Extremity Soft Tissue Sarcoma Reconstruction Nomograms: A Clinicoradiomic, Machine Learning-Powered Predictor of Postoperative Outcomes

Abstract

PURPOSE The choice of wound closure modality after limb-sparing extremity soft-tissue sarcoma (eSTS) resection is fraught with uncertainty. Leveraging machine learning and clinicoradiomic data, we developed Sarcoma Reconstruction Nomograms (SARCON), a tool that provides probabilistic estimates of five adverse outcomes on the basis of the selected reconstructive modality. METHODS This retrospective cohort study of limb-sparing eSTS resections integrated clinical variables and radiomic features, including eSTS and limb dimensions. Target outcomes included surgical site infections (SSI), wound dehiscence (WD), seroma formation, and minor and major complications. For each outcome, three machine learning classifiers—Logistic Regression with Lasso regularization, Naïve Bayes, and FasterRisk—were developed and evaluated using 10-fold cross-validation (CV), 50 random 80%-20% splits, leave-one-out CV, and a test data set. The best-performing model for each outcome was used to construct a respective nomogram. RESULTS A total of 316 limb-sparing eSTS resections were analyzed, predominantly located in the thigh (54%), lower leg (17%), and upper arm (11%). Postoperative outcomes included SSI (12%), WD (16%), seroma formation (8.5%), minor complications (34%), and major complications (25%). Logistic Regression with Lasso regularization consistently outperformed the other models across all outcomes, achieving area under the receiver operator curves ranging from 0.83 to 0.93 in all tests.