.6% of deaths. SHAP analyses showed that surgical complications and unplanned reoperation were driven primarily by procedural factorsparticularly reconstructive approach, extent of resection, and operative timewhereas medical complications and VTE reflected patient frailty and comorbidity burden. Mortality risk was dominated by markers of global physiologic vulnerability, with minimal influence from procedural variables. CONCLUSION: In a large, nationally representative surgical cohort, machine learning models accurately stratified risk for major 30-day postoperative outcomes following disease-directed breast surgery. These models enabled development of PRO-BREAST , an interpretable and publicly accessible risk prediction tool to support individualized preoperative counseling and surgical decision-making. External validation and prospective evaluation are warranted to assess clinical impact.© 2026. Plastic Surgery Research Council | All rights reserved |*Source: https://ps-rc.org/meeting/Program/2026/OS14.cgi*

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.